均衡性原则（Der Grundsatz der Angemessenheit）是公法“帝王原则”比例原则的子原则之一，它要求公权力行为手段增进的公共利益与其造成的损害成比例，故均衡性原则又称为狭义比例原则。某个能实现正当目的的手段虽然是最小损害的，但如果其所促进的公共利益与其所造成的损害不成比例，对公民权利造成过度损害，就不符合均衡性原则。可见，均衡性原则并非片面强调公共利益至上，它要求公权力行为者在追求公共利益的同时，认真对待公民权利，审慎权衡相关利益。
由此可见，司法实践并没有为均衡性原则的适用提供足够的判断均衡性与否的客观标准。学者近些年来开始不断尝试具体化均衡性原则的方法，他们从不同角度提出了一些使均衡性原则的适用更具有理性的方法。但是对于究竟该如何具体化均衡性原则，并没有形成一个统一的结论。归结起来，具体化均衡性原则的方法主要可分为两种模式：一种是从权衡者的角度出发，以德国学者罗伯特·阿列克西（Robert Alexy）为代表的数学计算模式；另一种是从当事人的角度出发，以加拿大学者戴维·M·贝蒂（David M. Beatty）为代表的事实问题商谈模式。
对于具体化均衡性原则中权衡的方法，阿列克西提出了权衡法则（Law of Balancing）:“对一个原则的未满足程度或损害程度越大，满足另外一个原则的重要性就必须越大。”此权衡法则为实质的权衡法则（Substantive Law of Balancing），可称为第一权衡法则。由于阿列克西认为权利即是原则，所以第一权衡法则也可以表述为：“对一个权利的未满足程度或损害程度越大，满足另外一个权利的重要性就必须越大。”
在对经验前提的可靠性进行赋值时，阿列克西提出了第二权衡法则：“对一项基本权利的干涉越大，其经验前提的可靠性就必须越大”。第二权衡法则属于认识的权衡法则（Epistemic Law of Balancing）。它是对经验事实认定的确定性的判断。对分量公式不同的变量赋值后，就可以进行计算。为了尽量避免僵局的出现，阿列克西又提出了更为精细的刻度划分，即双重的三刻度模式。对权利限制可以分为九重刻度：（1）轻度的轻度（ll）;（2）轻度的中度（lm）;（3）轻度的重度（Is）;（4）中度的轻度（ml）；（5）中度的中度（mm）;（6）中度的重度（ms）;（7）重度的轻度（sl）；（8）重度的中度（sm）;（9）重度的重度（ss）。对这九重刻度的赋值，分别为20、21、22、23、24、25、26、27、28。
近些年来，在司法实践中，国内外越来越多的司法案例已经提出，有必要通过成本收益分析方法来判断均衡性。例如，在Pfizer Animal Health SA v. Council一案中，欧洲法院认为，为了正确地判断争论的是非曲直，将成本收益分析与比例原则联系起来是适当的。在分析推理中，法院认为应当依据成本收益分析，判断风险管制所产生的侵害与不采取任何行为的益处是否不成比例。
“均衡性原则需要判断手段与目的是否成比例：与收益有关的成本是否过度。”关注手段的财政支出成本，已逐渐得到了法院的重视。在Schwerbehindertengesetz案中，德国联邦宪法法院认为：“如果某个可替代性的手段对权利损害更小，但要求更多的行政成本，与目的间不具有均衡关系，这个手段也应当被排除。”类似地，在Rückkehrgebot für Mietwagen案中，联邦宪法法院也认为，虽然对权利是更小损害，但如果要求更多的行政成本，不具有期待可能性，也应当否定此项手段。在其他一些案件中，联邦宪法法院也有许多相似表述。
1971年，罗尔斯《正义论》的出版被公认为是现代哲学界的一个标志性事件，它扭转了20世纪西方（尤其英美）哲学长时段沉浸于理论分析而疏于实践哲学的基本趋势，正义问题成为时代的哲学主题，创造了几乎整个哲学社会科学界的“罗尔斯产业”（Rawls’ industry）之学术景观。罗尔斯正义论的鼎兴并非偶然，其个人经历及其自觉的时代反思促使他最终聚焦并洞开现代社会正义的问题症结：制度性非正义及其校正。罗尔斯本人参加太平洋战争的切身经历和他对于第二次世界大战后欧美社会的观察分析，使他敏锐地意识到其所处时代最具危机和挑战意味的社会病症：诸如，为何人类历史上最为惨烈的两次世界大战都首先在“自由民主国家”之间爆发？为何以自由民主著称于世的美国直到20世纪60年代末还爆发以黑人为主流的少数族裔“民权运动”（civil rights movements）？为何在20世纪60-70年代的自由民主国家，青年一代不仅没有因为社会经济的增长繁荣而感受生活的快乐，反而会成为“垮掉的一代”（a broken generation）？凡此种种，在罗尔斯的哲学思考中渐渐汇聚成社会伦理暨政治哲学的严肃课题：即使在一个政治民主、经济繁荣的现代国家，若缺乏基本的社会公平正义，再多的财富和收入都无法消除社会冲突，也无法建立“一个良序社会”（a well-ordered society），更甭说维持其长治久安了。现代自由民主社会最为缺乏的不是社会财富的生产和效用的“充量化”（maximization），而是基于社会公平正义的普遍道义规范，作为普遍社会道义规范基础的基本制度架构，以及，由此所构成并长久维持的“社会稳定性”（social stability）。换言之，在罗尔斯看来，现代社会的“首要美德”或最高成就不是任何形式的社会总体化功效及其最大化增长，而是如何基于公平正义的原则分配或分享社会的“基本善品”，也就是社会基本自由、平等、权利、义务、财富和收入、自尊等的公平分享或分担，以及最为重要的是，确保自由民主社会长治久安的正义制度。
通过反省近代以来长期支配西方实践哲学传统的社会契约论和功利主义学说，罗尔斯断定，它们已无法满足现代自由民主国家对于正义秩序和公平制度的内在需求。传统的社会契约论已失去其令人信服的理论力量，而无论功利最大化到何种程度，单纯总体功利（价值）目的及其实现都因无法解决社会分配不公而难以建构现代社会持久稳定的正义秩序。任何良序社会的根基只能诉诸基于社会公平正义的普遍道义原则，而非功利或效用的最大化原则。社会契约论有助于现代民主国家在文化多元条件下达成普遍正义的原则共识，因为社会契约精神——亦即自由平等的协商精神——更适宜于社会公平正义原则的规范要求。然而，在罗尔斯看来，近代社会契约理论——无论是霍布斯、洛克式的，还是卢梭式的，甚或是康德式的——都不足以充分证成现代民主国家所需要的制度正义。它们或因过于消极而难以充分阐明作为一种积极的社会组织化成果的“社会合作体系”之国家理念，仿佛缔结社会契约或建立国家只不过是人类为了克制自身本性的“根本恶”——即自私本性——以摆脱恶劣的自然状态而不得不做出的被动选择；或因过于倚赖于传统自然法观念而缺少新理论的充分证明，使得原有的社会契约无法为任何普适的现代社会道义原则提供充分必要的理论论证，因之必须给予理论重铸和改进，使其获得充分的理论合理性和普遍有效性，成为足以证成“作为公平的正义”（justice as fairness）这一核心正义观的合理有效方式。于是，罗尔斯首要的哲学工作便是改进近代西方社会契约论，以建构其“高度康德式的”新社会契约论。
与传统社会契约论不同，新社会契约论不再从某种基于“趋利避害”之人性本恶假设的“自然状态”出发，而是基于一种信息预制过滤，也就是罗尔斯所谓的“原初状态”（an original position）下的正义原则认同。在“原初状态”中，参与社会契约的每一个人都被遮蔽于“无知之幕”（the veil of ignorance）背后，不仅相互“无知”彼此的实际差别，而且“互无利益关涉”（mutual disinterested），毫不介意诸如起点差异可能带来的契约预知困境或结果偏颇，仅仅依据参与各方对公平正义这一核心理念的广泛认同，达成并承诺正义之基本原则体系。罗尔斯的论证逻辑次序是：首先确认并确信无人反对“作为公平的正义”这一核心理念；其次就基本正义原则系统达成基本共识，以便为建构“一个良序社会”奠定正义的制度基础或“社会基本结构”（basic social structure）；最后是如何践行正义基本原则规范的规导问题和实践操作（策略步骤）问题。
罗尔斯当然不会看不到这些理论悖论。为此，他花费了巨大的精力来消除人们的质疑。罗尔斯首先申言，基于正义原则主导的制度安排所遵循的是普遍理性主义的社会道义论而非功利主义所坚持的价值目的论，因此它必须突破后者不可普遍化的局限，将社会正义制度的基点从高限“最大化”（maximization）降低为“底线最大化（普适化）”（mini-maximization），即一种兜底式的或惠顾尽可能多数人的普遍正义的制度安排，这是“差异原则”的宗旨所在。因为无论是在理论上还是在实践中，社会效用总量的最大化都不一定确保社会最大多数人都能得到公平的份额，更甭说因各种主客观原因可能造成的极少数社会弱者被边缘化，甚至被遗忘。其次，罗尔斯明确开出了一份社会“基本善（品）清单”（the list of primary goods），包括确保每一个社会个体过上“体面生活”所不可缺少的平等自由（权利）、基本收入和财富、人格尊严等。罗尔斯认为，任何“基本善（品）”的差别化分配，尤其是财富和收入的差别化分配，只有在其同时有利于改善社会弱者的前提下才具有普遍公平正义的道义性。否则，即便该分配合理合法，也未必正当合义。当然，“基本善（品）清单”的开列本身也表明，制度分配对极少数社会弱者的“惠顾”也并非毫无限度，它仅限于“基本善（品）清单”之内的制度分配。同时，为了回应激进自由主义和保守主义对“差异原则”的诘难，罗尔斯通过其“词典式次序”的自我辩护，证明其正义原则并未违背自由主义的基本信条。诺齐克等较为激进的自由主义思想家认为，罗尔斯的“差异原则”明显违反了个人自由权利神圣不可侵犯这一自由主义根本信条，且无法在不违背此一信条的情况下证成差异原则及其实践的正当合理性，若无“校正正义”（justice of correcting）“转移正义”（justice of transferring）的限制，任何通过制度安排所实施的权利转让（重新分配）都不正义。对此，罗尔斯反复申明，尽管“差异原则”表达了对惠顾少数处于最不利地位者的制度关照，但就整个正义原则体系而论，第一原则优先于第二原则，在第二原则的两部分之间，机会均等原则又优先于差异原则，最重要的是，正义原则体系内的这种优先性次序一如词典中诸词条的排序一样不可更改，否则，该辞典便无法使用。同样，正义的两个基本原则及其第二原则两个部分之间，也具有不可更改的次序规定。可问题是，如若这样，“差异原则”所诉求的惠顾少数弱者的制度偏好诉求又如何可能？
赞赏或批评罗尔斯正义论的现代哲学家和思想家甚多，但能直接对之做出关键性甚或替代性改进的当属其哈佛哲学系同事和朋友阿玛蒂亚·森。凭借对罗尔斯正义论的多学科透视分析，森在综合其多年有关罗尔斯《正义论》探讨和正义问题研究的基础上，于2009年发表其正义研究的总结性成果《正义的理念》（The Idea of Justice）专著，全面清算并改进了罗尔斯的正义论，堪称罗尔斯《正义论》之后又一部现代正义论经典。在当代哲学乃至整个人文社会科学界，森大概是最懂罗尔斯的学术知己和思想大家之一，也是能以独特的理论批评和创造性改进，来表达其学术敬意并证明其与罗尔斯之间崇高学术友谊的真诚道友。森对罗尔斯的敬仰如此真诚，在《正义的理念》一书的扉页上，他特别题词：“为了纪念约翰·罗尔斯”，并在该书第二章“罗尔斯及其超越”的结语中满含深情地写道：“……我们不能让罗尔斯的正义思考模式陷入一种理智‘停滞’。我们必须从罗尔斯那里获得的丰富思想中吸取有益的养分，然后继续前行，而不是‘休假’。我们需要‘正义’（justitia），而非‘休假’（justitium）。”
基于对社会选择理论优于社会契约论的理论确信，森首先将批评的矛头指向罗尔斯正义论的前提预设，认为罗尔斯所信奉的社会契约论进路并不合理，即使经过罗尔斯“擢升”后的新社会契约论，也未能达到更高合理性，犯了“先验制度主义”（transcendental institutionalism）的错误。罗尔斯假设，出于建构“良序社会”的目的，人们需要首先就正义的基本原则达成一致，这就需要他们首先超脱各种人际或群际差异，包括人生起点、生活境遇、个人主体条件，以便超越因诸种差异而产生的意见纷争，仿佛每一个参与社会契约的人都预先自觉罩上一层“无知之幕”，以此形成各社会成员得以开始的齐一化“原初状态”。从“同一起跑线”出发，大家才有可能对基于“作为公平的正义”这一核心理念所形成的正义原则体系达成共识——即：无论如何，一种能够体现公平正义的社会总是值得人类社会期待的，进而人们按照正义的基本原则体系来建构“良序社会”所必需的正义制度体系，这正是罗尔斯把正义珍视为社会制度之“第一美德”（the first virtue）的根本原因。然而，在森看来，这种前提假设既缺乏充足的逻辑理由，也没有任何现实可能。即使为了缔结公平正义的社会，也没有任何理由预先假定所有参与缔结契约的人都必须从“同一起跑线”出发，甚至为了“平整”（海德格尔术语）所有参与者的实际差别而强行给所有人罩上“无知之幕”。寻求结果的一致不是预设起点齐一的正当理由，更非充足理由。恰恰相反，任何社会契约的达成都是在厘清差异、相互妥协的前提下实现的，即所谓异中求同，同中存异。无差别的契约不是真正的契约，而是静止的无过程的想象性社会认同。现实生活中，既不可能强行制造“无知之幕”，也不存在一种无差别、“无利益关涉”的所谓“原初状态”。任何强制性加罩“无知之幕”的做法都是一种武断的“信息限制”，以此造成的“原初状态”无异于“信息茧房”，这显然不是“自由民主的国家”该有的选择。
可见，社会契约论并不是证成现代正义制度和正义原则体系的最佳方式，远不如社会选择理论更适合切入和解释现代正义问题。森指出：“一种正义理论必须针对实际可行的选择而论，而不只是让我们总沉溺于一种想象的、不可信的、无比壮美的世界而无法自拔。”任何合理可行的正义理论首先必须是从社会实际出发，针对现实生活中的非正义事实做出真实可行的理论反应，而非基于正义理想的美好想象，预定某种完型原则。罗尔斯显然落入了“高限”理想目的论的窠臼，无法避免地重复其理论前辈（如，康德）所遭遇的“先验制度主义”困境：它预设我们可以对个体行为和社会行为预先做出完全充分的评估，并借此预先制定一套完美的正义评价标准。可事实恰好相反，人类从来都是因为面临非正义的社会现实且无法回避，才开始思考和探寻社会正义制度和正义原则的。因此森断定，先验制度主义的“完全评估”之难正是社会行为选择理论的现实比较优势所在。“社会选择理论”（social choice theory）最早诞生于18世纪法国大革命时期，产生于人们对卢梭抱怨的人类不平等社会现实做出激进反应的“革命时刻”，它首先由巴黎著名数学家波达（Jean-Charles de Borda）和社会学家孔多塞（Marquis Condorcet）发明，也正是“通过他们对一群不同个人的个体评价之集合性学科调查”，社会选择理论才得以提出。社会正义及其实现的确同社会制度及其与个体行为模式的相互依赖性作用有关，但人类社会的制度选择具有某种“依势而定的本性”（contingent nature），人们无法穷尽对各种差异性个体行为的评估检测，却可以从实际存在的社会非正义现实中发现问题，在多种比较分析后做出切实可信的相对客观的比较评估，以此作为探寻建构社会正义的合理起点。森的结论是：“一种先验的进路本身无法讨论增进正义、无法为了更公正的社会而比较诸种替代性选择议案，它有的只是从一种先验进路跳跃到一个完美世界的乌托邦式议案。”与之相对，社会选择理论从各种现实具体的非正义问题——诸如，饥饿、贫穷、文盲、酷刑、种族不平等、妇女屈从、任意监禁、医疗保健政策中的排斥性不公等问题——及其解决入手，寻找并分析造成这些实际非正义后果的各种原因，选择并比较各种实际可能或可行的解决方案。这使得社会选择理论从一开始就具有新旧社会契约论都缺少的理论合理性和实际应用优势。
这六个方面的质疑不单关乎正义原则的推导，更关乎罗尔斯正义论的“狭隘地域主义”问题。在森看来，罗尔斯的正义论完全忽视了国际语境因而只具有“一国”或“某一社会”的内部适用性。坦诚地说，森对罗尔斯正义论的这些质疑看似温和，实则是颠覆性的，至少动摇了罗尔斯正义论的理论基础，揭橥了罗尔斯正义论的狭隘地域主义局限。这促使森进一步质疑罗尔斯著名的“基本善（品）清单”和罗尔斯晚年对“万民法”所做的最后也是最不成功的探究，并提出了自己的替代性方案。可见，对于罗尔斯的正义论来说，森的正义论更多的是批判性甚至革命性的，因而也是替代性的。正是通过深刻反省和批判罗尔斯的正义论，森才得以确立自己独特的“正义理念”，这便是他基于“可行能力进路”的正义论。与罗尔斯基于社会契约论所证成的制度正义论相对，森的正义论以个人可行能力为中心。“可行能力进路”（capability approach）是森的正义论的核心概念，其基本含义是指个人的“可行能力发挥”（capability functioning），“是个人做他或者她有理由珍视之事的能力”，具体包含个人实际可行的能力和该能力实际发挥的作用两个层面。早在1980年，森首次在其以“为了何种平等？”（Equality for what？）为题的“坦纳人类价值讲座”（Tanner Lecture on Human Value）中，提出“可行能力”概念，并在随后发表的著述中不断充实延展，以致后来成为森的政治经济学、伦理学和政治哲学中融会贯通的关键概念和基本方法。实际上，这一概念的提出首先是针对罗尔斯的“基本善（品）清单”的。以清单形式开列社会公共分配的基本善品是罗尔斯的一大发明，对现代实践哲学和诸多人文社会科学影响极大，甚至也影响到现代政府管理和公共政策操作。然而在森看来，与其开列一份“基本善（品）清单”，不如开出一份“基本可行能力清单”来得更为真实，更具实际可操作性。因为，公益善物的清单一如宴会餐桌上的食品清单，但对于社会大众来说，比宴会餐桌上诱人的食品清单更为实在的，是获得参加宴会的资格和食用餐桌上美味佳肴的能力，正如人们常说的那样，比机会均等的制度允诺更为实在的是人们获取并实现机会的能力。
森承认，任何一种政治哲学或伦理学都需要一种“信息焦点”（an informational focus）或前提预制，因此开列某种清单确乎必要。但对于关乎正义的政治哲学或伦理学来说，首要的工作不是从某种理想前提或终极结果出发来展开分析推导，因为选择从“社会实现的视角”（theperspectiveofsocialrealizations）出发，远比从“最终积累性结果”（the perspective of accumulating results）的视角出发更为优越可靠。实际上，人类社会的正义追寻总是从减少或消除现实生活中的非正义和反正义现象开始的。这种“社会实现的视角”所必需的，首先是对非正义现实的比较评估，是改变这些非正义现状的具体行动，以及，更为重要的是人们实施这种社会改变的行动能力。个人的可行能力直接关乎其自由（权利）及其实现之充分性程度，增强和改善个人的可行能力，意味着实现、增强或扩大个人的自由权利，这远比增加机会、财富和收入等社会资源更为根本，也更具社会正义的实践价值。罗尔斯一方面强调平等自由在两个正义原则中的优先排序，另一方面却又强调诸如机会、财富和收入这类社会资源的制度性再分配，实际上这两个方面不仅无法同时兼顾，而且极可能顾此失彼、“两头不讨好”。导致这种两难困境的原因至少有二：一是自由主义本身固有的自由与平等之间的张力；二是本文前所述的罗尔斯两个正义原则本身、它们之间，甚至正义第二原则的两个部分之间都存在难以消解的紧张。就前一方面而言，自由与平等二者所强调的价值取向并不能自然契合。自由所表达的是个人作为行动主体的权利自主与责任自律，而平等所表达的是不同权利主体之间的人格平等与平等相待；自由更强调个人主体的权利及其自主实现，而平等则更强调主体间平等相待的义务；两者之间相互制约。在资源有限的既定情形下，不同主体之间的自由权利及其实现，必定构成某种形式——或多或少、直接或间接——的竞争关系，既不可能是“利益无涉的”（disinterested），也不可能自然形成所谓“双赢”（win-win）或者同时兼顾、同等满足的理想局面。森认为，造成罗尔斯正义论的诸多难题的原因，除了社会契约论进路的局限、先验制度主义的前提误置、正义原则推导的六大疑问和“基本善（品）清单”的错开等原因之外，还由于其福利主义的社会偏好和地域主义的视域局限。“差异原则”的立意在于对社会少数弱者的制度援助，这诚然有利于彰显罗尔斯普遍道义论的政治伦理立场。然则，当罗尔斯将“财富和收入”列入其“基本善（品）清单”并强调基于财富和收入之重新分配的重要性时，人们难免对其制度分配或再分配的社会正义性产生警觉和怀疑，因为任何对财富和收入的制度性调节都可能面临一个难题：它不单关乎罗尔斯所说的基于代际正义考量的“储存正义”，更关乎同代人之间的“转移正义”和“校正正义”，也就是诺齐克反复追问罗尔斯的那个难以回答的问题：以何种方式且在何种程度上对人们的财富和收入进行制度调节才是正当合法的？正义社会并不必然等同于福利社会。
对于罗尔斯正义论的狭隘地域主义，大概是森最感不满的地方。罗尔斯反复申明其正义论仅限于“自由民主社会或民族”（liberal democracy societies or peoples），而对于其他“非自由非民主的社会或民族”（non-liberal and non-democracy states or peoples）则存而不论。这一学术立场受到包括他门下弟子（比如，Thomas Pogge）在内的许多学者的质疑。这促使晚年的罗尔斯有些仓促地写出《万民法》（The Law of Peoples，1999）一书，然而却并未改变其基本立场。在这部小册子中，罗尔斯似乎做出了些微让步，将其正义原则的推导从“自由民主社会”扩及“体面的协商等级制社会”（decent consultation hierarchy society），但仍不乐意谈论其他“非自由非民主的社会”。更有甚者，他甚至把世界上所有的社会或民族划分为五种类型——即：自由民主社会、体面的协商等级制社会、法外国家（outlaw states）、身负诸种不利条件之重负的社会（societies burdened by unfavorable conditions）和仁慈的专制主义社会（societies that are benevolent absolutisms），他不仅坚持认为差异原则不适合于“非自由非民主的社会”，而且主张当“法外国家”对“自由民主社会”构成直接威胁时，后者甚至可以采取“先发制人”的行动打击“法外国家”，以消除任何可能危及自由民主社会的危险境况。森认为，罗尔斯正义论的这种域外限制导致他无法达成其所标榜的普遍道义论目标，也不合乎自由主义的开放平等理想。森明确指出，民主并不仅仅是属于“西方的理智遗产”（Western intellectual inheritance），也是人类自古有之的国家治理经验或政治方式。我们不能把民主简单地等同于“选举+投票”，民主的本义是基于罗尔斯本人所说的“公共理性”的协商治理方式。因此，任何声称普遍合理的正义理论都不应该只限定于所谓“自由民主国家”区域，而应当放眼全球，寻求一种真正普适的全球正义。
纽斯鲍姆的学术著述极为丰富，在当代哲学同仁中鲜有比肩，其正义论的代表作是《正义的前沿：残障、国籍、物种成员资格》（Frontier of Justice, Disability Nationality, Species Membership，简称《正义的前沿》）。作者的特别献词也是“为了纪念罗尔斯”。从该书的献词和标题即见其理论抱负：作者开始正义主题探究的起点正是罗尔斯正义论探究的终点，其直接目标是扩展罗尔斯正义论的视域并将之推进到正义理论研究的“前沿”，也就是扩展到不单罗尔斯而且包括森在内的绝大多数正义论者都未（敢）涉猎的、包括不健全者和残障者、跨国籍全球公民、非人类生物在内的几乎所有生命主体。在《正义的前沿》的“致谢”中，纽斯鲍姆坦陈：“我的目标是对罗尔斯提出批判，我之所以单独挑出罗尔斯的理论进行批判，是因为他的理论是社会契约论传统中最强有力的政治理论，同时也因为他的理论是西方政治哲学传统中最杰出的理论。我关注的焦点是罗尔斯本人认为的、那些没有被解决的问题，这些问题对他的理论构成了挑战，他不确定他自己的理论能否解决这些问题。”“我的最终目标是扩展罗尔斯理论的核心观念，以解决这些新问题。尽管我认为，这种扩展肯定要在很大程度上改变他源自社会契约论传统的那部分理论，但我相信，当我们试图解决这些新的棘手的问题时，他这一理论本身、它的那些原则及其直觉性基础，都会给我们提供非常好的引导。我满怀尊敬、友爱和悲伤，谨将此书献给约翰·罗尔斯。”
与社会契约论的新老版本不同，“人类发展进路”吸取格劳修斯等人的自然法思想，既不从“民族—国家”及其由此衍生的“国家公民资格”（所谓“国籍”）出发，也不从“财富和收入”及其社会制度的再分配出发，而是回归最基本的人性理念，即人格尊严和人类社会性。“对格劳修斯而言，包括国家主权本身在内的国际社会中的所有权利资格，最终都源于人的尊严和社会性。”“人类发展进路”的基点和目标是世界和平和全人类的合作发展，这才是真正普遍的。如果说，全球正义确实需要达成某种形式的“全球契约”，那也不是基于“互利契约”，而毋宁是基于互助合作+和平发展的人类共识。纽斯鲍姆坦陈：“通过把国家的固定性设想为他的出发点，罗尔斯继而且实际上已然阻止了关于国家间的经济不平等和权力不平等的一切仔细思量。他已然从哲学上认可了世界上强大的国家——特别是美国——所做的任何事情：不管在人权问题、环境问题还是在经济政策问题上，不管是回应世界其他地方的处境还是回应国际协议和条约，它们都假装自己的体系是固定不变且具有决定性的，并且竭尽全力抵制关于在它们内部作改变的所有要求。”“然而，在现实世界中，我们看到了这种伎俩的真实面目：一种对严重问题漠不关心的、傲慢的思维方式，这种做法应受到谴责。人们不应在哲学上尊重它。”纽斯鲍姆清醒地意识到，我们现存的世界“并不是一个体面的、有着最低正义的世界，除非我们能保证这个世界上的所有人都能够在某种恰当的水平上具有那十种能力”。因此，关于全球正义的探讨“应当选择可行能力而不是富裕程度、效益或对个人的资源分配状况等，作为衡量的标准”。易言之，应当从每一个人的基本权利、尊严和平等开始，在此意义上，“可行能力进路与人类权利进路紧密相连”。我们可以怀疑，在一个多元文化/文明的世界上，不同国家和地区的人们有着不同的财富分配标准和评判标准，但却可以相信，无论身在何方，也无论我们有着怎样不同的生活条件、文化信念或宗教信仰，对于人类自由（权利）、尊严、平等合作，一定会有相同或相似的价值判断与价值确信，“毕竟，在当今的世界上，没有什么地方不流行人权观念、人类尊严和人类平等的观念以及平等合作的观念”。与之相类，我们不应且不能期待建构某种全球化的政治制度甚至是“世界国家”，任何企图建构这类政治制度或政治国家的尝试都是危险的。但我们可以合理期待，在保障每一个国家的国家主权之前提下，建立具有普适规范意义的“全球结构”（the global structure）或世界基本秩序。纽斯鲍姆由此推导出“全球结构的十大原则”：（1）避免国内政治永远无法逃避的“责任之过度决定”；（2）“应该在促进人类可行能力这一限制范围内尊重国家主权”；（3）富裕国家有责任将自己GDP（国内生产总值）中的相当一部分实质地赠送给贫穷国家；
基于上述分析，忠诚义务模式与公益义务模式各有利弊。总体上看，忠诚义务模式以美国为代表，而公益义务模式以德国为代表。但是，极端地强调辩护律师只维护当事人的利益或只维护公共利益的律师职业伦理在任何国家都是不存在的。无论是在美国还是在德国，均强调律师同时具有维护当事人利益与公共利益的义务，只不过是有所侧重，当存在利益冲突时，更加强调维护哪一方的利益而已。尤其是，当前世界范围内，当事人主义诉讼模式与职权主义诉讼模式相互融合已发展为一种趋势，在此影响下辩护律师职业伦理的两种模式也随之有相互融合的发展迹象。美国的辩护律师职业伦理中存在一种“法庭职员理论”。根据这一理论，律师是法庭的职员（officer of the court），律师是司法制度的重要一环，并且承担着重要的职责。该理论类似于德国关于律师是“独立司法机关”的观点，强调律师对于法庭负有职责。在该理论的影响下，律师的真实义务逐渐被重视起来，不仅仅是对抗制中的角色。辩护律师既要忠实于当事人，对于法庭也负有一定的真实义务。同样，在德国，对过于强调公共利益的辩护律师职业伦理理论的反思，促进了律师作为“限制的机关”理论、“一方利益代理人”理论以及“契约理论”的发展。这些理论的共同点在于一致强调辩护律师应当加强当事人的利益维护，在公共利益方面有所限制。
相对贫困的测量，一般以相对贫困线为标准。而相对贫困线的制定方法主要有以下四种：第一种是预算标准法，即由专家所研究的贫困群体的代表根据社会认可的生活水平制定的收入贫困线；第二种是社会指标法，即通过计算群体成员的剥夺程度、依据收入和剥夺程度的关系来计算贫困线；第三种是ELE法（extended linear expenditure system），即以拓展线性支出系统为理论基础制定的贫困线；第四种是收入法，即以社会收入集中趋势的一定比例作为相对贫困线，如均值和中位数，比如世界银行认为只要是低于平均收入1/3的社会成员即可视为相对贫困人口，欧盟则将收入水平位于中位收入60%之下的人口归入相对贫困人口。
在权利的概念分析上主要有两种理论，一是意志论，二是利益论。这两种理论既是关于权利之性质的概念分析，同时又指向了辩护权利的基本理据。权利的利益论和意志论反映了更为基础的道德分歧，比如意志论强调了自觉和自主性的重要性，利益论中的利益则被用来辩护某种主张成为权利的基础。剑桥大学的法哲学家克莱默(Matthew H. Kramer)对权利的利益论和意志论的基本观点作了如下总结：“对于利益论来说，一项权利的本质就在于对权利人某些方面之福祉的规范性保护。相反，对于意志论来说，一项权利的本质就在于权利人在规范性上做出重要选择的诸多机会，而这些选择涉及其他人的行为。”据此，利益构成了权利存在的一个必要条件，尽管不是充分条件。这表明利益是权利的概念性组成部分，而且尤为重要的是，利益是外在于我们对权利人本身的理解或界定的，即利益论诉诸一个外在于权利人自身(尽管和权利人相关)的因素来界定权利的本质。意志论反对利益是构成权利之存在的必要条件，遑论充分条件。权利人的能力和许可才是必要的或充分的条件，因为这两个因素都和权利人本身的某种性质相关。而在利益论者看来，这两个因素既非必要也非充分，因为他们对权利性质的理解已经不再受权利人自身之性质的限制。
宗教史提供了一个很好的例证。中世纪晚期的教派和新教之间存在着一个重大差别，它的重要性大于在那些被认为是宗教改革之先兆的学说中发现的相似之处，它也足以说明为什么后者和其他改革相比具有如此强大的生命力。威克里夫和胡斯仅反对天主教教义的某些细枝末节，而路德则抛弃教会的权威，赋予个体良知一种独立性，它必然使人持续不断地反抗。同样，在尼德兰革命、英国革命、美国独立战争或布拉班特起义(therising of Brabant)与法国大革命之间，也有类似的差别。
法国大革命的同情者并不支持在滑铁卢之役中胜利的那派势力。因为他们已经懂得把他们的学说和法国的事业联系在一起了。在英国的荷兰王室辉格党人(TheHollandHouse Whigs)、西班牙的亲法分子、意大利的缪拉党人(theMumtists)以及莱茵联盟(the Confederation of Rhine)的支持者们，将他们的爱国主义融化在他们的革命激情中，为法国势力的衰落感到惋惜。他们惊恐地看着解放战争(theWar of Deliverance)所产生的陌生的新势力，因为它们既威胁着法国的统治，也威胁着法国的自由主义。
CONTENTS 1 Introduction 2 GPT Diffusion Theory 3 The First Industrial Revolution and Britain’s Rise 4 The Second Industrial Revolution and America’s Ascent 5 Japan’s Challenge in the Third Industrial Revolution 6 A Statistical Analysis of Software Engineering Skill Infrastructure and Computerization 7 US-China Competition in AI and the Fourth Industrial Revolution 8 Conclusion
1 Introduction
IN JULY 2018, the BRICS nations (Brazil, Russia, India, China, and South Africa) convened in Johannesburg around a specific, noteworthy theme: “Collaboration for Inclusive Growth and Shared Prosperity in the Fourth Industrial Revolution.” The theme was noteworthy in part because of its specificity. Previous iterations of the BRICS summit, which gathers five nations that account for about 40 percent of the world’s population and 25 percent of the world’s GDP,1 had tackled fuzzy slogans such as “Stronger Partnership for a Brighter Future” and “Broad Vision, Shared Prosperity.” What stood out not only about that year’s theme but also in comments by BRICS leaders at the summit was an unambiguous conviction that the world was undergoing a momentous season of technological change—one warranting the title “Fourth Industrial Revolution.”2
Throughout the gathering, leaders of these five major emerging economies declared that the ongoing technological transition represented a rare opportunity for accelerating economic growth. When Chinese president Xi Jinping addressed the four other leaders of major emerging economies, he laid out the historical stakes of that belief:
From the mechanization of the first industrial revolution in the 18th century, to the electrification of the second industrial revolution in the 19th century, to the informatization of the third industrial revolution in the 20th century, rounds of disruptive technological innovation have … fundamentally changed the development trajectory of human history.3
Citing recent breakthroughs in cutting-edge technologies like artificial intelligence (AI), Xi proclaimed, “Today, we are experiencing a larger and deeper round of technological revolution and industrial transformation.”4
While the BRICS summit did not explicitly address how the Fourth Industrial Revolution could reshape the international economic order, the implications of Xi’s remarks loomed in the backdrop. In the following months, Chinese analysts and scholars expanded upon them, especially the connection he drew between technological disruption and global leadership transitions.5 One commentary on Xi’s speech, published on the website of the authoritative Chinese Communist Party publication Study Times, detailed the geopolitical consequences of past technological revolutions: “Britain seized the opportunity of the first industrial revolution and established a world-leading productivity advantage.… After the second industrial revolution, the United States seized the dominance of advanced productivity from Britain.”6 In his analysis of Xi’s address, Professor Jin Canrong of Renmin University, an influential Chinese international relations scholar, argued that China has a better chance than the United States of winning the competition over the Fourth Industrial Revolution.7
This broad sketch of power transition by way of technological revolution also resonates with US policymakers and leading thinkers. In his first press conference after taking office, President Joe Biden underscored the need to “own the future” as it relates to competition in emerging technologies, pledging that China’s goal to become “the most powerful country in the world” was “not going to happen on [his] watch.”8 In 2018, the US Congress stood up the National Security Commission on Artificial Intelligence (NSCAI), an influential body that convened leading government officials, technology experts, and social scientists to study the national security implications of AI. Comparing AI’s possible impact to past technologies like electricity, the NSCAI’s 756-page final report warned that the United States would soon lose its technological leadership to China if it did not adequately prepare for the “AI revolution.”9
Caught up in the latest technical advances coming out of Silicon Valley or Beijing’s Zhongguancun, these sweeping narratives disregard the process by which emerging technologies can influence a power transition. How do technological revolutions affect the rise and fall of great powers? Is there a discernible pattern that characterizes how previous industrial revolutions shaped the global balance of power? If such a pattern exists, how would it inform our understanding of the Fourth Industrial Revolution and US-China technological competition?
Conventional Wisdom on Technology-Driven Power Transitions
International relations scholars have long observed the link between disruptive technological breakthroughs and the rise and fall of great powers.10 At a general level, as Yale historian Paul Kennedy has established, this process involves “differentials in growth rates and technological change, leading to shifts in the global economic balances, which in turn gradually impinge upon the political and military balances.”11 Yet, as is the case with present-day speculation about the effects of new technologies on the US-China power balance, largely missing from the international relations literature is an explanation of how technological change creates the conditions for a great power to leapfrog its rival. Scholars have carefully scrutinized how shifts in economic balances affect global military power and political leadership, but there is a need for further investigation into the very first step of Kennedy’s causal chain: the link between technological change and differentials in long-term growth rates among great powers.12
Among studies that do examine the mechanics of how technological change shapes economic power transitions, the standard explanation stresses dominance over critical technological innovations in new, fast-growing industries (“leading sectors”). Britain became the world’s most productive economy, according to this logic, because it was home to new advances that transformed its burgeoning textile industry, such as James Hargreaves’s spinning jenny. In the same vein, Germany’s mastery of major breakthroughs in the chemical industry is seen as pivotal to its subsequent challenge to British economic leadership. Informed by historical analysis, the leading-sector (LS) perspective posits that, during major technological shifts, the global balance of economic power tips toward “the states which were the first to introduce the most important innovations.”13
Why do the benefits of leading sectors accrue to certain countries? Explanations vary, but most stress the goodness-of-fit between a nation’s domestic institutions and the demands of disruptive technologies. At a general level, some scholars argue that rising powers quickly adapt to new leading sectors because they are unburdened by the vested interests that have built up in more established powers.14 Others point to more specific factors, including the degree of government centralization or sectoral governance arrangements.15 Common to all these perspectives is a focus on the institutions that allow one country to first introduce major breakthroughs in an emerging industry. In the case of Britain’s rise, for example, many influential histories highlight institutions that supported “heroic” inventors.16 Likewise, accounts of Germany’s success with leading sectors focus on its investments in scientific education and industrial research laboratories.17
The broad outlines of LS theory exert substantial influence in academic and policymaking circles. Field-defining texts, including works by Robert Gilpin and Paul Kennedy, use the LS model to map out the rise and fall of great powers.18 In a review of international relations scholarship, Daniel Drezner summarizes their conclusions: “Historically, a great power has acquired hegemon status through a near-monopoly on innovation in leading sectors.”19
The LS template also informs contemporary discussion of China’s challenge to US technological leadership. In another speech about how China could leverage this new round of industrial revolution to become a “science and technology superpower,” President Xi called for China to develop into “the world’s primary center for science and high ground for innovation.”20 As US policymakers confront China’s growing strength in emerging technologies like AI, they also frame the competition in terms of which country will be able to generate radical advances in new leading sectors.21
Who did it first? Which country innovated it first? Presented with technical breakthroughs that inspire astonishment, it is only natural to gravitate toward the moment of initial discovery. When today’s leaders evoke past industrial revolutions, as Xi did in his speech to the BRICS nations, they tap into historical accounts of technological progress that also center the moment of innovation.22 The economist and historian Nathan Rosenberg diagnoses the problem with these innovation-centric perspectives: “Much less attention … if any at all, has been accorded to the rate at which new technologies have been adopted and embedded in the productive process. Indeed the diffusion process has often been assumed out of existence.”23 Yet, without the humble undertaking of diffusion, even the most extraordinary advances will not matter.
Taking diffusion seriously leads to a different explanation for how technological revolutions affect the rise and fall of great powers. A diffusion-centric framework probes what comes after the hype. Less concerned with which state first introduced major innovations, it instead asks why some states were more successful at adapting and embracing new technologies at scale. As outlined in the next section, this alternative pathway points toward a different set of institutional factors that underpin leadership in times of technological leadership, in particular institutions that widen the base of engineering skills and knowledge linked to foundational technologies.
GPT Diffusion Theory
In September 2020, the Guardian published an opinion piece arguing that humans should not fear new breakthroughs in AI. Noting that “Stephen Hawking has warned that AI could ‘spell the end of the human race,’ ” the article’s “author” contends that “I am here to convince you not to worry. Artificial intelligence will not destroy humans. Believe me.”24 If one came away from this piece with the feeling that the author had a rose-tinted view of the future of AI, it would be a perfectly reasonable judgment. After all, the author was GPT-3, an AI model that can understand and produce humanlike text.
Released earlier that year by OpenAI, a San Francisco–based AI lab, GPT-3 surprised everyone—including its designers—with its versatility. In addition to generating poetry and essays like the Guardian op-ed from scratch, early users demonstrated GPT-3’s impressive capabilities in writing code, translating languages, and building chatbots.25 Six months after its launch, one compilation listed sixty-six unique use cases of GPT-3, which ranged from automatically updating spreadsheets to generating website landing pages.26 Two years later, OpenAI’s acclaimed ChatGPT model, built on an improved version of GPT-3, would set the internet aflame with its wide-ranging capabilities.27
While the name “GPT-3” derives from a class of language models known as “generative pre-trained transformers,” the abbreviation, coincidentally, also speaks to the broader significance of recent breakthroughs in AI: the possible arrival of the next general-purpose technology (GPT). Foundational breakthroughs in the ability of computers to perform tasks that usually require human intelligence have the potential to transform countless industries. Hence, scholars and policymakers often compare advances in AI to electricity, the prototypical GPT.28 As Kevin Kelly, the former editor of WIRED, once put it, “Everything that we formerly electrified we will now cognitize … business plans of the next 10,000 startups are easy to forecast: Take X and add AI.”29
In this book, I argue that patterns in how GPTs diffuse throughout the economy illuminate a novel explanation for how and when technological changes affect power transitions. The emergence of GPTs—fundamental advances that can transform many application sectors—provides an opening for major shifts in economic leadership. Characterized by their scope for continuous improvement, pervasive applicability across the economy, and synergies with other technological advances, GPTs carry an immense potential for boosting productivity.30 Carefully tracking how the various applications of GPTs are adopted across various industries, a process I refer to as “GPT diffusion,” is essential to understanding how technological revolutions disrupt economic power balances.
Based on the experience of past GPTs, this potential productivity boost comes with one notable caveat: the full impact of a GPT manifests only after a gradual process of diffusion into pervasive use.31 GPTs demand structural changes across a range of technology systems, which involve complementary innovations, organizational adaptations, and workforce adjustments.32 For example, electrification’s boost to productivity materialized about five decades after the introduction of the first electric dynamo, occurring only after factories had restructured their layouts and there had been interrelated breakthroughs in steam turbines.33 Fittingly, after the release of GPT-3, OpenAI CEO Sam Altman alluded to this extended trajectory: “The GPT-3 hype is way too much … it still has serious weaknesses and sometimes makes very silly mistakes. AI is going to change the world, but GPT-3 is just a very early glimpse. We have a lot still to figure out.”34
Informed by historical patterns of GPT diffusion, my explanation for technology-driven power transitions diverges significantly from the standard LS account. Specifically, these two causal mechanisms differ along three key dimensions, which relate to the technological revolution’s impact timeframe, phase of relative advantage, and breadth of growth. First, while the GPT mechanism involves a protracted gestation period between a GPT’s emergence and resulting productivity boosts, the LS mechanism assumes that there is only a brief window during which countries can capture profits in leading sectors. “The greatest marginal stimulation to growth may therefore come early in the sector’s development at the time when the sector itself is expanding rapidly,” William Thompson reasons.35 By contrast, the most pronounced effects on growth arrive late in a GPT’s development.
Second, the GPT and LS mechanisms also assign disparate weights to innovation and diffusion. Technological change involves a phase when the technology is first incubated as a viable commercial application (“innovation”) and a phase when the innovation permeates across a population of potential users (“diffusion”). The LS mechanism is primarily concerned about which country dominates innovation in leading sectors, capturing the accompanying monopoly profits.36 Under the GPT mechanism, successful adaptation to technological revolutions is less about being the first to introduce major innovations and more about effectively adopting GPTs across a wide range of economic sectors.
Third, regarding the breadth of technological transformation and economic growth, the LS mechanism focuses on the contributions of a limited number of leading sectors and new industries to economic growth in a particular period.37 In contrast, GPT-fueled productivity growth is spread across a broad range of industries.38 Dispersed productivity increases from many industries and sectors come from the extension and generalization of localized advances in GPTs.39 Thus, the LS mechanism expects the breadth of growth in a particular period to be concentrated in leading sectors, whereas the GPT mechanism expects technological complementarities to be dispersed across many sectors.
A clearer understanding of the contours of technological change in times of economic power transition informs which institutional variables matter most. If the LS trajectory holds, then the most important institutional endowments and responses are those that support a monopoly on innovation in leading sectors. In the context of skill formation, institutional competencies in science and basic research gain priority. For instance, the conventional explanation of Germany’s industrial rise in the late nineteenth century attributes its technological leadership to investments in industrial research labs and highly skilled chemists. These supported Germany’s dominance of the chemical industry, a key LS of the period.40
The impact pathway of GPTs brings another set of institutional complementarities to the fore. GPT diffusion theory highlights the importance of “GPT skill infrastructure”: education and training systems that widen the pool of engineering skills and knowledge linked to a GPT. When widespread adoption of GPTs is the priority, it is ordinary engineers, not heroic inventors, who matter most. Widening the base of engineering skills associated with a GPT cultivates a more interconnected technological system, spurring cross-fertilization between institutions optimized for applied technology and those oriented toward foundational research.41
Returning to the example of late-nineteenth-century advances in chemicals, GPT diffusion spotlights institutional adjustments that differ from those of the LS mechanism. In a decades-long process, innovations in chemical engineering practices gradually enabled the chemicalization of procedures common to many industries beyond synthetic dyes, which was controlled by Germany. Despite trailing Germany in the capacity to produce elite chemists and frontier chemical research, the United States was more effective at adapting to chemicalization because it first institutionalized the discipline of chemical engineering.42
Of course, since GPT diffusion depends on factors aside from human capital, GPT skill infrastructure represents one of many institutional forces at work. Standards-setting organizations, financing bodies, and the competitiveness of markets can all influence the flow of information between the GPT domain and application sectors.43 Since institutions of skill formation produce impacts that spill over into and complement other institutional arrangements, they comprise the focus of my analysis.44
Assessing GPT Diffusion across Industrial Revolutions
To test this argument, I employ a mixed-methods approach that pairs qualitative historical analysis with quantitative methods. Historical case studies permit me to thoroughly trace the interactions between technologies and institutions among great powers in previous industrial revolutions. I then explore the generalizability of GPT diffusion theory beyond the chosen set of great powers. Using data on nineteen countries from 1995 to 2020, I analyze the theorized connection between GPT skill infrastructure in software engineering and computerization rates.
To investigate the causal processes that connect technological changes to economic power transitions, I set the LS mechanism against the GPT diffusion mechanism across three historical case studies: Britain’s rise to preeminence in the First Industrial Revolution (IR-1); America’s and Germany’s overtaking of Britain in the Second Industrial Revolution (IR-2); and Japan’s challenge to America’s technological dominance in the Third Industrial Revolution (IR-3), or what is sometimes called the “information revolution.” This case setup allows for a fair and decisive assessment of the explanatory relevance of GPT diffusion theory in comparison to LS theory. Because the IR-1 and IR-2 function as typical cases where the cause and outcome are clearly present, they are ideal for developing and testing mechanism-based theories.45 The IR-3, a deviant case in that a technological revolution is not followed by an economic power transition, provides a different but still useful way to compare the two mechanisms.
The IR-1 (1780–1840) is a paradigmatic case of technology-driven power transition. It is well established that the IR-1’s technological advances propelled Great Britain to unrivaled economic supremacy. As for the specific causal pathway, international relations scholarship tends to attribute Britain’s rise to its monopoly over innovation in cotton textiles and other leading sectors. According to these accounts, Britain’s technological leadership in the IR-1 sprang from its institutional capacity to nurture genius inventors in these sectors. Since the publication of these field-defining works, economic and technology historians have uncovered that the impacts on British industrialization of the two most prominent areas of technological change, cotton textiles and iron, followed different trajectories. Often relying on formal econometric methods to understand the impact of key technologies, these historical accounts question the prevailing narrative of the IR-1.
The IR-2 (1870–1914) supplies another opportunity to pit GPT diffusion theory against the LS account. International relations scholars interpret the IR-2 as a case in which Britain’s rivals challenged its economic leadership because they first introduced significant technological advances in leading sectors. Particular emphasis is placed on Germany’s ability to corner market shares in chemicals, which is linked to its strengths in scientific education and industrial research institutions. More granular data on cross-national differences in engineering education suggest that the U.S. technological advantage rested on the country’s wide base of mechanical engineers. Combined with detailed tracing of the pace and extent of technology adoption during this period, this chapter’s evidence suggests modifications to conventional understandings of the IR-2.
In the IR-3 (1960–2000), fundamental breakthroughs in information and communication technologies presented another opening for a shift in economic leadership. During this period, prominent thinkers warned that Japan’s lead in industries experiencing rapid technological change, including semiconductors and consumer electronics, would threaten U.S. economic leadership. Influential scholars and policymakers advocated for the United States to adopt Japan’s keiretsu system of industrial organization and its aggressive industrial policy approach. Ultimately, Japan’s productivity growth stalled in the 1990s. Given the absence of an economic power transition, the primary function of the IR-3 case therefore is to provide disconfirming evidence of the two explanations. If the components of the LS mechanism were present, then the fact that an economic power transition did not occur would damage the credibility of the LS mechanism. The same condition applies to GPT diffusion theory.
In each of the cases, I follow the same standardized procedures. First, I test three pairs of competing propositions about the key technological trajectories, derived from the different expectations of the LS and GPT mechanisms related to the impact timeframe, phase of relative advantage, and breadth of growth. Then, depending on whether the LS or GPT trajectory better accords with the historical evidence, I analyze the goodness-of-fit between the institutional competencies of leading industrial powers and the prevailing trajectory. For instance, if an industrial revolution is better characterized by the GPT trajectory, then the corresponding case analysis should show that differences in GPT skill infrastructure determine which powers rise and fall. Although I primarily distinguish GPT diffusion theory from the LS model, I also examine alternative factors unique to the particular case, as well as two other prominent explanations of how advanced economies differentially benefit from technological changes (the varieties of capitalism and threat-based approaches).
The historical case analysis supports the explanatory power of the GPT mechanism over the LS mechanism. In all three periods, technological changes affected the rise and fall of great powers in a gradual, decades-long impact pathway that advantaged those that effectively diffused GPTs across a broad range of sectors. Education and training systems that cultivated broad pools of engineering skills proved crucial to GPT diffusion.
Evaluating these two competing explanations requires a clear understanding of the cause and outcome that bracket both the GPT and LS mechanisms. The hypothesized cause is a “technological revolution,” or a period characterized by particularly disruptive technological advances.46 Since the shape of technological change is uneven, not all improvements in useful knowledge are relevant for power transitions.47 However, some extraordinary clusters of technological breakthroughs, often deemed industrial revolutions by historians, do have ramifications for the rise and fall of great powers.48 I am primarily interested in the pathway by which these technological revolutions influence the global distribution of power.
The outcome variable of interest is an economic power transition, in which one great power sustains productivity growth at higher levels than its rivals. The balance of power can shift in many ways; here I focus on relative economic growth rates because they are catalysts for intensifying hegemonic rivalries.49 Productivity growth, in particular, determines economic growth over the long run. Unique in its fungibility with other forms of power, sustained economic growth is central to a state’s ability to exert political and military influence. As demonstrated by the outcomes of interstate conflicts between great powers, economic and productive capacity is the foundation of military power.50
Lastly, the quantitative analysis supplements the historical case studies by scrutinizing the generalizability of GPT diffusion theory outside of great powers. A key observable implication of my argument is that the rate at which a GPT spreads throughout the economy owes much to that country’s institutional capacity to widen the pool of pertinent engineering skills and knowledge. Using a novel method to estimate the breadth of software engineering education at a cross-national level, I analyze the theorized connection between GPT skill infrastructure and computerization rates across nineteen advanced and emerging economies from 1995 to 2020. I supplement my time-series cross-sectional models with a duration analysis and cross-sectional regressions. Robust to many alternative specifications, my results show that, at least for computing technology, advanced economies that have higher levels of GPT skill infrastructure preside over higher rates of GPT diffusion.
Key Contributions
The book makes several contributions to scholarship on power transitions and the effects of technological change on international politics. First, it puts forward a novel explanation for how and when significant technological breakthroughs generate a power transition in the international system. GPT diffusion theory revises the dominant theory based on leading sectors, which holds significant sway over academic and policymaking circles. By deepening our understanding of how technological revolutions influence shifts in economic leadership, this book also contributes to long-standing debates about the causes of power transitions.51
Second, the findings of this book bear directly on present-day technological competition between the United States and China. Emphasizing where fundamental breakthroughs are first seeded, the LS template strongly informs not only assessments of the US-China competition for technological leadership but also the ways in which leading policymakers in both countries formulate technology strategies. It is no coincidence that the three cases in this study match the three technological revolutions referenced by Chinese president Xi in his speech on the IR-4 to the BRICS summit.
As chapter 7 explores in detail, GPT diffusion theory suggests that Xi, along with other leading policymakers and thinkers in both the United States and China, has learned the wrong lessons from previous industrial revolutions. If the IR-4 reshapes the economic power balance, the impact will materialize through a protracted period during which a GPT, such as AI, acquires a variety of uses in a wide range of productive processes. GPT skill infrastructure, not the flashy efforts to secure the high ground in innovation, will decide which nation owns the future in the IR-4.
Beyond power transitions, Technology and the Rise of Great Powers serves as a template for studying the politics of emerging technologies. An enduring dilemma is that scholars either assign too much weight to technological change or underestimate the effects of new technologies.52 Approaches that emphasize the social shaping of technology neglect that not all technologies are created equal, whereas technologically deterministic approaches discount the influence of political factors on technological development. By first distinguishing GPTs, together with their pattern of diffusion, from other technologies and technological trajectories, and then showing how social and political factors shape the pace and direction of GPT diffusion, my approach demonstrates a middle way forward.
Roadmap for the Book
The book proceeds as follows. Chapter 2 fleshes out the key differences between GPT diffusion theory and the LS-based account, as well as the case analysis procedures and selection strategy that allow me to systematically evaluate these two causal mechanisms. The bulk of the evidence follows in three case studies that trace how technological progress affected economic power transitions in the First, Second, and Third Industrial Revolutions.
The first two case studies, the IR-1 and IR-2, show that a gap in the adoption of GPTs, as opposed to monopoly profits from dominating LS innovations, was the crucial driver of an economic power transition. In both cases, the country that outpaced its industrial rivals made institutional adjustments to cultivate engineering skills related to the key GPT of the period. The IR-1 case, discussed in chapter 3, reveals that Britain was the most successful in fostering a wide pool of machinists who enabled the widespread diffusion of advances in iron machinery. In considering the IR-2 case, chapter 4 highlights how the United States surpassed Britain as the preeminent economic power by fostering a wide base of mechanical engineering talent to spread interchangeable manufacturing methods.
The IR-3 case, presented in chapter 5, demonstrates that technological revolutions do not necessarily always produce an economic power transition. The fact that Japan did not overtake the United States as the economic leader would provide disconfirming evidence of both the LS and GPT mechanisms, if the components of these mechanisms were present. In the case of the LS mechanism, Japan did dominate innovations in the IR-3’s leading sectors, including consumer electronics and semiconductor components. In contrast, the IR-3 does not discredit the GPT mechanism because Japan did not lead the United States in the diffusion of information and communications technology across a wide variety of economic sectors.
Chapter 6 uses large-n quantitative analysis to explore how GPT diffusion applies beyond great powers. Chapter 7 applies the GPT diffusion framework to the implications of modern technological breakthroughs for the US-China power balance. Focusing on AI technology as the next GPT that could transform the international balance of power, I explore the extent to which my findings generalize to the contemporary US-China case. I conclude in chapter 8 by underscoring the broader ramifications of the book.
2 GPT Diffusion Theory
HOW AND WHEN do technological changes affect the rise and fall of great powers? Specifically, how do significant technological breakthroughs result in differential rates of economic growth among great powers? International relations scholars have long observed that rounds of technological revolution lead to upheaval in global economic leadership, bringing about a power transition in the international system. However, few studies explore how this process occurs.
Those that do tend to fixate on the most dramatic aspects of technological change—the eureka moments and first implementations of radical inventions. Consequently, the standard account of technology-driven power transitions stresses a country’s ability to dominate innovation in leading sectors. By exploiting brief windows in which to monopolize profits in new industries, the country that dominates innovation in these sectors rises to become the world’s most productive economy. Explanations vary regarding why the benefits of leading sectors tend to accrue in certain nations. Some scholars argue that national systems of political economy that accommodate rising challengers can more readily accept and support new industries. Leading economies, by contrast, are victims of their past success, burdened by powerful vested interests that resist adaptation to disruptive technologies.1 Other studies point to more specific institutional factors that account for why some countries monopolize leading sectors, such as the degree of government centralization or industrial governance structures.2
An alternative explanation, based on the diffusion of general-purpose technologies (GPTs), draws attention to the less spectacular process by which fundamental innovations gradually diffuse throughout many industries. The rate and scope of diffusion is particularly relevant for GPTs, which are distinguished by their scope for continual improvement, broad applicability across many sectors, and synergies with other technological advances. Recognized by economists and historians as “engines of growth,” GPTs hold immense potential for boosting productivity.3 Realizing this promise, however, necessitates major structural changes across the technology systems linked to the GPT, including complementary innovations, organizational changes, and an upgrading of technical skills. Thus, GPTs lead to a productivity boost only after a “gradual and protracted process of diffusion into widespread use.”4 This is why more than five decades passed before key innovations in electricity, the quintessential GPT, significantly transformed manufacturing productivity.5
The process of GPT diffusion illuminates a pathway from technological change to power transition that diverges from the LS account (figure 2.1). Under the GPT mechanism, some great powers sustain economic growth at higher levels than their rivals do because, during a gradual process spanning decades, they more intensively adopt GPTs across a broad range of industries. This is analogous to a marathon run on a wide road. The LS mechanism, in contrast, specifies that one great power rises to economic leadership because it dominates innovations in a limited set of leading sectors and captures the accompanying monopoly profits. This is more like a sprint through a narrow running lane.
Why are some countries more successful at GPT diffusion? Building from scholarship arguing that a nation’s success in adapting to emerging technologies is determined by the fit between its institutions and the demands of evolving technologies, I argue that the GPT diffusion pathway informs the institutional adaptations crucial to success in technological revolutions.6 Unlike institutions oriented toward cornering profits in leading sectors, those optimized for GPT diffusion help standardize and spread novel best practices between the GPT sector and application sectors. Education and training systems that widen the base of engineering skills and knowledge linked to new GPTs, or what I call “GPT skill infrastructure,” are essential to all of these institutions.
FIGURE 2.1. Causal Diagrams for LS and GPT Mechanisms
The differences between these two theories of technological change and power transition are made clear when one country excels in the institutional competencies for LS product cycles but does not dominate GPT diffusion. Take, for example, chemical innovations and Germany’s economic rise in the late nineteenth century. Germany dominated major innovations in chemicals and captured nearly 90 percent of all global exports of synthetic dyestuffs.7 In line with the LS mechanism, this success was backed by Germany’s investments in building R&D labs and training doctoral students in chemistry, as well as a system of industrial organization that facilitated the rise of three chemical giants.8 Yet it was the United States that held an advantage in adopting basic chemical processes across many industries. As expected by GPT diffusion theory, the United States held institutional advantages in widening the base of engineering skills and knowledge necessary for chemicalization on a wide scale.9 This is when the ordinary tweakers and the implementers come to the fore, and the star scientists and inventors recede to the background.10
The rest of this chapter fleshes out my theoretical framework. It first clarifies the outcome I seek to explain: an economic power transition in which one great power becomes the economic leader by sustaining productivity growth at higher levels than its rivals. The starting point of my argument is that the diffusion of GPTs is central to the relationship between technological change and productivity leadership. This chapter explicates this argument by justifying the emphasis on both GPTs and diffusion, highlighting the differences between the GPT and LS mechanisms. It then extends the analysis to the institutional competencies that synergize with GPT trajectories. From the rich set of technology-institution interactions identified by evolutionary economists and comparative institutionalists, I justify my focus on institutions that enable countries to widen the skill base required to spread GPTs across industries. After differentiating my argument from alternative explanations, the chapter closes with a description of the book’s research methodology.11
The Outcome: Long-Term Economic Growth Differentials and Power Transitions
Power transitions are to the international system as earthquakes are to the geological landscape. Shifts in the relative power of leading nations send shock waves throughout the international system. What often follows is conflict, the most devastating form of which is a war waged by coalitions of great powers for hegemony over the globe.12 Beyond heightened risks of conflict, the aftershocks of power transitions reverberate in the architecture of the international order as victorious powers remake international institutions in their own images.13
While the power transition literature largely tackles the consequences of power transitions, I treat the rise and fall of great powers as the outcome to be explained. This follows David Baldwin’s instruction for international relations scholars “to devote more attention to treating power as a dependent variable and less to treating it as an independent variable.”14 Specifically, I explore the causes of “economic power transitions,” in which one great power sustains economic growth rates at higher levels than its rivals.15
It might not be obvious, at first glance, why I focus on economic power. After all, power is a multidimensional, contested concept that comes in many other forms. The salience of certain power resources depends on the context in which a country draws upon them to exert influence.16 For my purposes, differentials in economic growth are the most relevant considerations for intensifying hegemonic rivalry. An extensive literature has demonstrated that changes in relative economic growth often precede hegemonic wars.17
Moreover, changes in global political and military leadership often follow shifts in economic leadership. As the most fungible mode of power, economic strength undergirds a nation’s influence in global politics and its military capabilities.18 The outcomes of interstate conflicts bear out that economic and productive capacity is the foundation of military power.19 Paul Kennedy concludes that
all of the major shifts in the world’s military-power balances have followed alterations in the productive balances … the rising and falling of the various empires and states in the international system has been confirmed by the outcomes of the major Great Power wars, where victory has always gone to the side with the greatest material resources.20
How does one identify if or when an economic power transition occurs? Phrased differently, how many years does a great power need to lead its rivals in economic growth rates? How large does that gap have to reach? Throughout this book, I judge whether an economic power transition has occurred based on one great power attaining a lead in overall economic productivity over its rivals by sustaining higher levels of productivity growth rates.21 Productivity growth ensures that efficient and sustainable processes are fueling growth in total economic output. Additionally, productivity is the most important determinant of economic growth in the long run, which is the appropriate time horizon for understanding power transitions. “Productivity isn’t everything, but in the long run it is almost everything,” states Nobel Prize–winning economist Paul Krugman.22
Alternative conceptualizations of economic power cannot capture how effectively a country translates technological advance into national economic growth. Theories of geo-economics, for instance, highlight a state’s balance of trade in certain technologically advanced industries.23 Other studies emphasize a state’s share of world-leading firms.24 National rates of innovation, while more inclusive, measure the generation of novel technologies but not diffusion across commercial applications, thereby neglecting the ultimate impact of technological change.25 Compared to these indicators, which account for only a small portion of the value-added activities in the economy, productivity provides a more comprehensive measure of economic leadership.26
This focus on productivity is supported by recent work on power measurement, which questions measures of power resources based on economic size. Without accounting for economic efficiency, solely relying on measures of gross economic and industrial output provides a distorted view of the balance of power, particularly where one side is populous but poor.27 If national power was measured by GDP alone, China was the world’s most powerful country during the first industrial revolution. However, China’s economy was far from the productivity frontier. In fact, as the introduction chapter spotlighted, the view that China fell behind the West because it could not capitalize on productivity-boosting technological breakthroughs is firmly entrenched in the minds of leading Chinese policymakers and thinkers.
Lastly, it is important to clarify that I limit my analysis of productivity differentials to great powers.28 In some measures of productivity, other countries may rank highly or even outrank the countries I study in my cases. In the current period, Switzerland and other countries have higher GDP per capita than the United States; before World War I, Australia was the world leader in productivity, as measured by GDP per labor-hour.29 However, smaller powers like pre–World War I Australia and present-day Switzerland are excluded from my study of economic power transitions, as they lack the baseline economic and population size to be great powers.30
There is no exact line that distinguishes great powers from other countries.31 Kennedy’s seminal text The Rise and Fall of the Great Powers, for instance, has been challenged for not providing a precise definition of great power.32 Fortunately, across all the case studies in this book, there is substantial agreement on the great powers of the period. According to one measure of the distribution of power resources, which spans 1816 to 2012 and incorporates both economic size and efficiency, all the countries I study rank among the top six at the beginning of the case.33
The Diffusion of GPTs
Scholars often gravitate to technological change as the source of a power transition in which the mantle of industrial preeminence changes hands. However, there is less clarity over the process by which technical breakthroughs translate into this power shift among countries at the technological frontier. I argue that the diffusion of GPTs is the key to this mechanism. In this section, I first outline why my theory privileges GPTs over other types of technology. I then set forth why diffusion should be prioritized over other phases of technological change, especially innovation. Finally, I position GPT diffusion theory against the leading sector (LS) model, which is the standard explanation in the international relations literature.
Why GPTs?
Not all technologies are created equal. When assessed on their potential to transform the productivity of nations, some technical advances, such as the electric dynamo, rank higher than others, such as an improved sleeping bag. My theory gives pride of place to GPTs, such as electricity and the steam engine, which have historically generated waves of economy-wide productivity growth.34 Assessed on their own merits alone, even the most transformative technological changes do not tip the scale far enough to significantly affect aggregate economic productivity.35 GPTs are different because their impact on productivity comes from accumulated improvements across a wide range of complementary sectors; that is, they cannot be judged on their own merits alone.
Recognized by economists and economic historians as “engines of growth,” GPTs are defined by three characteristics.36 First, they offer great potential for continual improvement. While all technologies offer some scope for improvement, a GPT “has implicit in it a major research program for improvements, adaptations, and modifications.”37 Second, GPTs acquire pervasiveness. As a GPT evolves, it finds a “wide variety of uses” and a “wide range of uses.”38 The former refers to the diversity of a GPT’s use cases, while the latter alludes to the breadth of industries and individuals using a GPT.39 Third, GPTs have strong technological complementarities. In other words, the benefits from innovations in GPTs come from how other linked technologies are changed in response and cannot be modeled from a mere reduction in the costs of inputs to the existing production function. For example, the overall energy efficiency gains from merely replacing a steam engine with an electric motor were minimal; the major benefits from factory electrification came from electric “unit drive,” which enabled machines to be driven individually by electric motors, and a radical redesign of plants.40
Taken together, these characteristics suggest that the full impact of a GPT materializes via an “extended trajectory” that differs from those associated with other technologies. Economic historian Paul David explains:
We can recognize the emergence of an extended trajectory of incremental technical improvements, the gradual and protracted process of diffusion into widespread use, and the confluence with other streams of technological innovation, all of which are interdependent features of the dynamic process through which a general purpose engine acquires a broad domain of specific applications.41
For example, the first dynamo for industrial application was introduced in the 1870s, but the major boost of electricity to overall manufacturing productivity did not occur until the 1920s. Like other GPT trajectories, electrification required a protracted process of workforce skill adjustments, organizational adaptations, such as changes in factory layout, and complementary innovations like the steam turbine, which enabled central power generation in the form of utilities.42 To track the full impact of these engines of growth, one must travel the long roads of their diffusion.
Why Diffusion?
All technological trajectories can be divided into a phase when the technology is incubated and then first introduced as a viable commercial application (“innovation”) and a phase when the innovation spreads through a population of potential users, both nationally and internationally (“diffusion”).43 Recognizing this commonly accepted distinction, other studies of the scientific and technological capabilities of nations primarily focus on innovation.44 I depart from other works by giving priority to diffusion, since that is the phase of technological change most significant for GPTs.45
Undeniably, the activities and conditions that produce innovation can also spur diffusion.46 A firm’s ability to conduct breakthrough R&D does not just create new knowledge but also boosts its capacity to assimilate innovations from external sources (“absorptive capacity”).47 Faced with an ever-shifting technological frontier, building competency at producing new innovations gives a firm the requisite prior knowledge for identifying and commercializing external innovations. Other studies extend these insights beyond firms to regional and national systems.48 In order to absorb and diffuse technological advances first incubated elsewhere, they argue, nations must invest in a certain level of innovative activities.
This connection between innovation capacity and absorptive capacity could question the GPT mechanism’s attention to diffusion. Possibly, a country’s lead in GPT innovation could also translate directly into a relative advantage in GPT diffusion.49 Scholarship on the agglomeration benefits of innovation hot spots, such as Silicon Valley, support this case to some extent. Empirical analyses of patent citations indicate that knowledge spillovers from GPTs tend to cluster within a geographic region.50 In the case of electricity, Robert Fox and Anna Guagnini underscore that it was easier for countries with firms at the forefront of electrical innovation to embrace electric power at scale. The interconnections between the “learning by doing” gained on the job in these leading firms and academic labs separated nations in the “fast lane” and “slow lane” of electrification.51
Being the first to pioneer new technologies could benefit a state’s capacity to absorb and diffuse GPTs, but it is not determinative. A country’s absorptive capacity also depends on many other factors, including institutions for technology transfer, human capital endowments, openness to trade, and information and communication infrastructure.52 Sometimes the “advantages of backwardness” allow laggard states to adopt new technologies faster than the states that pioneer such advances.53 In theory and practice, a country’s ability to generate fundamental, new-to-the-world innovations can widely diverge from its ability to diffuse such advances.
This potential divergence is especially relevant for advanced economies, which encompass the great powers that are the subject of this research. Although innovation-centered explanations do well at sorting the advantages of technological breakthroughs to countries at the technological frontier compared to those trying to catch up, they are less effective at differentiating among advanced economies. As supported by a wealth of econometric research, divergences in the long-term economic growth of countries at the technology frontier are shaped more by imitation than innovation.54 These advanced countries have firms that can quickly copy or license innovations; first mover advantages from innovations are thus limited even in industries, like pharmaceuticals, that enforce intellectual property rights most strictly.55 Nevertheless, advanced countries that are evenly matched in their capacity for radical innovation can undertake vastly different growth trajectories in the wake of technological revolutions. Differences in diffusion pathways are central to explaining this puzzle.
This diffusion-centered approach is especially well suited for GPTs. Since GPTs entail gradual evolution into widespread use, there is a longer window for competitors to adopt GPTs more intensively than the leading innovation center. In other technologies, first-mover benefits from pioneering initial breakthroughs are more significant. For instance, leadership in the innovation of electric power technologies was fiercely contested among the industrial powers. The United States, Germany, Great Britain, and France all built their first central power stations within a span of three years (1882–1884), their first electric trams within a span of nine years (1887–1896), and their first three-phase AC power systems within a span of eight years (1891–1899).56 However, the United States clearly led in the diffusion of these systems: by 1912, its electricity production per capita had more than doubled that of Germany, its closest competitor.57 Thus, while most countries at the technological frontier will be able to compete in the production and innovation of GPTs, the hardest hurdles in the GPT trajectory are in the diffusion phase.
GPT Diffusion and LS Product Cycles
GPT diffusion challenges the LS-based account of how technological change drives power transitions. The standard explanation in the international relations literature emphasizes a country’s dominance in leading sectors, defined as new industries that experience rapid growth on the back of new technologies.58 Cotton textiles, steel, chemicals, and the automobile industry form a “classic sequence” of “great leading sectors,” developed initially by economist Walt Rostow and later adapted by political scientists.59 Under the LS mechanism, a country’s ability to maintain a monopoly on innovation in these emerging industries determines the rise and fall of lead economies.60
This model of technological change and power transition builds on the international product life cycle, a concept pioneered by Raymond Vernon. Constructed to explain patterns of international trade, the cycle begins with a product innovation and subsequent sales growth in the domestic market. Once the domestic market is saturated, the new product is exported to foreign markets. Over time, production shifts to these markets, as the original innovating country loses its comparative advantage.61
LS-based studies frequently invoke the product cycle model.62 Analyzing the effects of leading sectors on the structure of the international system, Gilpin states, “Every state, rightly, or wrongly, wants to be as close as possible to the innovative end of ‘the product cycle.’ ”63 One scholar described Gilpin’s US Power and the Multinational Corporation, one of the first texts that outlines the LS mechanism, as “[having] drawn on the concept of the product cycle, expanded it into the concept of the growth and decline of entire national economies, and analyzed the relations between this economic cycle, national power, and international politics.”64
The product cycle’s assumptions illuminate the differences between the GPT and LS mechanisms along three key dimensions. In the first stage of the product cycle, a firm generates the initial product innovation and profits from sales in the domestic market before saturation. Extending this model to national economies, the LS mechanism emphasizes the clustering of LS innovations and attendant monopoly profits in a single nation.65 “The extent of national success that we have in mind is of the fairly extreme sort,” write George Modelski and William Thompson. “One national economy literally dominates the leading sector during its phase of high growth and is the primary beneficiary of the immediate profits.”66 The GPT trajectory, in contrast, places more value on where technologies are diffused than where an innovation is first pioneered.67 I refer to this dimension as the “phase of relative advantage.”
In the next stage, the product innovation spreads to global markets and the technology gradually diffuses to foreign competitors. Monopoly rents associated with a product innovation dissipate as production becomes routinized and transfers to other countries.68 Mirroring this logic, Modelski and Thompson write, “[Leading sectors] bestow the benefits of monopoly profits on the pioneer until diffusion and imitation transform industries that were once considered radically innovative into fairly routine and widespread components of the world economy.”69 Thompson also states that “the sector’s impact on growth tends to be disproportionate in its early stages of development.”70
The GPT trajectory assumes a different impact timeframe. The more wide-ranging the potential applications of a technology, the longer the lag between its initial emergence and its ultimate economic impact. This explains why the envisioned transformative impact of GPTs does not appear straightaway in the productivity statistics.71 Time for complementary innovations, organizational restructuring, and institutional adjustments such as human capital formation is needed before the full impact of a GPT can be known. It is precisely the period when diffusion transforms radical innovations into routine components of the economy—the stage when the causal effects of leading sectors are expected to dissipate—that generates the productivity gap between nations.
The product cycle also reveals differences between the LS and GPT mechanisms regarding the “breadth of growth.” Like the product cycle’s focus on an innovation’s life cycle within a singular industry, the LS mechanism emphasizes the contributions of a limited number of new industries to economic growth in a particular period. GPT-fueled productivity growth, on the other hand, is dispersed across a broad range of industries.72 Table 2.1 specifies how LS product cycles differ from GPT diffusion along the three dimensions outlined here. As the following section will show, the differences in these two technological trajectories shape the institutional factors that are most important for national success in adapting to periods of technological revolution.
Table 2.1 Two Mechanisms of Technological Change and Power Transitions
Mechanisms
Impact Timeframe
Phase of Relative Advantage
Breadth of Growth
Institutional Complements
LS product cycles
Lopsided in early stages
Monopoly on innovation
Concentrated
Deepen skill base in LS innovations
GPT diffusion
Lopsided in later stages
Edge in diffusion
Dispersed
Widen skill base in spreading GPTs
While I have highlighted the differences between GPT diffusion and LS product cycles, it is important to recognize that there are similarities between the two pathways.73 Some scholars, for example, associate leading sectors with broad spillovers across economic sectors.74 In addition, lists of leading sectors and lists of GPTs sometimes overlap, as evidenced by the fact that electricity is a consensus inclusion on both lists. Moreover, both explanations begin with the same premise: to fully uncover the dynamics of technology-driven power transitions, it is essential to specify which new technologies are the key drivers of economic growth in a particular time window.75
At the same time, these resemblances should not be overstated. Many classic leading sectors do not have general-purpose applications. For instance, cotton textiles and automobiles both feature on Rostow’s series of leading sectors, and they are studied as leading sectors because each has “been the largest industry for several major industrial nations in the West at one time or another.”76 Although these were certainly both fast-growing large industries, the underlying technological advances do not fulfill the characteristics of GPTs. In addition, many of the GPTs I examine do not qualify as leading sectors. The machine tool industry in the mid-nineteenth century, for instance, was not a new industry, and it was never even close to being the largest industry in any of the major economies. Most importantly, though the GPT and LS mechanisms sometimes point to similar technological changes, they present very different understandings of how revolutionary technologies bring about an economic power transition. As the next section reveals, these differences also map onto varied institutional adaptations.
GPT Skill Infrastructure
New technologies agitate existing institutional patterns.77 They appeal for government support, generate new collective interests in the form of technical societies, and induce organizations to train people in relevant fields. If institutional environments are slow or fail to adapt, the development of emerging technologies is hindered. As Gilpin articulates, a nation’s technological fitness is rooted in the “extent of the congruence” between its institutions and the demands of evolving technologies.78 This approach is rooted in a rich tradition of work on the coevolution of technology and institutions.79
Understanding the demands of GPTs helps filter which institutional factors are most salient for how technological revolutions bring about economic power transitions. Which institutional factors dictate disparities in GPT adoption among great powers? Specifically, I emphasize the role of education and training systems that broaden the base of engineering skills linked to a particular GPT. This set of institutions, which I call “GPT skill infrastructure,” is most crucial for facilitating the widespread adoption of a GPT.
To be sure, GPT diffusion is dependent on institutional adjustments beyond GPT skill infrastructure. Intellectual property regimes, industrial relations, financial institutions, and other institutional factors could affect GPT diffusion. Probing inter-industry differences in technology adoption, some studies find that less concentrated industry structures are positively linked to GPT adoption.80 I limit my analysis to institutions of skill formation because their effects permeate other institutional arrangements.81 GPT skill infrastructure provides a useful indicator for other institutions that standardize and spread the novel best practices associated with GPTs.82
It should also be noted that the institutional approach is one of three main categories of explanation for cross-country differences in economic performance over the long term.83 Other studies document the importance of geography and culture to persistent cross-country income differences.84 I prioritize institutional explanations for two reasons. First, natural experiments from certain historical settings, in which institutional divergence occurs but geographical and cultural factors are held constant, suggest that institutional differences are particularly influential sources of long-term economic growth differentials.85 Second, since LS-based accounts of power transitions also prioritize institutional adaptations to technological change, my approach provides a more level test of GPT diffusion against the standard explanation.86
One final note about limits to my argument’s scope. I do not investigate the deeper origins of why some countries are more effective than others at developing GPT skill infrastructure. Possibly, the intensity of political competition and the inclusiveness of political institutions influence the development of skill formation institutions.87 Other fruitful lines of inquiry stress the importance of government capacity to make intertemporal bargains and adopt long time horizons in making technology investments.88 It is worth noting that a necessary first step to productively exploring these underlying causes is to establish which types of technological trajectories and institutional adaptations are at work. For instance, LS product cycles may be more closely linked to mercantilist or state capitalist approaches that favor narrow interest groups, whereas, political systems that incorporate a broader group of stakeholders may better accommodate GPT diffusion pathways.
Institutions Fit for GPT Diffusion
If GPTs drive economic power transitions, which institutions fit best with their demands? Institutional adaptations for GPT diffusion must solve two problems. First, since the economic benefits of GPTs materialize through improvements across a broad range of industries, capturing these benefits requires extensive coordination between the GPT sector and numerous application sectors. Given the sheer scope of potential applications, it is infeasible for firms in the GPT sector to commercialize the technology on their own, as the necessary complementary assets are embedded with different firms and industries.89 In the AI domain, as one example of a potential GPT, firms that develop general machine learning algorithms will not have access to all the industry-specific data needed to fine-tune those algorithms to particular application scenarios. Thus, coordination between the GPT sector and other organizations that provide complementary capital and skills, such as academia and competitor firms, is crucial. In contrast, for technologies that are not general-purpose, this type of coordination is less conducive and could even be detrimental to a nation’s competitive advantage, as the innovating firm could leak its technical secrets.90
Second, GPTs pose demanding conditions for human capital adjustments. In describing the connection between skill formation and technological fitness, scholars often delineate between general skills and industry-specific skills. According to this perspective, skill formation institutions that optimize for the former are more conducive to technological domains characterized by radical innovation, while institutions that optimize for the latter are more favorable for domains marked by incremental innovation.91 GPT diffusion entails both types of skill formation. The skills must be specific to a rapidly changing GPT domain but also broad enough to enable a GPT’s advance across many industries.92 Strong linkages between R&D-intensive organizations at the technological frontier and application areas far from the frontier also play a key role in GPT diffusion. This draws attention to the interactions between researchers who produce innovations and technicians who help absorb them into specific contexts.93
Education and training systems that foster relevant engineering skills for a GPT, or what I call GPT skill infrastructure, address both constraints. Engineering talent fulfills the need for skills that are rooted in a GPT yet sufficiently flexible to implement GPT advances in a wide range of sectors. Broadening the base of engineering knowledge also helps standardize best practices with GPTs, thereby coordinating information flows between the GPT sector and application sectors. Standardization fosters GPT diffusion by committing application sectors to specific technological trajectories and encouraging complementary innovations.94 This unlocks the horizontal spillovers associated with GPTs.95
Indeed, distinct engineering specializations have emerged in the wake of a new GPT. New disciplines, such as chemical engineering and electrical engineering, have proved essential in widening knowledge bases in the wake of a new GPT.96 Computer science, another engineering-oriented field, was central to US leadership in the information revolution.97 These professions developed alongside new technical societies—ranging from the American Society of Mechanical Engineers to the Internet Engineering Task Force—that formulated and disseminated guidelines and benchmarks for GPT development.98
Clearly, the features of GPT skill infrastructure have changed over time. Whereas informal associations systematized the skills crucial for mechanization in the eighteenth century, formal higher education has become increasingly integral to computerization in the twenty-first century.99 Some evidence suggests that computers and other technologies are skill-biased, in the sense that they favor workers with more years of schooling.100 These trends complicate but do not undercut the concept of GPT skill infrastructure. Regardless of the extent of formal training, all configurations of GPT skill infrastructure perform the same function: to widen the pool of engineering skills and knowledge associated with a GPT. This can take place in universities as well as in informal associations, provided these settings train engineers and facilitate the flow of engineering knowledge between knowledge-creation centers and application sectors.101
Which Institutions Matter?
The institutional competencies for exploiting LS product cycles are different. Historical analysis informed by this frame highlights heroic inventors like James Watt and pioneering research labs at large companies.102 Studying which countries benefited most from emerging technologies over the past two centuries, Herbert Kitschelt prioritizes the match between the properties of new technologies and sectoral governance structures. Under his framework, for example, tightly coupled technological systems with high causal complexity, such as nuclear power systems and aerospace platforms, are more likely to flourish in countries that allow for extensive state support.103 In other studies, the key institutional factors behind success in LS product cycles are education systems that subsidize scientific training and R&D facilities in new industries.104
These approaches equate technological leadership with a state’s success in capturing market shares and monopoly profits in new industries.105 In short, they use LS product cycles as the filter for which institutional variables matter. Existing scholarship lacks an institutional explanation for why some great powers are more successful at GPT diffusion.
Competing interpretations of technological leadership in chemicals during the late nineteenth century crystallize these differences. Based on the LS template, the standard account accredits Germany’s dominance in the chemical industry—as represented by its control over 90 percent of global production of synthetic dyes—to its investments in scientific research and highly skilled chemists.106 Germany’s dynamism in this leading sector is taken to explain its overall industrial dominance.107
GPT diffusion spotlights a different relationship between technological change and institutional adaptation. The focus turns toward institutions that complemented the extension of chemical processes to a wide range of industries beyond synthetic dye, such as food production, metals, and textiles. Under the GPT mechanism, the United States, not Germany, achieved leadership in chemicals because it first institutionalized chemical engineering as a discipline. Despite its disadvantages in synthetic dye production and chemical research, the United States was more effective in broadening the base of chemical engineering talent and coordinating information flows between fundamental breakthroughs and industrial applications.108
It is important to note that some parts of the GPT and LS mechanisms can coexist without conflict. A state’s capacity to pioneer new technologies can correlate with its capacity to absorb and diffuse GPTs. Countries that are home to cutting-edge R&D infrastructure may also be fertile ground for education systems that widen the pool of GPT-linked engineering skills. However, these aspects of the LS mechanism are not necessary for the GPT mechanism to operate. In accordance with GPT diffusion theory, a state can capitalize on GPTs to become the most powerful economy without monopolizing LS innovation.
Moreover, other dimensions of these two mechanisms directly conflict. When it comes to impact timeframe and breadth of growth, the GPT and LS mechanisms advance opposing expectations. Institutions suited for GPT diffusion can diverge from those optimized for creating new-to-the-world innovations. Research on human capital and long-term growth separates the effects of engineering capacity, which is commonly tied to adoptive activities, and other forms of human capital that are more often connected to inventive activities.109 This divergence can also be seen in debates over the effects of competition on technological activity. On the one hand, Joseph Schumpeter and others have argued that monopoly structures incentivize more R&D activity because the monopolists can appropriate all the gains from technological innovation.110 On the other hand, empirical work demonstrates that more competitive market structures increase the rate of technological adoption across firms.111 Thus, while there is some overlap between these two mechanisms, they can still be set against each other in a way that improves our understanding of technological revolutions and power transitions.
This theoretical framework differs from related work on the political economy of technological change.112 Scholars attribute the international competitiveness of nations to broader institutional contexts, including democracy, national innovation systems, and property rights enforcement.113 Since this book is limited to the study of shifts in productivity leadership at the technological frontier, many of these factors, such as those related to basic infrastructure and property rights, will not explain differences among technologically advanced nations.
In addition, most of the institutional theories put forth to explain the productivity of nations are technology-agnostic, in that they treat all forms of technological change equally. To borrow language from a former chairman of the US Council of Economic Advisers, they do not differentiate between an innovation in potato chips and an innovation in microchips.114 In contrast, I am specific about GPTs as the sources of shifts in competitiveness at the technological frontier.
Other theories identify key technologies but leave institutional factors at a high level of abstraction. Some scholars, for instance, posit that the lead economy’s monopoly on leading-sector innovation eventually erodes because of “ubiquitous institutional rigidities.”115 Unencumbered by the vested interests that resist disruptive technologies, rising challengers inevitably overtake established powers. Because these explanations are underspecified, they cannot account for cases where rich economies expand their lead or where poorer countries do not catch up.116
When interpreting great power competition at the technological frontier, adjudicating between the GPT and LS mechanisms represents a choice between two different visions. The latter prioritizes being the first country to introduce novel technologies, whereas the former places more value on disseminating and transforming innovations after their inception. In sum, industrial competition among great powers is not a sprint to determine which one can create the most brilliant Silicon Valley; it is a marathon won by the country that can cultivate the closest connections between its Silicon Valleys and its Iowa Citys.
Alternative Explanations
Although I primarily set GPT diffusion theory against the LS model, I also consider two other prominent explanations that make specific claims about how technological breakthroughs differentially advantage leading economies. Crucially, these two lines of thinking could account for differences in GPT diffusion, nullifying the import of GPT skill infrastructure.
Threat-Based Arguments
According to one school of thought, international security threats motivate states to invest in science and technology.117 When confronted with more threatening geopolitical landscapes, states are more incentivized to break down status quo interests and build institutions conducive to technological innovation.118 Militaries hold outsized influence in these accounts. For example, Vernon Ruttan argues that military investment, mobilized against war or the threat of war, fueled commercial advances in six technologies designated as GPTs.119 Studies of the success of the United States and Japan with emerging technologies also stress interconnections between military and civilian technological development.120 I group these related arguments under the category of threat-based theories.
Related explanations link technological progress with the balance of external threats and domestic roadblocks. Mark Taylor’s “creative insecurity” theory describes how external economic and military pressures permit governments to break from status quo interest groups and promote technological innovation. He argues that the difference between a nation’s external threats and its internal rivalries determines its propensity for innovation: the greater the difference, the greater the national innovation rate.121 Similarly, “systemic vulnerability” theory emphasizes the influence of external security and domestic pressures on the will of leaders to invest in institutions conducive to innovation, as well as the effect of “veto players” on their ability to do so.122
Certainly, external threats could impel states to invest more in GPTs, and military investment can help bring forth new GPTs; however, there are several issues with adapting threat-based theories to explain differences in GPT diffusion across great powers. First, threat-based arguments tend to focus on the initial incubation of GPTs, as opposed to the gradual spread of GPTs throughout a national economy. During the latter phase, a great deal of evidence suggests that civilian and military needs can greatly conflict.123 Besides, some GPTs, such as electricity in the United States, developed without substantial military investment. Since other civilian institutions could fill in as strong sources of demand for GPTs, military procurement may not be necessary for spurring GPT diffusion. Institutional adjustments to GPTs therefore can be motivated by factors other than threats. Ultimately, to further probe these points of clash, the impact of security threats and military investment must be traced within the historical cases.
Varieties of Capitalism
The “varieties of capitalism” (VoC) explanation highlights differences among developed democracies in labor markets, industrial organization, and interfirm relations and separates them into coordinated market economies (CMEs) and liberal market economies (LMEs). VoC scholars argue that CMEs are more suited for incremental innovations because their thick intercorporate networks and protected labor markets favor gradual adoption of new technological advances. LMEs, in contrast, are more adept at radical innovation because their fluid labor markets and corporate organization make it easier for firms to reorganize themselves around disruptive technologies. Most relevant to GPT diffusion theory, VoC scholars argue that LMEs incentivize workers to acquire general skills, which are more accommodative to radical innovation, whereas CMEs support industry-specific training, which is more favorable for incremental innovation.124
It is possible that differences between market-based capitalism and strategically coordinated capitalism account for GPT diffusion gaps between nations. Based on the expectations of the VoC approach, LMEs should be more likely to generate innovations with the potential to become GPTs, and workers in LMEs should possess more general skills that could spread GPTs across firms.125 Examining the pattern of innovation during the information revolution, scholars find that the United States, an LME, concentrated on industries experiencing radical innovation, such as semiconductors and telecommunications, while Germany, a CME, specialized in domains characterized by incremental innovation, such as mechanical engineering and transport.126
Despite bringing vital attention to the diversity of skill formation institutions, VoC theory’s dichotomy between general and industry-specific skills does not dovetail with the skills demanded by specific GPTs.127 Cutting across this sometimes arbitrary distinction, the engineering skills highlighted in GPT diffusion theory are specific to a fast-evolving GPT field and general enough to transfer ideas from the GPT sector across various sectors. Software engineering skills, for instance, are portable across multiple industries, but their reach is not as ubiquitous as critical thinking skills or mathematics knowledge. To address similar gaps in skill classifications, many political economists have appealed for “a more fine-grained analysis of cross-national differences in the particular mix of jobs and qualifications that characterize different political economies.”128 In line with this move, GPT skill infrastructure stands in for institutions that supply the particular mix of jobs and qualifications for enabling GPT diffusion. The empirical analysis provides an opportunity to examine whether this approach should be preferred to the VoC explanation for understanding technology-driven power transitions.
Research Methodology
My evaluation of the GPT and LS mechanisms primarily relies on historical case studies, which allow for detailed exploration of the causal processes that connect technological change to economic power transitions. Employing congruence-analysis techniques, I select cases and assess the historical evidence in a way that ensures a fair and rigorous test of the relative explanatory power of the two mechanisms.129 This sets up a “three-cornered fight” among GPT diffusion theory, the rival LS-based explanation, and the set of empirical information.130
The universe of cases most useful for assessing the GPT and LS mechanisms are technological revolutions (cause) that produced an economic power transition (outcome) in the industrial period. Following guidance on testing competing mechanisms that prioritize typical cases where the cause and outcome are clearly present, I investigate the First Industrial Revolution (IR-1) and the Second Industrial Revolution (IR-2).131 Both cases featured clusters of disruptive technological advances, highlighted by some studies as “technological revolutions” or “technology waves.”132 They also saw economic power transitions, when one great power sustained growth rates at substantially higher levels than its rivals.133 I also study Japan’s challenge to American economic leadership—which ultimately failed—in the Third Industrial Revolution (IR-3). This deviant case can disconfirm mechanisms and help explain why they break down.134
These cases are highly crucial and relevant for testing the GPT mechanism against the LS mechanism. All three cases favor the latter in terms of background conditions and existing theoretical explanations. Scholarship has attributed shifts in economic power during this period to the rise of new leading sectors.135 Thus, if the empirical results support the GPT mechanism, then my findings would suggest a need for major modifications to our understanding of how technological revolutions affect the rise and fall of great powers. The qualitative analysis appendix provides further details on case selection, including the universe of cases, the justification for these cases as “most likely cases” for the LS mechanism, and relevant scope conditions.136
This overall approach adapts the methodology of process-tracing, often conducted at the individual or micro level, to macro-level mechanisms that involve structural factors and evolutionary interactions.137 Existing scholarship on diffusion mechanisms, which the GPT mechanism builds from, emphasizes the influence of macro-level processes. In these accounts the diffusion trajectory depends not just on the overall distribution of individual-level receptivity but also on structural and institutional features, such as the degree of interconnectedness in a population.138 This approach aligns with a view of mechanistic thinking that allows for mechanisms to be set at different levels of abstraction.139 As Tulia Falleti and Julia Lynch point out, “Micro-level mechanisms are no more fundamental than macro-level ones.”140
To judge the explanatory strength of the LS and GPT mechanisms, I employ within-case congruence tests and process-tracing principles to evaluate the predictions of the two theoretical approaches against the empirical record.141 In each historical case, I first trace how leading sectors and GPTs developed in the major economies, paying particular attention to adoption timeframes, the technological phase of relative advantage, and the breadth of growth—three dimensions that differentiate GPT diffusion from LS product cycles.142
For example, my assessment of the two mechanisms along the impact time-frame dimension follows consistent procedures in each case.143 To evaluate when certain technologies were most influential, I establish when they initially emerged (based on dates of key breakthroughs), when their associated industries were growing fastest, and when they diffused across a wide range of application sectors. When data are available, I estimate a GPT’s initial arrival date by also factoring in the point at which it reached a 1 percent adoption rate in the median sector.144 Industry growth rates, diffusion curves, and output trends all help measure the timeline along which technological breakthroughs substantially influenced the overall economy. The growth trajectory of each candidate GPT and LS is then set against a detailed timeline of when a major shift in productivity leadership occurs.
I then turn to the institutional factors that could explain why some countries were more successful in adapting to a technological revolution, with a focus on the institutions best suited to the demands of GPTs and leading sectors.145 If the GPT mechanism is operative, the state that attains economic leadership should have an advantage in institutions that broaden the base of engineering human capital and spread best practices linked to GPTs. Additional evidence of the GPT diffusion theory’s explanatory power would be that other countries had advantages in institutions that complement LS product cycles, such as scientific research infrastructure and sectoral governance structures.
These evaluation procedures are effective because I have organized the competing mechanisms “so that they are composed of the same number of diametrically opposite parts with observable implications that rule each other out.”146 This allows for evidence in favor of one explanation to be doubly decisive in that it also undermines the competing theory.147 In sum, each case study is structured around investigating a set of four standardized questions that correspond to the three dimensions of the LS and GPT mechanisms as well as the institutional complements to technological trajectories (table 2.2).148
table 2.2 Testable Propositions of the LS and GPT Mechanisms
Dimensions
Key Questions
LS Propositions
GPT Propositions
Impact timeframe
When do revolutionary technologies make their greatest marginal impact on the economic balance of power?
New industries make their greatest impact on growth differentials in early stages.
GPTs do not make a significant impact on growth differentials until multiple decades after emergence.
Key phase of relative advantage
Do monopoly profits from innovation or benefits from more successful diffusion drive growth differentials?
A state’s monopoly on innovation in leading sectors propels it to economic leadership.
A state’s success in widespread adoption of GPTs propels it to economic leadership.
Breadth of growth
What is the breadth of technology-driven growth?
Technological advances concentrated in a few leading sectors drive growth.
Technological advances dispersed across a broad range of GPT-linked industries drive growth.
Institutional complements
Which types of institutions are most advantageous for national success in technological revolutions?
Key institutional adaptations help a state capture market shares and monopoly profits in new industries.
Key institutional adaptations widen the base of engineering skills and knowledge for GPT diffusion.
In each case study, I consider alternative theories of technology-driven power transitions. Countless studies have examined the rise and fall of great powers. My aim is not to sort through all possible causes of one nation’s rise or another’s decline. Rather, I am probing the causal processes behind an established connection between technological advances in each industrial revolution and an economic power transition. The VoC framework and threat-based theories outline alternative explanations for how significant technological advances translated into growth differentials among great powers. Across all the cases, I assess whether they provide a better explanation for the historical case evidence than the GPT and LS mechanisms.
I also address case-specific confounding factors. For example, some scholars argue that abundant inputs of wood and metals induced the United States to embrace more machine-intensive technology in the IR-2, reasoning that Britain’s slower adoption of interchangeable parts manufacturing was an efficient choice given its natural resource constraints.149 For each case, I determine whether these types of circumstantial factors could nullify the validity of the GPT and LS mechanisms.
In tracing these mechanisms, I benefit from a wealth of empirical evidence on past industrial revolutions, which have been the subject of many interdisciplinary inquiries. Since the cases I study are well-traversed terrain, my research is primarily based on secondary sources.150 I rely on histories of technology and general economic histories to trace how and when technological breakthroughs affected economic power balances. Notably, my analysis takes advantage of the application of formal statistical and econometric methods to assess the impact of significant technological advances, part of the “cliometric revolution” in economic history.151 Some of these works have challenged the dominant narrative of previous industrial revolutions. For instance, Nick von Tunzelmann found that the steam engine made minimal contributions to British productivity growth before 1830, raising the issue that earlier accounts of British industrialization “tended to conflate the economic significance of the steam engine with its early diffusion.”152
I supplement these historical perspectives with primary sources. These include statistical series on industrial production, census statistics, discussions of engineers in contemporary trade journals, and firsthand accounts from commissions and study teams of cross-national differences in technology systems. In the absence of standardized measures of engineering education, archival evidence helps fill in details about GPT skill infrastructure for each of the cases. In the IR-1 case, I benefit from materials from the National Archives (United Kingdom), the British Newspaper Archive, and the University of Nottingham Libraries, Manuscripts, and Special Collections. My IR-2 case analysis relies on collections based at the Bodleian Library (United Kingdom), the Library of Congress (United States), and the University of Leipzig and on British diplomatic and consular reports.153 In the IR-3 case analysis, the Edward A. Feigenbaum Papers collection, held at Stanford University, helps inform US-Japan comparisons in computer science education.
My research also benefits greatly from new data on historical technological development. I take advantage of improved datasets, such as the Maddison Project Database.154 New ones, such as the Cross-Country Historical Adoption of Technology dataset, were also beneficial.155 Sometimes hype about exciting new technologies influences the perceptions of commentators and historians about the pace and extent of technology adoption. More granular data can help substantiate or cut through these narratives. Like the reassessments of the impact of previous technologies, these data were released after the publication of the field-defining works on technology and power transitions in international relations. Making extensive use of these sources therefore provides leverage to revise conventional understandings.
When assessing these two mechanisms, one of the main challenges is to identify the key technological changes to trace. I take a broad view of technology that encompasses not just technical designs but also organizational and managerial innovations.156 Concretely, I follow Harvey Brooks, a pioneer of the science and technology policy field, in defining technology as “knowledge of how to fulfill certain human purposes in a specifiable and reproducible way.”157 The LS and GPT mechanisms both call attention to the outsized import of particular technical breakthroughs and their interactions with social systems, but they differ on which ones are more important. Therefore, a deep and wide understanding of advances in hardware and organizational practices in each historical period is required to properly sort them by their potential to spark LS or GPT trajectories.
This task is complicated by substantial disagreements over which technologies are leading sectors and GPTs. Proposed lists of GPTs often conflict, raising questions about the criteria used for GPT selection.158 Reacting to the length of such lists, other scholars fear that “the [GPT] concept may be getting out of hand.”159 According to one review of eleven studies that identified past GPTs, twenty-six different innovations appeared on at least one list but only three appeared on all eleven.160
The LS concept is even more susceptible to these criticisms because the characteristics that define leading sectors are inconsistent across existing studies. Though most scholars agree that leading sectors are new industries that grow faster than the rest of the economy, there is marked disagreement on other criteria. Some scholars select leading sectors based on the criterion that they have been the largest industry in several major industrial nations for a period of time.161 Others emphasize that leading sectors attract significant investments in R&D.162 To illustrate this variability, I reviewed five key texts that analyze the effect of leading sectors on economic power transitions. Limiting the lists of proposed leading sectors to those that emerged during the three case study periods, I find that fifteen leading sectors appeared on at least one list and only two appeared on all five.163
My process for selecting leading sectors and GPTs to trace helps alleviate concerns that I cherry-pick the technologies that best fit my preferred explanation. In each historical case, most studies that explicitly identify leading sectors or GPTs agree on a few obvious GPTs and leading sectors. To ensure that I do not omit any GPTs, I consider all technologies singled out by at least two of five key texts that identify GPTs across multiple historical periods.164 I apply the same approach to LS selection, using the aforementioned list I compiled.
Following classification schemes that differentiate GPTs from “near-GPTs” and “multipurpose technologies,” I resolve many of the conflicts over what counts as a GPT or leading sector by referring to a set of defining criteria.165 For instance, while some accounts include the railroad and the automobile as GPTs, I do not analyze them as candidate GPTs because they lack a variety of uses.166 I support my choices with empirical methods for LS and GPT identification. To confirm certain leading sectors, I examine the rate of growth across various industry sectors. I also leverage recent studies that identify GPTs with patent-based indicators.167 Taken together, these procedures limit the risk of omitting certain technologies while guarding against GPT and LS concept creep.168 The qualitative analysis appendix outlines how I address other issues related to LS and GPT identification, including concerns about omitting important single-purpose technologies and scenarios when certain technological breakthroughs are linked to both LS and GPT trajectories.
These considerations underscore that taking stock of the key technological drivers is only the first step in the case analysis. To judge whether these breakthroughs actually brought about the impacts that are often claimed for them, it is important to carefully trace how these technologies evolved in close relation with societal systems.
As a complement to the historical case studies, this book’s research design includes a large-n quantitative analysis of the relationship between the breadth of software engineering skill formation institutions and computerization rates. This tests a key observable implication of GPT diffusion theory, using time-series cross-sectional data on nineteen advanced and emerging economies across three decades. I leave the more detailed description of the statistical methodology to chapter 6.
Summary
The technological fitness of nations is determined by how they adapt to the demands of new technical advances. I have developed a theory to explain how revolutionary technological breakthroughs affect the rise and fall of great powers. My approach is akin to that of an investigator tasked with figuring out why one ship sailed across the ocean faster than all the others. As though differentiating the winning ship’s route from possible sea-lanes in terms of trade wind conditions and course changes, I first contrast the GPT and LS trajectories with regard to the timing, phase, and breadth of technological change. Once the superior route has been mapped, attention turns to the attributes of the winning ship, such as its navigation equipment and sailors’ skills, that enabled it to take advantage of this fast lane across the ocean. In similar fashion, having set out the GPT trajectory as the superior route from technological revolution to economic leadership, my argument then highlights GPT skill infrastructure as the key institutional attribute that dictates which great power capitalizes best on this route.
3 The First Industrial Revolution and Britain’s Rise
FEW HISTORICAL EVENTS have shaken the world like the First Industrial Revolution (IR-1, 1780–1840). Extraordinary upheaval marked the contours and consequences of the IR-1. For the first time in history, productivity growth accelerated dramatically, allowing large numbers of people to experience sustained improvements in their living standards. Small towns transformed into large cities, new ideologies gathered momentum, and emerging economic and social classes reshaped the fabric of society. These changes reverberated in the international sphere, where the ramifications of the IR-1 included the transition to industrialized mass warfare, the decline of the absolutist state, and the birth of the modern international system.
Among these transformations, two phenomena stand out. The first is the remarkable technological progress that inaugurated the IR-1 period. Everything was changing in part because so many things were changing—water frames, steam engines, and puddling processes not least among them. The second is Britain’s rise to unrivaled economic leadership, during which it sustained productivity growth at higher levels than its rivals, France and the Netherlands. The following sections adjudicate the debates over the exact timeline of Britain’s industrialization, but there is no doubt that Britain, propelled by the IR-1, became the world’s most advanced economic power by the mid-nineteenth century.
No study of technological change and power transitions is complete without an account of the IR-1. For both the LS and GPT mechanisms, the IR-1 functions as a typical case that is held up as paradigmatic of technology-driven power transitions. The standard account in international relations scholarship attributes Britain’s industrial ascent to its dominance of innovation in the IR-1’s leading sectors, including cotton textiles, iron metallurgy, and steam power.1 Present-day scholarship and policy discussions often draw upon stylized views of the IR-1, analogizing present developments in information technology and biotechnology to the effects of steam power and cotton textiles in the industrial revolution.2
A deeper inquiry into the IR-1 and Britain’s economic rise challenges many of these conventional views. First, it reveals that general-purpose transformations linked to advances in iron metallurgy diffused widely enough to significantly affect economy-wide productivity only after 1815—a timeline that aligns with the period when Britain significantly outpaced its rivals in industrialization. Other prominent advances, including the steam engine, made only limited contributions to Britain’s rise to industrial prominence in this period owing to a prolonged period of gestation before widespread adoption. Second, the IR-1 case also demonstrates that it was Britain’s advantage in extending mechanization throughout the economy, not monopoly profits from innovations in cotton textiles, that proved crucial to its industrial ascendancy. Third, the historical data illustrate that the dispersion of mechanical innovations across many sectors fueled British productivity growth. Across these three dimensions, the IR-1 case matches the GPT trajectory better than the LS trajectory.
Since no country monopolized innovations in metalworking processes and Britain’s competitors could also absorb innovations from abroad, why did Britain gain the most from this GPT trajectory? In all countries, as technical advances surged forward, institutional adjustments raced to cultivate the skills required to keep pace. Importantly, France and the Netherlands were competitive with Britain—and even surpassed it in some respects—in scientific research infrastructure and education systems for training expert engineers. These institutional settings in France and the Netherlands, however, produced knowledge and skills that were rather divorced from practical applications.
Britain’s comparative advantage rested on another type of skill infrastructure. It depended less on heroic innovators like James Watt, the famed creator of the modern steam engine, and more on competent engineers who could build and maintain new technological systems, as well as make incremental adaptations to implement these systems in many different settings.3 As expected by GPT diffusion theory, Britain benefited from education systems that expanded the base of mechanically skilled engineers and disseminated knowledge of applied mechanics. Britain’s competitors could not match its system for cultivating a common technical language in applied mechanics that encouraged the knowledge exchanges between engineers and entrepreneurs needed for advancing mechanization from one industry to the next.
To trace these mechanisms, I gathered and sorted through a wealth of evidence on the IR-1. Historical accounts served as the foundational materials, including general economic histories of the IR-1, histories of influential technologies and industries like the steam engine and the iron industry, country-specific histories, and comparative histories of Britain, France, and the Netherlands. I also benefited from contemporary assessments of the IR-1’s institutional features provided by trade journals, proceedings of mechanics’ institutes, recruitment advertisements published in local newspapers, and essays by leading engineers. This evidence stems from archival materials at the British Newspaper Archive, the National Archives (United Kingdom), and the University of Nottingham Libraries, Manuscripts, and Special Collections. Triangulating a variety of sources, I endeavored to back up my claims with statistical evidence in the form of industrial output estimates, patenting rates, and detailed biographical information on British engineers.
The assessment of the GPT and LS mechanisms against historical evidence from the IR-1 proceeds as follows. To begin, the chapter reviews Britain’s rise to industrial preeminence, which is the outcome of the case. Next, it sorts the key technological breakthroughs of the period by their potential to drive two types of trajectories—LS product cycles and GPT diffusion. I then assess whether Britain’s rise in this period is better explained by the GPT or LS mechanism, tracing the development of candidate leading sectors and GPTs in terms of impact timeframe, phase of relative advantage, and breadth of growth. If the GPT trajectory holds for this period, there should be evidence that Britain was better equipped than its competitors in GPT skill infrastructure. Another section evaluates whether the historical data support this expectation. Before concluding the chapter, I address alternative factors and explanations.
A Power Transition: Britain’s Rise
When did Britain ascend to industrial hegemony? The broad outlines of the story are well known. Between the mid-eighteenth century and the mid-nineteenth century, the industrial revolution propelled Great Britain to global preeminence. Although Britain did not boast the world’s largest economy—China held that title during this period—it did capitalize on the technologies of the industrial revolution to become “the world’s most advanced productive power.”4 France and the Netherlands, its economic rivals, did not keep pace with Britain’s productivity growth.
While both the LS and GPT models agree that Britain established itself as the preeminent industrial power in this period, a clearer sense of when this shift occurred is essential for testing the explanatory power of the LS and GPT mechanisms during this period. One common view of Britain’s industrialization, brought to prominence by Rostow, depicts an accelerated takeoff into sustained growth. Rostow’s timeline dates this takeoff to the last two decades of the eighteenth century.5 In alignment with this periodization, some scholars writing in the LS tradition claim that Britain achieved its industrial ascent by the late eighteenth century.6
A different perspective, better supported by the evidence that follows, favors a delayed timeline for Britain’s ascent to industrial preeminence. Under this view, Britain did not sustain economic and productivity advances at levels substantially higher than its rivals until the 1820s and after. To clarify the chronology of Britain’s industrial ascent, the following sections survey three proxies for productivity leadership: GDP per capita, industrialization, and total factor productivity.
GDP PER-CAPITA INDICATORS
Trend lines in GDP per capita, a standard proxy for productivity, confirm the broad outlines of Britain’s rise. Evidence from the Maddison Project Database (MPD) points to the decades after 1800, not before, as the key transition period (figure 3.1).7 These trends come from the 2020 version of the MPD, which updates Angus Maddison’s data and incorporates new annual estimates of GDP per capita in the IR-1 period for France, the Netherlands, and the United Kingdom.8 In 1760, the Netherlands boasted the world’s highest per-capita income, approximately 35 percent higher than Britain’s.9 The Dutch held this lead for the rest of the eighteenth century through to 1808, when Britain first overtook the Netherlands in GDP per capita. By 1840, Britain’s GDP per capita was about 75 percent higher than that of France and about 10 percent ahead of that of the Netherlands.10
figure 3.1 Economic Power Transition in the IR-1. Source: Maddison Project Database, version 2020 (Bolt and van Zanden 2020).
It should be noted that GDP per-capita information for the early years of the IR-1 is sometimes missing or only partially available. For years prior to 1807, the MPD bases Dutch GDP per-capita estimates on data for just the Holland region, so the Dutch economy’s decline during this time could be an artifact of changes in data sources.11 At the same time, to ensure that the MPD data can be used to provide accurate information on historical patterns of economic growth and decline, researchers have made adjustments to partial data series and consulted experts to assess their representativeness.12 Furthermore, the Holland-based data in the early 1800s already indicated a decline in Dutch GDP per capita. Although data scarcity makes it difficult to mark out exactly when Britain’s GDP per capita surpassed that of the Netherlands, the MPD remains the best source for cross-country comparisons of national income in this period.
INDUSTRIALIZATION INDICATORS
Industrialization indicators depict a mixed picture of when Britain sustained leadership in economic efficiency. By one influential set of metrics compiled by economic historian Paul Bairoch, Britain’s per-capita industrialization levels had grown to 50 percent higher than those of France in 1800, from a position of near-equality with France and Belgium in 1750. For scholars who map the trajectories of great powers, these estimates have assumed a prominent role in shaping the timeline of British industrial ascendance.13 For instance, Paul Kennedy employs Bairoch’s estimates to argue that the industrial revolution transformed Britain into a different kind of world power.14
Further examination of Bairoch’s estimates qualifies their support for an accelerated timeline of Britain’s industrial ascendance. First, by limiting his definition of “industrial output” to manufacturing industry products, Bairoch excludes the contribution of notable sectors such as construction and mining, a distinction even he admits is “rather arbitrary.”15 Second, the gap between Britain and France in per-capita industrialization levels in 1800 still falls within the margin of error for Bairoch’s estimates.16
Moreover, a delayed timeframe is supported by industrialization measures that encompass more than the manufacturing industries. In 1700, the Netherlands had a substantially higher proportion of its population employed in industry (33 percent) compared to the United Kingdom (22 percent). In 1820, the proportion of people employed in UK industry had risen to 33 percent—higher than the Dutch corresponding rate of 28 percent.17 One expert on the pre-industrial revolution in Europe notes that the Netherlands was at least as industrialized as England, if not more so, throughout the eighteenth century.18 Lastly, aggregate industrial production trends map out a post-1815 surge in British industrialization, providing further evidence that Britain did not solidify its productivity advantage until decades into the nineteenth century.19
PRODUCTIVITY INDICATORS
Total factor productivity (TFP) indicators, which capture the efficiency by which production factors are converted into useful outputs, further back the delayed ascent story. As was the case with trends in aggregate industrial output, TFP growth in Britain did not take off until after 1820.20 In truth, British TFP growth was very modest throughout the eighteenth century, averaging less than 1 percent per year.21
While the paucity of reliable data on total factor productivity in France hinders cross-national comparisons in this period, some evidence suggests that Britain did not surpass the Netherlands in TFP until after 1800.22 Historian Robert Allen estimated TFP in agriculture by calculating the ratio between actual output per worker and the output per worker predicted by a country’s available agricultural population and land. On this metric for the year 1800, the Netherlands ranked higher than Britain and all other European nations.23 The Dutch also attained the highest absolute TFP in Europe for almost all of the eighteenth century.24
Which periodization of Britain’s industrial ascent better reflects the empirical evidence? On balance, measures of per-capita GDP, industrialization levels, and total factor productivity support a deferred timeline for Britain’s industrial rise. This clarification of when an economic power transition occurred during the IR-1 provides a stable target to test the competing LS and GPT mechanisms.
Key Technological Changes in the IR-1
Before evaluating the LS and GPT mechanisms in greater depth, the technological elements of the IR-1 must be further specified. Hargreaves’s spinning jenny (1764), Arkwright’s water frame (1769), Watt’s steam engine (1769), Cort’s puddling process (1784), and many other significant technical advances emerged during the First Industrial Revolution. The most likely sources of GPT and LS trajectories can be identified with guidance from existing work that calls attention to key technologies and accepted criteria for these two categories. Narrowing the assessment of these two mechanisms to a limited set of technologies makes for a more viable exercise.
Candidate Leading Sectors: Cotton Textiles, Iron, and Steam Power
A strong degree of consensus on the leading sectors that powered Britain’s rise in the IR-1 makes it relatively easy to identify three candidate sectors: cotton textiles, iron, and steam power.25 Among these, historians widely recognize the cotton textile industry as the original leading sector of the First Industrial Revolution.26 New inventions propelled the industry’s rapid growth, as its share of total value added to British industry rose from 2.6 percent in 1770 to 17 percent in 1801.27 In characterizing the significance of the cotton industry, Schumpeter went as far as to claim, “English industrial history can, in the epoch 1787–1842 … be almost resolved into the history of a single industry.”28
If the cotton textile industry places first in the canon of the IR-1’s leading sectors, then the iron industry follows close behind. In their account of Britain’s rise, Modelski and Thompson single out these two major industries, employing pig iron production and cotton consumption, as indicators for Britain’s leading sector growth rates.29 According to the traditional view of the IR-1, the iron and cotton industries were the only two that experienced “highly successful, rapidly diffused technical change” before the 1820s.30
I also evaluate the steam power industry as a third possible leading sector. A wide range of LS-based scholarship identifies steam power as one of the technological foundations of Britain’s leadership in the nineteenth century.31 Most of this literature labels only the steam engine itself as the leading sector, but since leading sectors are new industries, the steam engine–producing industry is the more precise understanding of the leading sector related to major advances in steam engine technology. Compared to the iron and cotton textile industries, it is much more uncertain whether the steam engine–producing industry, which experienced relatively slow growth in output and productivity, meets the analytical criteria for a leading sector during the IR-1 case.32 Still, I include the steam engine–producing industry as a potential leading sector, leaving it to the case analysis to further study its growth trajectory.
Candidate GPTs: Iron, Steam Engine, and the Factory System
Since the possible sources of GPT trajectories in the IR-1 are less established, I drew on previous studies that mapped out technological paradigms in this period to select three candidate GPTs: the steam engine, mechanization based on advances in iron machinery, and the factory system.33 As a possible source of GPT-style effects, the steam engine is a clear choice. Alongside electricity and (ICT) technology, it has been described as one of the “Big Three” GPTs, appearing in nearly all catalogs of GPTs.34 Here the emphasis is on the capacity of steam engines to transform a wide variety of industrial processes across many sectors, as opposed to the potential growth of the steam engine–producing industry.
Of the two paradigmatic industries of the IR-1, cotton and iron, the latter was a more plausible driver of GPT-style effects for Britain. As the demand for iron-made machinery grew, iron foundries development of new generations of machine tools, such as cylinder-boring machines, contributed to the creation of a mechanical engineering industry.35 This spurred the mechanization of production processes in a wide range of industries, including agriculture, food processing, printing, and textiles.36 Although both cotton textiles and iron were fast-growing industries, developments in iron better resembled a “motive branch” driving pervasive effects across the economy.37
In addition, the late eighteenth century saw the emergence of centralized factories, which significantly increased the scale of goods production. The factory system offered the potential to change the techniques of production across many industries. One widely cited classification scheme for GPTs picks out the factory system as an “organizational GPT” in the late eighteenth to early nineteenth century period.38 Other scholars describe this organizational innovation as “one of the most fundamental changes of ‘metabolism’ in the Industrial Revolution.”39
I also considered but ultimately decided against including developments in railroads as a candidate GPT. Among five core texts that classify GPTs across many historical periods, at least two highlighted the significance of the railroad to the IR-1.40 In my view, the railroad represented a disruptive advance, but it did not acquire the variety of uses to qualify as a GPT. Railways carried many types of freight and made new business models possible, but their function was limited to transport.41
Sources of LS and GPT Trajectories
Table 3.1 recaps the potential technological sources for both the GPT and LS mechanisms. It is important to clarify three points about the sorting process. First, it is notable but not surprising that the candidate leading sectors and GPTs draw from similar technological wellsprings. Both mechanisms agree that some inventions, like Cort’s puddling process for making wrought iron, mattered much more than others in terms of their impact on the economic balance of power.
Table 3.1 Key Sources of Technological Trajectories in the IR-1
Candidate Leading Sectors
Candidate GPTs
Cotton textile industry
Factory system
Iron industry
Mechanization
Steam engine–producing industry
Steam engine
Where the mechanisms separate is in how this process transpired. Cort’s puddling process and other ironmaking innovations, under the GPT model, are expected to follow an impact pathway characterized by three features: an extended lag before they affect productivity growth, the spread of mechanization across the economy, and widespread complementary innovations in many machine-using industries. Under the LS model, the same technological sources are expected to affect economic growth in a way that is lopsided in the early stages of development, fueled by monopolizing iron exports, and limited to technological innovations in the iron industry.
Second, it is still useful to classify three distinct candidate GPTs in the IR-1, despite the fact that developments in factory systems, mechanization, and steam engines were mutually reinforcing in many respects. Steam engines depended on the transition from hand-tool processes to machinery-based production systems; at the same time, the impact of steam engines on coal mining was to boost the iron industry, spurring mechanization. Yet a number of historians distinguish the expansion of mechanization in the British industrial revolution from transformations linked to the steam engine, arguing that the latter’s economic impact materialized much later than the former.42 Thus, while these candidate GPT trajectories are interconnected, it is still possible to locate various GPTs at different stages of their life cycle.
Third, not all of these technological changes necessarily had a decisive impact on Britain’s capacity to sustain higher productivity levels than its rivals during the period of interest. They are labeled as candidates for a reason. As this chapter will show, the steam engine did not achieve widespread diffusion until after Britain had already established economic leadership. When subjected to more rigorous empirical analysis, developments in some technologies may not track well with the proposed LS and GPT trajectories for this period.
GPT vs. LS Trajectories in the IR-1
Spelling out possible sources of technological trajectories in the IR-1 provides a bounded terrain for testing the validity of the GPT and LS mechanisms. By leveraging differences between the two mechanisms with respect to impact timeframe, phase of relative advantage, and breadth of growth, I derive three sets of opposing predictions for how technological changes translated into an economic power transition in this period. I then assess whether, and to what extent, the developments in the IR-1 supported these predictions.
OBSERVABLE IMPLICATIONS RELATED TO IMPACT TIMEFRAME
When did the revolutionary technologies of the IR-1 disrupt the economic balance of power? If the impact timeframe of the LS mechanism holds, then radical technical advances in the cotton textile, iron, and/or steam engine–producing industries should have substantially stimulated British economic growth shortly after the emergence of major technological advances in the 1760s and 1770s.43 Accordingly, scholars theorize that leading sectors propelled Great Britain to industrial superiority in the late eighteenth century.44 In line with this conception of a rapid timeline, Modelski and Thompson expect that the growth of two lead industries, cotton and iron, peaked in the 1780s.45
On the other hand, if the GPT mechanism was operational, the impact of major technological breakthroughs on Britain’s industrial ascent should have arrived on a more gradual timeline. Key advances tied to mechanization, steam power, and the factory system emerged in the 1770s and 1780s. Given that GPTs require a long period of delay before they diffuse and achieve widespread adoption, the candidate GPTs of the IR-1 should not have had substantial economy-wide repercussions until the early decades of the nineteenth century and after. I use the year 1815 as a rough cut-point to separate the accelerated impact timeframe of leading sectors from that of GPTs in this period.
OBSERVABLE IMPLICATIONS RELATED TO THE PHASE OF RELATIVE ADVANTAGE
The LS mechanism places high value on the innovation phase of technological change. Where major breakthroughs arise is key. Accordingly, Britain’s capacity to pioneer major technological advances should explain economic growth differentials in the IR-1. Concretely, the LS mechanism expects that Britain’s rise was fueled by its dominance of innovation in the cotton textile, iron, and steam engine–producing industries, as well as the resultant monopoly rents from exports in these sectors.
The GPT mechanism emphasizes a less-celebrated phase of technological change. Where innovations diffuse is key. Differentials in the rate and intensiveness of GPT adoption generate the gap between an ascending industrial leader and other competitors. The GPT mechanism suggests that Britain’s rise to industrial preeminence can be traced to its superior ability to diffuse generic technological changes across the economy.
OBSERVABLE IMPLICATIONS RELATED TO BREADTH OF GROWTH
The last set of observable implications relate to the breadth of growth during the IR-1. As illustrated in the descriptions here of the candidate leading sectors, many accounts attribute Britain’s industrial ascent to a narrow set of critical advances.46 In one of the first texts dedicated to the investigation of technology and international relations, William Ogburn declared, “The coming of the steam engine … is the variable which explains the increase of Britain as a power in the nineteenth century.”47 According to GPT diffusion theory, Britain’s rise to industrial preeminence came from the advance of GPTs through many linked sectors.
Taken together, these three sets of diverging predictions guide my assessment of the GPT mechanism against the LS mechanism. I make expectations specific to the IR-1 case by using the relevant information on particular technologies and the timeline of British industrialization. Table 3.2 lays out the specific, testable predictions that provide the framework of evaluation in the following sections.
Table 3.2 Testable Predictions for the IR-1 Case Analysis
Prediction 1: LS (impact timeframe)
The cotton textile, iron, and/or* steam engine–producing industries made a significant impact on Britain’s rise to industrial preeminence before 1815.
Prediction 1: GPT
Mechanization, the steam engine, and / or the factory system made a significant impact on Britain’s rise to industrial preeminence only after 1815.
Prediction 2: LS (relative advantage)
Innovations in cotton textile, iron, and/or the steam engine–producing industries were concentrated in Britain.British advantages in the production and exports of textiles, iron, and/or steam engines were crucial to its industrial superiority.
Prediction 2: GPT
Innovations in iron, the steam engine, and/or the factory system were not concentrated in Britain.
British advantages in the diffusion of mechanization, steam engines, and/or the factory system were crucial to its industrial superiority.
Prediction 3: LS (breadth of growth)
Productivity growth in Britain was limited to the cotton textile, iron, and/or steam engine – producing industries.
Prediction 3: GPT
Productivity growth in Britain was spread across a broad range of industries linked to mechanization, the steam engine, and/or the factory system.
* The operator “and/or” links all candidate leading sectors and GPTs because it could be the case that only some of these technologies drove the trajectories of the period.
Impact Timeframe: Delayed Surge vs. Fast Rise of British Industrialization
The painstaking reconstruction of temporal chronology is at the heart of tracing mechanisms. Tremendous technological changes occurred during this period, but when exactly did they make their mark on Britain’s industrial superiority? The period when significant technological innovations emerge often does not match up with the time when their impacts are felt. Unfortunately, when drawing lessons from the IR-1 on the effect of technology on international politics, scholars have conflated the overall significance of certain technologies with near-immediate impact.48 Establishing a clear timeline of when technological changes catalyzed a shift in productivity leadership during the IR-1 is therefore an important first step in comparing the LS and GPT mechanisms.
DIVERGING TIMELINES: COTTON VS. IRON
Time-series data on the output growth of twenty-six industries that accounted for around 60 percent of Britain’s industrial production help differentiate the growth schedules of the cotton textiles and iron industries. According to these data, the major upswing in British industrialization took place after 1815, when the aggregate growth trend increased from 2 percent to a peak of 3.8 percent by 1825. In line with the expectations of the LS model, the cotton textile industry grew exceptionally fast following major technological innovations in the 1760s, but from the 1780s there was a deceleration in the output growth of cotton textiles. Based on the relatively early peak of the cotton industry’s expansion, David Greasley and Les Oxley conclude that “it appears unlikely that cotton played the major role in the post-1815 upswing in British industrialization.”49
Following a completely different trajectory, growth in iron goods was more in line with the GPT model. Starting in the 1780s, the growth rate of the British iron industry accelerated to a peak of about 5.3 percent in the 1840s.50 “Compared to cotton textiles, change in iron was gradual, incremental, and spread out over a longer period of time,” Espen Moe writes.51 With a limited role for cotton, the gradual expansion of the iron industry led Britain’s post-1815 industrial surge, as its trend output tracked much more closely with that of aggregate industry. In sum, the cotton industry followed the growth path of a leading sector, whereas developments in the iron industry reflected the impact timeline of a GPT.
The timing of Britain’s mechanization, linked to the expanded uses of iron in machine-making, also aligned with the GPT trajectory. The first metalworking tools for precision engineering, including Wilkinson’s boring mill of 1774 and Maudslay’s all-iron lathe, appeared in the late eighteenth century, but they would remain in a “comparatively rudimentary state” until about 1815.52 According to accounts of qualified engineers and the 1841 Select Committee on Exportation of Machinery, over the course of the next two decades improvements and standardization in such machine tools ushered in a “revolution” in machine-making.53 The gradual evolution of the mechanical engineering industry provides additional support for a delayed impact timeframe for mechanization. According to British patent data from 1780 to 1849, the share of mechanical engineering patents among the total number of patents increased from an average of 18 percent in the decade starting in 1780 to a peak of 34 percent in the one starting in 1830.54
DELAYED, OUT-OF-PERIOD EFFECTS: STEAM ENGINE AND FACTORY SYSTEM
Compared to mechanization, the steam engine had not diffused widely enough through Britain’s economy by the mid-nineteenth century to make a substantial impact on overall industrial productivity. Detailed investigations of steam engine adoption have forced a reassessment of commonly held assumptions about the rapid impact of steam power on British productivity growth.55 According to one growth accounting analysis, which compares the impact of steam against water power as a close substitute source of power, steam power’s contribution to British productivity growth was modest until at least the 1830s and most influential in the second half of the nineteenth century.56 This revised impact timeframe conflicts with international relations scholarship, which advances faster trajectories for steam’s impact as a leading sector.57
Steam engine adoption was slow. In 1800, thirty years after Watt patented his steam engine, there were only about thirty-two engines operating in Manchester, which was a booming center of industrialization.58 Even into the 1870s, many important sectors in the British economy, such as agriculture and services, were virtually unaffected by steam, as most of steam power’s applications were concentrated in mining and in cotton textiles.59 During the period when LS accounts expect its peak growth, the steam engine could not claim generality of use.
The process by which the steam engine gained a broad variety and range of uses entailed complementary innovations that followed many years after the introduction of Watt’s steam engine. It took sixty years for steam to become the prime driver of maritime transport; that was made possible only after cumulative enhancements to the power of steam engines and the replacement of paddle wheels by screw propellers, which increased the speed of steam-powered ships.60 Watt’s original low-pressure design engines consumed large amounts of coal, which hindered widespread adoption. After 1840, aided by inventions such as the Lancashire boiler and discoveries in thermodynamics, it became economically viable to deploy steam engines that could handle higher pressures and temperatures.61 In sum, steam power may be the quintessential example of the long delay between the introduction of a GPT and significant economy-wide effects.
It is worth assessing whether interconnections between developments in steam and those in iron and cotton give grounds for an earlier impact trajectory.62 Yet both forward and backward linkages were limited in the early stages of the steam engine’s evolution. Regarding the latter, the steam engine–producing industry did not substantially stimulate the iron industry’s expansion. In the late 1790s, at a peak in sales, Boulton and Watt steam engines consumed less than 0.25 percent of Britain’s annual iron output.63 Forward linkages to textiles, the most likely sector to benefit from access to cheaper steam power, were also delayed. Steam-powered inventions in textiles did not overtake water-powered textile processes until after 1830.64 Of course, over the long run steam power was a critical technological breakthrough that changed the energy budget of the British economy.65 However, for investigating the mechanisms that facilitated Britain’s rise to become the clear productivity leader in First Industrial Revolution, the steam engine played a modest role and most of its effects were out-of-period.
A similar timeline characterizes the progression of the factory system, another candidate GPT considered for this period. The factory system diffused slowly and took hold in only a limited number of trades during the time when Britain was establishing its industrial preeminence. The British textile industry, as the earliest adopter of this organizational innovation, had established nearly five thousand steam- or water-powered factories by the 1850s.66 Other industries, however, were much slower to adopt the factory system. In the first decades of the nineteenth century, small workshops and domestic production still dominated the metal trades as well as other hardware and engineering trades.67
Moreover, factories were still relatively small even into the mid-nineteenth century, and some industries adopted a mixed factory system in which many processes were outsourced to household workers.68 It was not until steam power overtook water power in the 1830s and 1840s as a source of power in factories that the subsequent redesigns of factory layouts led to large gains in productivity.69
What does this clarified chronology of technological impacts in the IR-1 mean for the explanatory power of the GPT and LS mechanisms? Of the three candidate leading sectors, only cotton textiles, which expanded rapidly and peaked in terms of output growth in the 1780s and 1790s, followed the impact timeframe of a leading sector. As figure 3.2 shows, by 1814 British cotton exports had already surpassed 75 percent of the value they would attain in 1840. Yet Britain sustained productivity growth rates at higher levels than its rivals only in the first decades of the nineteenth century. Thus, the period when cotton should have made its greatest impact on Britain’s industrial ascent does not accord with the timeline of Britain’s industrialization surge.
The hurried timeline of the LS mechanism contrasts with the more delayed impact of other technological advances. As predicted by the GPT mechanism, all three candidate GPTs—mechanization, the steam engine, and the factory system—had little impact on Britain’s industrial rise until after 1815. In fact, the diffusion timelines for the steam engine and factory system were so elongated that their impact on Britain’s rise to industrial preeminence was limited in this period. In 1830, steam engine adoption, as measured by total horsepower installed, was only one-quarter of the 1840 level, whereas by 1830 the level of iron production had reached about 50 percent of its corresponding value in 1840 (see figure 3.2). This is consistent with the steady expansion of mechanization across industry in the early decades of the 1800s as the GPT trajectory most attuned with the timing of Britain’s rise to economic leadership.
figure 3.2 Technological Impact Timeframes in the IR-1. Note: British cotton exports, iron production, and steam engine adoption over time. Source: Robson 1957, 331–35; Mitchell 1988, 280–81; Crafts 2004, 342.
Phase of Relative Advantage: Diffusion of Iron vs. Monopoly Profits from Cotton
Thus far, the empirical evidence has presented a bird’s-eye view of the overall timeline of technological change and industrialization in the IR-1, but there are two other dimensions on which the GPT and LS trajectories diverge. According to the expectations of the LS mechanism, the phase of technological development central to Britain’s relative economic rise was its dominance of key innovations in cotton textiles, iron, or the steam engine. The GPT mechanism predicts, in contrast, that Britain’s advantage in the diffusion of mechanization, the steam engine, or the factory system was the key driver.
The rest of this section tests two sets of predictions derived from the diverging assumptions of the two mechanisms. First, regarding the geographic clustering of major technological breakthroughs in the IR-1, I assess whether innovations in candidate leading sectors and GPTs were concentrated in Britain. Next, regarding the comparative consequences of these technologies, I evaluate whether Britain’s industrial superiority drew more from its advantages in the production and exports of the IR-1’s leading sectors or from its advantage in the diffusion of the IR-1’s GPTs.
INNOVATION CLUSTERING IN THE IR-1’S BREAKTHROUGHS
Did Britain dominate innovation in the leading sectors of the IR-1? At first glance, there is no question that radical advances in candidate leading sectors clustered in Britain. This list includes Watt’s steam engine, Arkwright’s water frame, Cort’s puddling process, and many more. Per one analysis of 160 major innovations introduced during the nineteenth and twentieth centuries, Britain was home to 44 percent of major innovations from 1811 to 1849—a rate that was double that of the closest competitor (the United States at 22 percent).70
Further investigation into British superiority in technological innovation paints a more mixed picture. According to another list of technical advances by country of origin, Britain accounted for only 29 percent of major innovations in the years from 1826 to 1850, a period that corresponds to when it cemented its productivity leadership.71 Moreover, the European continent introduced many significant innovations, including the Jacquard loom, mechanical flax spinning, chlorine bleaching, the Leblanc soda–making process, and the Robert continuous papermaking machine.72 France, in particular, generated many of the major industrial discoveries, such as in chemicals, clocks, glass, papermaking, and textiles.73
Thus, some scholars argue that Britain’s comparative edge was in more incremental improvements. Reflecting on technological creativity in the IR-1, economic historian Joel Mokyr argues, “Britain seems to have no particular advantage in generating macroinventions … the key to British technological success was that it had a comparative advantage in microinventions.”74 A proverb from the time captured this distinction: “For a thing to be perfect it must be invented in France and worked out in England.”75 This suggests that digging deeper into the different phases of technological development can help uncover the roots of Britain’s industrial leadership.
MONOPOLY PROFITS VS. DIFFUSION DEFICIT
First, as was the case with the period when they made their impact, developments in cotton and iron followed very different paths with respect to the phase of technological change that determined economic differentials. Britain’s cotton textile industry, the most likely source of monopoly profits, grew faster than other industries before 1800, and it sold most of its goods abroad. Technological innovations such as the spinning jenny and the water frame triggered exponential increases in the efficiency of cotton production, and Britain’s cotton output increased by 2,200 percent from 1770 to 1815.76 From 1750 to 1801, cotton’s share of Britain’s major exports increased from 1 percent to 39.6 percent.77
Certainly, the growth of British cotton exports was remarkable, but what was the impact of associated monopoly rents on overall growth differentials? Supported by improved quantitative estimates of the cotton industry’s impact on the British economy, historians generally accept that the cotton industry was much more significant for enhancing Britain’s trade balance than for boosting its economic productivity.78 According to one estimate, between 1800 and 1860, the cotton industry accounted for 43 percent of the threefold increase in the value of exports but only 8 percent of the threefold increase in national income.79
Overall, exports constituted a small proportion of British economic activity during the IR-1. From 1770 to 1841, British exports as a percentage of overall industrial demand increased only from 13 to 16 percent.80 Now, these figures probably underrate trade as a critical engine of growth for Britain in the IR-1, as they ignore gains from the reinvestment of profits from overseas trade.81 But the impact of reinvestment has been challenged, and it is not apparent why reinvestments from exports were more important than reinvestments from the profits generated by domestic production.82
The iron industry’s impact on Britain’s economic rise did not run through monopoly profits from innovation. From 1794 to 1796, British ironmakers contributed 11 percent of Britain’s manufacturing exports. This proportion actually declined to just 2 percent by the 1814–1816 period and stayed around that rate into the 1830s.83 It is also questionable whether Britain held a relative advantage in iron exports during the late eighteenth century, which is when LS accounts expect monopoly profits to drive British industrialization.84 In fact, British industries continued to rely on imports of high-grade iron from Sweden and Russia well into the nineteenth century.85
An alternative pathway, captured by the GPT trajectory, posits that Britain’s advantage came from the diffusion of iron machinery advances across a wide range of sectors. To trace this trajectory, it is necessary to pay more attention to what a prominent historian of the IR-1 calls one of the astonishing things about the phenomenon: the gap between “innovation as ‘best practice’ technique and the diffusion of innovation to become ‘representative’ technique.”86
Britain was more successful than its industrial rivals in the diffusion of mechanization. Contemporary observers from the European continent often remarked upon Britain’s ability to bridge this gap between best practice and representative practice.87 Writing in 1786 in their Voyages aux Montagnes, French observers F. and A. de la Rochefoucauld-Liancourt, commenting on Britain’s relative advantage in the widespread adaptation of the use of iron, noted
the great advantage [their skill in working iron] gives them as regards the motion, lastingness and accuracy of machinery. All driving wheels and in fact almost all things are made of cast iron, of such a fine and hard quality that when rubbed up it polishes just like steel. There is no doubt but that the working of iron is one of the most essential of trades and the one in which we are most deficient.88
But France’s deficiency in iron machinery was not a product of its lack of access to key innovations. In fact, France was the world’s center of science from the late eighteenth century until the 1830s.89 Rather, as the following quote illustrates, Britain’s industrial rivals fell behind in “diffused average technology” and the “effective spread of technical change more widely.” Economic historian Peter Mathias writes:
It is remarkable how quickly formal knowledge of “dramatic” instances of new technology, in particular steam engines, was diffused, and how quickly individual examples of “best-practice” technology in “show piece” innovations were exported. The blockage lay in the effective spread of technical change more widely—diffused average technology rather than single instances of best-practice technology in “dramatic” well-publicized machines.90
Advances in iron metallurgy played a crucial role in a GPT trajectory that spread from a sector that improved the efficiency of producing capital goods. The GPT trajectory unfolds as the technology becomes more general-purpose through interactions between the upstream capital goods sector and the user industries that enlarge the range of applications. Rosenberg’s depiction of this type of system highlights the nineteenth-century American machine tool industry as the innovative capital goods sector.91 In this case, Britain’s metal-processing works were the crucial wellspring. Specifically, technical advances in iron fed into metalworking industries from which broadly similar production processes diffused over a large number of industries.92 Maxine Berg, a professor of history at the University of Warwick, pinpoints these industries as the “prime mechanism for technological diffusion.”93
Scholars also identify Watt’s improved steam engine as a potential source of both LS- and GPT-based effects. Here I focus on testing the LS prediction about the steam engine–producing industry because the previous section showed that the steam engine and the factory system, as candidate GPTs, diffused too slowly to make a meaningful impact on the economic power transition in the IR-1.
It is tough to make a case that the growth of the steam engine–producing industry generated a substantial source of monopoly profits for Britain. Equipped with an exclusive patent, James Watt and Matthew Boulton set up a firm in 1775 to sell steam engines.94 In the period from 1775 to 1825, however, the firm sold only 110 steam engines to overseas customers.95 By 1825, France and the United States were manufacturing the Watt engine and high-pressure engines at more competitive prices, and overseas demand declined sharply.96 Thus, the international sales history of this firm severely weakens the significance of the monopoly profits associated with the innovation of the steam engine.97
In sum, the evidence from this section supports two conclusions. British advantages in the production and export of iron, steam engines, and cotton textiles (the best representative of the LS trajectory) had muted effects on its overall industrialization and productivity advances. Second, the contributions of technological breakthroughs in iron metallurgy and steam power to Britain’s industrial rise track better with the GPT mechanism, based on relative advantages in widespread technological diffusion as opposed to monopoly profits from innovation.
Breadth of Growth: Complementarities of Iron vs. Spillovers from Cotton Textiles
The breadth of growth in the IR-1 is the last dimension on which the LS and GPT trajectories disagree. Was Britain’s industrial rise driven by technological changes confined to a narrow range of leading sectors, or was it based on extensive, complementary innovations that enabled the spread of GPTs? Making use of data on sectoral sources of productivity growth, trade flows, and patents, I evaluate these competing propositions about the breadth of technological change in the industrial revolution.
WIDESPREAD PRODUCTIVITY GROWTH
In differentiating between the narrow view and the broad view of technical change during the IR-1, a natural starting point is to estimate the contribution of various industries to British productivity growth. Deirdre McCloskey’s calculations of sectoral contributions to productivity growth support the broad view. Though cotton accounted for a remarkable 15 percent of Britain’s total productivity growth, nonmodernized sectors still drove the lion’s share (56 percent) of productivity gains.98
Manufacturing trade data provide another testing ground. If other manufacturing industries outside of textiles and iron were technologically stagnant during the first fifty years of the nineteenth century, then British competitiveness in these industries should decline relative to textiles and iron. The narrow view implies that Britain should have imported other manufactures. Peter Temin’s analysis of British trade data, however, finds the opposite. Throughout the first half of the nineteenth century, British manufacturing exports matched the increase in cotton exports throughout the first half of the nineteenth century.99 In a wide range of manufactures, such as arms and ammunition, carriages, glass, and machinery and metals, Britain held a clear comparative advantage. This pattern points to some general pattern of change that spanned industries. “The spirit that motivated cotton manufactures extended also to activities as varied as hardware and haberdashery, arms, and apparel,” Temin concludes.100
The patent record also depicts a landscape of extensive technological change.101 From 1780 to 1840, about 80 percent of all patented inventions came from outside the textiles and metals industries.102 Per Christine MacLeod’s data on British patents covering the 1750–1799 period, most capital goods patents originated from sectors outside of textile machinery and power sources.103 As summed up by historian Kristine Bruland, the historical evidence supports “the empirical fact that this was an economy with extensive technological change, change that was not confined to leading sectors or highly visible areas of activity.”104
GPTS AND COMPLEMENTARY INNOVATIONS
At this point, indicators of the multisectoral spread of innovation in the IR-1 should not be sufficient to convince a skeptical reader of the GPT mechanism’s validity. Broad-based growth could be a product of macroeconomic factors, such as sound fiscal and monetary policy or labor market reforms, rather than a GPT trajectory.105 Proving that the dispersion of technological change in Britain’s economy reflected a GPT at work requires evidence that connects this broad front to mechanization.106
Input-output analysis, which sheds light on the linkages between industries, suggests that improvements in the working of iron had broader economic significance. To better understand the interrelationships among industries during the industrial revolution, Sara Horrell, Jane Humphries, and Martin Weale constructed an input-output table for the British economy in 1841. Across the seventeen industries included in the analysis, the two industries most closely associated with mechanization—metal manufacture and metal goods—scored the highest on combined backward and forward linkages.107 These two domains were “lynchpins of linkage effects.”108
Patent indicators confirm these results. When patents are grouped according to standard industry taxonomies, the resulting distribution shows that the textile industry contributed to 15 percent of the patents issued between 1711 and 1850, making it the most inventive industry in aggregate terms.109 However, when patents are sorted by general techniques as opposed to industry sectors, the same data reveal the underlying drive force of mechanical technology: it is linked to almost 50 percent of all British patents during this period.110
Along all three dimensions of technological trajectories in the IR-1, the process-tracing evidence bolsters the validity of the GPT mechanism. First, slower-moving developments in mechanization lined up with a delayed timeline of Britain’s industrialization. Other candidate leading sectors and GPTs either peaked too early (cotton) or got started too late (steam engine, factory system). Second, Britain gained its industrial dominance from a relative advantage in widespread adoption of iron metalworking and linked machinery. Third, the benefits from this GPT advantage circulated throughout the economy, rather than remaining concentrated in the iron industry.
The standard explanation of how the IR-1 gave rise to a power transition, as captured by the LS mechanism, analyzes technological change at the level of industries that grow faster than others. The historical evidence reveals the limitations of these industry taxonomies. Instead, advantages in the diffusion of production machinery—a general pattern of change that extended across a wide range of economic activities—propelled Britain to industrial dominance.111
Institutional Complementarities: GPT Skill Infrastructure in the IR-1
Having mapped Britain’s industrial rise to a GPT trajectory linked to mechanization, there is still a need to explain why Britain was best positioned to exploit this trajectory. If other countries at the technological frontier can also cultivate mechanical innovations at home and absorb them from abroad, why were Britain’s competitors unable to benefit from the diffusion of metalworking processes to the same extent? This section supports an explanation based on Britain’s institutional competencies in widening the pool of engineering skills and knowledge linked to mechanization.
Which types of institutions for skill provision were most conducive to national success in the IR-1? One common refrain is that Britain’s leadership was rooted in the genius of individual innovators like James Watt, and such genius did not transfer as quickly across borders during the IR-1.112 Though recent scholarship has weakened this view, many influential histories center on the “heroic” inventors of the industrial revolution.113 Consistent with the LS template, this view focuses on the institutions that helped drive heroic invention in Britain, such as the development of a patent system.114
The pathway by which mechanization propelled Britain’s industrial ascent, established as a GPT trajectory in the previous section, emphasizes another set of institutions for skill formation. In line with GPT diffusion theory, Britain owed its relative success in the IR-1 to mechanics, instrument-makers, and engineers who could build machines according to blueprints and improve upon them depending on the application context. Under this view, the institutions that trained the “tweakers” and “implementers,” rather than those that cultivated genius inventors, take center stage.115
Widening the Base of Mechanical “Tweakers” and “Implementers”
At first, rapid advances in precise metalworking exposed a skills shortage in applied mechanics. Beginning in the 1770s, a cascade of recruitment advertisements in local newspapers sought out an “engine-maker” or a “machine-maker.”116 Reflecting on this skills mismatch, the president of Britain’s Institute of Civil Engineers stated that the use of cast iron in machine parts “called for more workmen than the millwright class could supply.”117
A number of institutional adjustments helped Britain meet this demand for mechanically skilled tweakers and implementers. Initially, Britain benefited from a flexible apprenticeship system that empowered workers in related domains to get trained in applied mechanics.118 Thus, to develop the workforce to build and maintain the machinery of the IR-1, Britain could draw from a wide pool of blacksmiths, millwrights, gunsmiths and locksmiths, instrument-makers, mechanics, and toolmakers.119
In addition, institutes dedicated to broadening the base of mechanical expertise helped diffuse ironmaking and machine-making skills. Starting in the 1790s, private and informal initiatives created a flurry of trade associations that supported a new class of mechanical and civil engineers and helped connect them with scientific societies and entrepreneurs.120 Critical centers included the Andersonian Institution in Glasgow, the Manchester College of Arts and Sciences, the School of Arts in Edinburgh, the Mechanical Institution in London, the Society for the Diffusion of Useful Knowledge, and hundreds of mechanics’ institutes.121 These institutes helped to absorb knowledge from foreign publications on science and engineering, recruit and upskill new mechanical engineers from a variety of trades, and spread mechanical engineering knowledge more widely.122
It is important to note that these institutional features differed from those more suited to the “heroic inventor” model. In Britain’s cotton textile industry, the classic leading sector of the IR-1, the key institutional complements deviated greatly from the education and training systems that widened the base of mechanical expertise in the IR-1. Collating information from biographies of British engineers, online databases, and detailed economic histories, Ralf Meisenzahl and Joel Mokyr constructed a database of 759 British individuals who made incremental improvements to existing inventions during the industrial revolution.123 Notably, based on their analysis of interactions between tweakers and their institutional surroundings, they found that the textile industry was an outlier in terms of protectiveness over intellectual property rights and reluctance to share information about new techniques. Less than one-tenth of tweakers in textiles published their knowledge to a broader audience or joined professional societies, in stark contrast to the two-thirds of tweakers in mechanically inclined fields who did so.124 Over 80 percent of the tweakers who were active primarily in textiles took out at least one patent, compared to just 60 percent for tweakers overall.125
These trends in applied mechanics underscore the significance for British mechanization of “collective invention,” a process that involved firms sharing information freely with one another and engineers publishing technical procedures in journals to spur the rapid diffusion of best-practice techniques. According to one analysis of how various districts adapted to the early phases of industrialization, areas that practiced collective invention often cultivated “a much higher degree of technological dynamism than locations which relied extensively on the patent system.”126
Britain’s Comparative Advantage over France and the Netherlands in GPT Skill Infrastructure
Britain’s competitors also grasped the significance of Britain’s wide pool of mechanical skills. Whereas codified knowledge crisscrossed western Europe and North America via patent specifications, global exchanges among scientific societies, and extensive visits by foreign observers to British workshops and industrial plants, the European continent struggled greatly to absorb tacit knowledge, especially the know-how embodied in the practical engineering skills of British mechanical tweakers and implementers.127 France and the Netherlands fiercely poached British engineers, as the transfer of tacit knowledge in the fields of large-scale ironworking and machine construction almost always necessitated the migration of skilled workers from Britain.128 “It was exactly in the skills associated with the strategic new industries of iron and engineering that [Britain’s] lead over other countries was most marked,” argues Mathias.129
Why did this repository of engineering skills develop more fruitfully in Britain than in its industrial rivals? A growing body of evidence suggests that Britain’s institutions better adapted its distribution of skills to mechanization. Britain’s institutional advantage was rooted in the system of knowledge diffusion that connected engineers with entrepreneurs, cities with the countryside, and one social class with another. Institutes that trained mechanics took part in a broader “mushrooming of associations” that spread technical knowledge in early-nineteenth-century Britain.130 By the mid-nineteenth century, there were 1,020 such associations in Britain, with a total membership of approximately 200,000; clearly, these networks are essential to any explanation that links human capital to Britain’s industrial ascent.131 Compared to their peers on the continent, British mechanics had superior access to scientific and technical publications.132 As a result, the British system of the early nineteenth century had no match in its abundance of people with “technical literacy.”133
The French system, by way of comparison, lacked similar linkages and collaborations between highly educated engineers and local entrepreneurs.134 Though France produced elite engineers at schools like the École Polytechnique, it trained too few practitioners to widen the base of mechanical skills.135 For example, Napoleon’s early-nineteenth-century reform of France’s higher education system encouraged the training of experts for narrow political and military ends, thereby limiting the ability of trainees to build connections with industry.136 These reforms and other industrial policies directed French engineers toward projects associated with luxury industries and specialized military purposes, which “tended to become locked away from the rest of the economy in special enclaves of high cost.”137 To illustrate, through the mid-1830s, only one-third of École Polytechnique graduates entered the private sector.138 France’s system for disseminating mechanical knowledge and skills was vastly inferior to that of the British.
The Netherlands also failed to develop a base of mechanical skills that linked scientific research to practical ends. In some mechanical sciences, the Dutch generated plenty of potentially useful innovations, even pioneering key breakthroughs that eventually improved the steam engine.139 Yet the Dutch struggled to translate these scientific achievements into practical engineering knowledge because they trailed the British in forming institutional settings that made widespread knowledge of applied mechanics possible. Records of Dutch educational systems, the dearth of societies that held lectures and demonstrations for mechanical learning, and the materials available at libraries in technical colleges all “reflected a profound lack of interest in applied mechanics.”140 In his study of Dutch technological leadership, Karel Davids argues that, during the first three-quarters of the eighteenth century, “collaboration between science and industry in the Netherlands failed to merge in the very period that relations between the two became rapidly closer in Britain.”141
Britain’s advantage in GPT diffusion was not rooted in its higher education system, which lagged far behind the French education system during the IR-1 period.142 France had already established more than twenty universities before the French Revolution. The French system of higher technical education, from the late eighteenth century through the 1830s, had no rival. The Grande Écoles system, including the elite École Polytechnique (established in 1794), trained expert scientists and engineers to take on top-level positions as industrial managers and high-level political personnel.143 Up until 1826, England had set up only two universities, Oxford and Cambridge. These institutions made limited contributions to training the workforce necessary for industrialization. One study with a sample of 498 British applied scientists and engineers born between 1700 and 1850 found that only 50 were educated at Oxford or Cambridge; 329 were not university-educated.144
At this point, curiosity naturally leads us to ask why Britain accumulated an advantage in GPT skill infrastructure. Due to practical constraints of time and space, I acknowledge but do not delve into the deeper causes for the notable responsiveness of Britain’s institutions to the skill demands of mechanization. In surveying valuable lines of inquiry on this subject, chapter 2 points to government capacity to adopt long time horizons and reach intertemporal bargains. In the IR-1 case, two specific factors are also worthy of consideration. Attributing Britain’s later success to pre-industrial training practices, some studies suggest that Britain’s apprenticeship system allowed for agile and flexible adaptation to fluctuations in the demand for skills, especially in mechanical trades.145 Looking even further back, other scholars probe the geographical origins of Britain’s mechanical skills, underscoring the lasting effects of Britain’s adoption of watermills in the early Middle Ages.146
Alternative Explanations of Britain’s Rise
The history of the IR-1 is certainly not a neglected topic, and the literature features enthusiastic debates over a wide range of possible causes for Britain’s rise. Prominent explanations tie Britain’s early industrialization to population growth,147 demand and consumption standards,148 access to raw materials from the colonies,149 slavery,150 and trade.151 The obvious concern is that various contextual factors may confound the analysis of the LS and GPT mechanisms.
I am not rewriting the history of the IR-1. I am drawing from one particularly influential and widely held view of the IR-1—that technological advances drove Britain’s industrial ascent—and investigating how technological change and institutional adaptations produced this outcome. The most relevant alternative factors, therefore, are those that provide a different interpretation of how technologies and institutions coevolved to result in Britain’s industrial hegemony. Although I primarily focus on the LS mechanism as the most formidable alternative explanation to the GPT diffusion theory, other explanations also warrant further investigation.
Threat-Based Explanations
Threat-based theories assert that external threats are necessary to incentivize states to innovate and diffuse new technologies. Did Britain owe its technological leadership to war and its military’s impetus to modernize? During the IR-1 period, Britain was embroiled in the Revolutionary and Napoleonic Wars (1793–1815), a near-continuous stretch of conflicts involving France and other European states. If threat-based explanations stand up in the IR-1 case, then the historical record should show that these wars made an essential and positive contribution to Britain’s adoption of iron machinery and mechanization.
Some evidence supports this argument. By 1805, the British government’s needs for iron in the war effort accounted for 17 percent of the total British iron output in 1805.152 This wartime stimulus to iron production facilitated improvements in iron railways, iron ships, and steam engines.153 In particular, military investments in gunmaking produced important spin-offs in textiles and machine tools, most famously encapsulated by Watt’s dependence on John Wilkinson’s cannon boring techniques to make the condenser cylinders for his steam engine.154
On the flip side, war’s disruptive costs are likely to have offset any stimulus to Britain’s mechanization. Aside from Wilkinson’s cannon boring device and some incremental improvements, wartime pressures did not produce any major technological breakthroughs for the civilian economy.155 Military needs absorbed productive laborers from Britain’s civilian economy, resulting in labor shortages.156 War also limited both the domestic demand for iron, by halting investment in construction, agriculture, and other industries, and the foreign demand for British iron, by cutting off foreign trade. Historian Charles Hyde notes, “In the absence of fighting, overall demand for iron might have been higher than it was.”157
Furthermore, any temporary benefits that accrued to Britain’s iron industry in wartime were wiped out in the transition to peacetime. In one influential text, historian Thomas Ashton reviewed how each of the wars of the eighteenth century, including the Revolutionary and Napoleonic Wars, affected Britain’s iron industry.158 He observed a similar pattern in each case. At first, the outbreak of hostilities boosts demand for iron in the form of armament, and trade disruptions protect domestic producers against foreign competitors. This initial boom is followed by a severe crash, however, when the iron industry adjusts to the conflict’s aftermath. A trade depression follows. Converting foundries to make plowshares instead of cannons incurs heavy losses, made even more painful by the fact that war conditions promoted “feverish” developments that were unsustainable in the long run.159
On a more fundamental level, threat-based theories have limited leverage in explaining Britain’s relative rise because its economic competitors were also embroiled in conflicts—in many cases, against Britain. The Dutch fought Britain in the fourth Anglo-Dutch War (1780–1784) as well as in the Napoleonic Wars.160 Of course, during this time, France was Britain’s main military opponent. Thus, since the Netherlands and France also faced a threatening external environment, the net effect of the war on economic growth differentials should have been minimal.161 If anything, since France fought on many more fronts than Britain during this period, proponents of threat-based explanations would expect France to have experienced more effective and widespread diffusion of iron machinery throughout its economy. The case analysis clearly discredits that expected outcome.
VoC Explanations
Can Britain’s particular brand of capitalism account for its technological rise? The varieties of capitalism (VoC) approach posits that liberal market economies (LMEs) are particularly suited to radical innovation. Consistent with this framework, international political economy scholars emphasize that Britain’s free market economy supported gains in rapidly changing technological domains like consumer goods, light machine tools, and textiles.162 During the IR-1 period, Britain began to develop the institutional features that would cement it as a LME, including decentralized collective bargaining and high levels of corporatization.163 Most pertinent to GPT diffusion theory, VoC scholars expect LMEs like Britain to excel at cultivating general skills, which help transfer GPT-related knowledge and techniques across firms.
Taking measure of Britain’s human capital development in general skills in this period is therefore central to evaluating whether its technological leadership can be explained by its form of capitalism. Overall, estimates of literacy rates and school attendance demonstrate that the general level of human capital in Britain was notably low for an industrial leader.164 British literacy rates for males were relatively stagnant between 1750 and 1850, and average literacy rates in Britain were much lower than rates in the Netherlands and barely higher than those in France around the turn of the nineteenth century.165 In fact, general levels of educational attainment in Britain, as measured by average years of schooling, declined from around 1.4 years in 1740 to 1.25 years in 1820.166 Contrary to VoC theory’s expectations, Britain did not hold an advantage in general skills during this period.
The VoC explanation’s applicability to the IR-1 period is further limited by issues with designating Britain as the only LME in this period. Like Britain, the Netherlands functioned as a relatively open economy and exhibited tendencies toward economic liberalism, but it was not able to adapt and diffuse significant technological changes.167 Though France is now considered a coordinated market economy, in the early nineteenth century it took on some of the characteristics of LMEs by implementing capital market reforms and trade liberalization.168 The VoC approach therefore struggles to resolve why these two LMEs diverged so greatly in their adaptation to mechanization.
Case-Specific Factors
Among other factors specific to the IR-1 setting, one alternative explanation emphasizes Britain’s fortunate geographic circumstances. More specifically, classic works have argued that proximity to plentiful coalfields was essential to British industrialization.169 These natural resource endowments enabled the expansion of coal-intensive industries, such as the iron industry. In this line of thinking, the fact that coal was cheaper in Britain than elsewhere in Europe explains why Britain was the first to sustain productivity leadership.170
The relationship between coal and industrialization does not necessarily undermine the GPT mechanism. For one, in principle, Britain’s competitors could also have effectively leveraged coal resources. The southern provinces of the Netherlands were located close to Belgian coalfields.171 Over the course of the eighteenth century, Dutch industry had mostly shifted to coal, and away from peat stocks, as a key source of energy.172 Even if Britain’s industrial rivals had to pay more by importing coal, the expected productivity gains associated with adopting new technologies should have outweighed these costs. Moreover, GPT skill infrastructure could have mediated the relationship between coal and mechanization, as Britain’s edge in metalworking skills spurred the adoption of new coal-using technologies, which strengthened the connection between proximity to coal and economic growth.173
Summary
In many ways, the industrial revolution marked an exceptional transformation. It is to any number of historical trends what the birth of Jesus is to the Gregorian calendar—an inflection point that separates “before” and “after.” For my purposes, however, the industrial revolution is a typical case showing how technological revolutions influence the rise and fall of great powers. Evidence from great powers’ different adaptations to technological changes in this period therefore helps test GPT diffusion theory against the LS mechanism.
In sum, GPT diffusion theory best explains why Britain led Europe’s industrial transformation in this period. Britain effectively capitalized on general-purpose improvements in mechanization owing to its institutional advantages that were conducive to widening the pool of mechanical skills and knowledge. According to GPT diffusion theory, countries like this disproportionately benefit from technological revolutions because they adapt more successfully to the GPT trajectories that transform productivity. In line with these expectations, Britain was more successful than its industrial rivals in sustaining long-term economic growth, which became the foundation of its unrivaled power in the early and mid-nineteenth century.
On the flip side, this chapter’s case analysis undercuts the LS-based explanation. The timeframe for when leading sectors were expected to stimulate Britain’s productivity growth did not align with when Britain’s industrialization took off. Britain’s economic ascent owed more to the widespread adoption of iron metalworking and linked production machinery than to monopoly profits from cotton textiles. The key institutional complements were not those that produced heroic inventions—Britain’s rivals held their own in these areas—but rather those that fostered widespread knowledge of applied mechanics.
Do these findings hold in other periods of technological and geopolitical upheaval? The IR-1 was one of the most extraordinary phases in history, but it was not the only era to attain the title of an “industrial revolution.” To further explore these dynamics, it is only appropriate to turn to the period some have labeled the Second Industrial Revolution.
4 The Second Industrial Revolution and America’s Ascent
IN THE LATE nineteenth and early twentieth centuries, the technological and geopolitical landscape transformed in ways familiar to observers of today’s environment. “AI is the new electricity” goes a common refrain that compares current advances in machine intelligence to electrical innovations 150 years ago. Those fundamental breakthroughs, alongside others in steel, chemicals, and machine tools, sparked the Second Industrial Revolution (IR-2), which unfolded from 1870 to 1914.1 Studies of how present-day technological advances could change the balance of power draw on geopolitical competition for technological leadership in the IR-2 as a key reference point.2
Often overshadowed by its predecessor, the IR-2 is equally important for investigating causal patterns that connect technological revolutions and economic power transitions. The presence of both cause and outcome ensures a fruitful test of the GPT and LS mechanisms. The beginning of the period featured remarkable technological innovations, including the universal milling machine, the electric dynamo, the synthesis of indigo dye, and the internal combustion engine. According to some scholars, one would be hard-pressed to find another period with a higher density of important scientific advances.3 By the end of the period, Britain’s decline and the rise of Germany and the United States had yielded a new balance of economic power, which one historian describes as a “shift from monarchy to oligarchy, from a one-nation to a multi-nation industrial system.”4 Arguably, British industrial decline in the IR-2 was the ultimate cause of World War I.5
International relations scholars hold up the IR-2 as a classic case of a power transition caused by LS product cycles.6 According to this view, Britain’s rivals cornered market shares in the new, fast-growing industries arising from major technological innovations in electricity, chemicals, and steel.7 Specifically, scholars argue that Germany surpassed Britain in the IR-2 because it was “the first to introduce the most important innovations” in these key sectors.8 Analysis of emerging technologies and today’s rising powers follows a similar template when it compares China’s scientific and technological capabilities to Germany’s ability to develop major innovations in chemicals.9 Thus, as a most likely case for the LS mechanism, which is favored by background conditions and existing theoretical explanations, the IR-2 acts as a good test for the GPT mechanism.
Historical evidence from this period challenges this conventional narrative. No country monopolized innovation in leading sectors such as chemicals, electricity, steel, and motor vehicles. Productivity growth in the United States, which overtook Britain in productivity leadership during the IR-2, was not dominated by a few R&D-based sectors. Moreover, major breakthroughs in electricity and chemicals, prominent factors in LS accounts, required a protracted process of diffusion across many sectors before their impact was felt. This made them unlikely key drivers of the economic rise of the United States before 1914.
Instead, the IR-2 case evidence supports GPT diffusion theory. Spurred by inventions in machine tools, the industrial production of interchangeable parts, known as the “American system of manufacturing,” embodied the key GPT trajectory.10 The United States did not lead the world in producing the most advanced machinery; rather, it had an advantage over Britain in adapting machine tools across almost all branches of industry. Though the American system’s diffusion also required a long gestation period, the timing matches America’s industrial rise. Incubated by the growing specialization of machine tools in the mid-nineteenth century, the application of interchangeable parts across a broad range of manufacturing industries was the key driving force of America’s relative economic success in the IR-2.11
Since a nation’s efficacy in adapting to technological revolutions is determined by how well its institutions complement the demands of emerging technologies, the GPT model of the IR-2 highlights institutional factors that differ from those featured in standard accounts. LS-based theories tend to highlight Germany’s institutional competencies in scientific education and industrial R&D.12 In contrast, the case analysis points toward the American ability to develop a broad base of mechanical engineering skills and standardize best practices in mechanical engineering. Practice-oriented technical education at American land-grant colleges and technical institutes enabled the United States to take better advantage of interchangeable manufacturing methods than its rivals.
This chapter’s evidence comes from a variety of sources. In tracing the contours of technological trajectories and the economic power transition in this period, I relied on histories of technology, categorization schemes from the long-cycle literature, general accounts of economic historians, and revised versions of historical productivity measures. I investigated the fit between institutions and technology in leading economies using annual reports of the US Commissioner of Education, British diplomatic and consular reports, cross-national data on technological diffusion, German engineering periodicals, and firsthand accounts from inspection teams commissioned to study related issues.13 My analysis benefited from archival materials based at the Bodleian Library’s Marconi Archives (United Kingdom), the Library of Congress (United States), and the University of Leipzig and from records of the British Foreign Office.
This chapter proceeds as follows. I begin by chronicling the economic power transition that took place during the IR-2 to clarify that the United States, not Germany, ascended to industrial preeminence. I then identify the key technological breakthroughs, which I sort according to their ties to GPT and LS trajectories. Along the dimensions of impact timeframe, phase of relative advantage, and breadth of growth, this chapter demonstrates that the GPT trajectory aligns better with how the IR-2 enabled the economic rise of the United States. Next, I evaluate whether differences in GPT skill infrastructure can account for the American edge over Britain and Germany in interchangeable manufacturing methods. Toward the chapter’s end, I also tackle alternative explanations.14
A Power Transition: America’s Ascent
To begin, tracing when an economic power transition takes place is critical. In 1860, Britain was still at the apogee of its industrial power.15 Most historians agree that British industrial preeminence eroded in the late nineteenth century. By 1913, both the United States and Germany had emerged as formidable rivals to Britain with respect to the industrial and productive foundations of national power. According to Paul Kennedy’s influential account, before World War I Britain was “in third place,” and “in terms of industrial muscle, both the United States and imperial Germany had moved ahead.”16 Aided with more data than was available for the IR-1 case, I map the timeline of this economic power transition with various measures of industrial output and efficiency.
In the IR-2 case, clarifying who surpassed Britain in economic efficiency takes on added gravity. Whereas in the IR-1 Britain separated itself from the rest, both the United States and Germany challenged British industrial power in the IR-2. But studies of this case often neglect the rise of the United States. Preoccupied with debates over whether Germany’s overtaking of Britain sparked World War I, the power transition literature has directed most of its attention to the Anglo-German competition for economic leadership.17 Some LS-based accounts explain only Germany’s rise in this period without investigating America’s ascent.18
As the rest of this section will show, Germany and the United States both surpassed Britain on some measures of economic power, but the United States emerged as the clear productivity leader. Therefore, any explanation of the rise and fall of technological leadership in this period must be centered on the US experience. The following sections trace the contours of the IR-2’s economic power transition with a range of indicators for productivity leadership, including GDP per capita, industrialization, labor productivity, and total factor productivity.
GDP PER-CAPITA INDICATORS
Changes in total GDP over the course of the IR-2 provide a useful departure point for understanding changes in the balance of productive power. At the beginning of the period in 1871, Germany’s economy was around three-quarters the size of the British economy; by the end of the period, in 1913, Germany’s economy was approximately 14 percent larger than Britain’s. The growth trajectory of the American economy was even starker. Over the same time period, overall economic output in the United States increased from 1.2 times to around 3.4 times that of Britain’s total GDP.19 This trend is further confirmed by the growth rates of overall GDP for the three countries. In the period between 1870 and 1913, the US GDP grew roughly 5.3 times over, compared to 3.3 for Germany and 2.2 for the United Kingdom.20
While gross economic size puts countries in contention for economic leadership, the most crucial outcome is sustained economic efficiency. Compared to total output, trend lines in real GDP per capita mark out a broadly similar picture of the IR-2 period, but they also differ in two significant respects (figure 4.1). First, whereas the United States was already the largest economy by total output in 1870, the United Kingdom maintained a slight lead in real GDP per capita over the United States in the 1870s. The United Kingdom’s average GDP per capita over the decade was about 15 percent higher than the US equivalent.21 US GDP per capita was roughly on par with Britain’s throughout the 1880s and 1890s, but the United States established a substantial lead starting around 1900.22
FIGURE 4.1. Economic Power Transition during the IR-2. Source: Maddison Project Database, version 2020 (Bolt and van Zanden 2020).
Second, in contrast to trend lines in aggregate economic output, Germany did not surpass Britain in GDP per capita before World War I. Germany certainly closed the gap, as its GDP per capita increased from around 50 percent of British GDP per capita in 1870 to around 70 percent in the years before World War I. However, Germany never even came close to overtaking the United Kingdom in GDP per capita during this period.23 This is an important distinction that justifies the focus on US technological success in the IR-2, since surpassing at the technological frontier is a different challenge than merely catching up to the technological frontier.
INDUSTRIALIZATION INDICATORS
Industrialization indicators back up the findings from the GDP per-capita data. The United States emerged as the preeminent industrial power, boasting an aggregate industrial output in 1913 that equaled 36 percent of the global total—a figure that exceeded the combined share of both Great Britain and Germany.24 More importantly, the United States became the leading country in terms of industrial efficiency, with a per-capita industrialization level about 10 percent higher than Britain’s in 1913.25
Once again, the emphasis on productivity over aggregate output reveals that the economic gap between Germany and Britain narrowed but did not disappear. In aggregate terms, Germany’s share of the world’s industrial production rose to 16 percent in 1913. This eclipsed Britain’s share, which declined from 32 percent of the world’s industrial production in 1870 to just 15 percent in 1913.26 However, Germany did not overtake Britain in industrial efficiency. In 1913, its per-capita industrialization level was about 75 percent of Britain’s.27 The magnitude of this gap was approximately the same as the gap between German per-capita GDP and British per-capita GDP.
PRODUCTIVITY INDICATORS
Lastly, I consider various productivity statistics. Stephen Broadberry’s work on the “productivity race” contains the most comprehensive and rigorous assessments of productivity levels in Britain, Germany, and the United States in this period.28 Comparative statistics on labor productivity line up with findings from other indicators (figure 4.2). The United States surpassed Britain in aggregate labor productivity during the 1890s or 1900s, whereas Germany’s aggregate labor productivity increased relative to but did not fully overtake Britain’s over the IR-2 period.29
FIGURE 4.2. Comparative Labor Productivity Levels in the IR-2. Source: Broadberry 2006, 110.
Another set of productivity indicators, Maddison’s well-known and oft-cited historical data on comparative GDP per hour worked, supports Broadberry’s comparative measures of labor productivity levels.30 According to Maddison’s estimates of the average rate of productivity growth from 1870 to 1913, the American and German economies were both growing more productive relative to the British economy. The growth rate of America’s GDP per hour worked was 1.9 percent compared to 1.8 percent for the German rate and 1.2 percent for the UK rate.31
It should be noted that the United Kingdom may have retained a total factor productivity lead in this period. Based on 1909 figures, the last measurements available before World War I, the US aggregate TFP was a little over 90 percent of Britain’s. By 1919, US aggregate TFP was nearly 10 percent larger than Britain’s.32 The United States could have surpassed Britain in overall TFP before World War I, but the data do not clearly demonstrate this outcome. Still, the TFP data track well with the general trends found in other measures of economic efficiency, including a marked increase in US TFP in the 1890s and 1900s as well as a steady narrowing of the gap between UK and German TFP throughout the period. Issues related to the availability, reliability, and comparability of capital stock estimates during this period, however, caution against concluding too much from the TFP trends alone.33
Albeit with some caveats, the general thrust of evidence confirms that the United States overtook Britain in productivity leadership around the turn of the twentieth century. In productive efficiency, Germany significantly narrowed the gap but did not surpass Britain. A clarified picture of the outcome also helps guide the assessment of the LS and GPT mechanisms. In contrast to work that focuses on Anglo-German rivalry in this period, I prioritize explaining why the United States became the preeminent economic power. Moreover, if GPT diffusion theory holds for this period, it should also explain why the United States was more successful than Germany in overtaking Britain in productivity during this period.
Key Technological Changes in the IR-2
Which technological changes could have sparked the economic power transition before World War I? The IR-2 was an age of dizzying technological breakthroughs, including but not limited to the electric dynamo (1871), the first internal combustion engine (1876), the Thomas process for steel manufacturing (1877), and the synthesis of indigo dye (1880).34 Tracking down how every single technical advance could have affected the growth differentials among Britain, Germany, and the United States is an unmanageable task. I narrow the scope of analysis to the most likely sources of LS and GPT trajectories based on previous scholarship that calls attention to the significance of certain technological developments in the IR-2. Once confirmed to meet the established criteria for leading sectors and GPTs, these technological drivers serve as the fields of reference for assessing the validity of the GPT and LS mechanisms in this case.
Candidate Leading Sectors
I focus on the chemicals, electrical equipment, motor vehicles, and steel industries as the leading sectors of the IR-2. These choices are informed by scholars who study the implications of technological change during this period from a LS perspective. The first three sectors feature in the standard rendering of the IR-2 by prominent historical accounts, which centers major discoveries in chemistry and electricity as well as the invention of the internal combustion engine.35 Among those who study the effect of technological revolutions on the balance of power, there is near-consensus that the chemicals and electrical industries were technologically advanced, fast-growing industries during this time.36 Some scholars also identify the automobile industry as a key industry in this period.37 Others reason, however, that automobiles did not emerge as a leading sector until a later period.38
The automobile, chemicals, and electrical industries all experienced prodigious growth during the IR-2, meeting the primary qualification for leading sectors. According to statistics from the US census, the percentage increase in value added by manufacture in each of the chemicals, electrical, and automobile industries was much higher than the average across all industries from 1899 through 1909. In fact, the automobile and electrical equipment industries boasted the two highest rates of percentage growth in value added over this period among sectors with a market size over $100 million.39
I also consider developments in steel as a possible source of leading-sector product cycles. It is hard to ignore the explosive growth of the steel industry in both Germany, where it multiplied over 100-fold from 1870 to 1913, and the United States, where it multiplied around 450 times over the same period.40 In addition, many scholars list steel as one of the leading sectors that affected the economic power balance in the IR-2.41 Rostow identifies steel as part of “the classic sequence” of “great leading sectors.”42 In sum, I consider four candidate leading sectors in this period: the automobile, chemicals, electrical equipment, and steel industries.
Candidate GPTs
I analyze chemicalization, electrification, the internal combustion engine, and interchangeable manufacture as potential drivers of GPT-style transformations in the IR-2. Of these four, electricity is the prototypical GPT. It is “unanimously seen in the literature as a historical example of a GPT.”43 Electricity is one of three technologies, alongside the steam engine and information and communications (ICT) technology, that feature in nearly every article that seeks to identify GPTs throughout history.44 Electrical technologies possessed an enormous scope for improvement, fed into a variety of products and processes, and synergized with many other streams of technological development. Empirical efforts to identify GPTs with patent data provide further evidence of electricity as a GPT in this period.45
Like advances in electricity, clusters of innovations in chemicals and the internal combustion engine not only spurred the rapid growth of new industries but also served as a potential source of GPT trajectories. Historians of technology pick out chemicalization, alongside electrification, as one of two central processes that transformed production routines in the early twentieth century.46 Historical patent data confirm that chemical inventions could influence a wide variety of products and processes.47
In line with GPT classification schemes by other scholars, I also evaluate the internal combustion engine as a candidate GPT, with the potential to replace the steam engine as a prime mover of many industrial processes.48 After its introduction, many believed that the internal combustion engine would transform a range of manufacturing processes with smaller, divisible power units.49
Lastly, I examine the advance of interchangeable manufacture, spurred by innovations in machine tools, as a candidate GPT in this period. Though the machine tool industry was neither new nor especially fast-growing, it did play a central role in extending the mechanization of machine-making first incubated in the IR-1. The diffusion of interchangeable manufacture, or the “American system,” owed much to advances in turret lathes, milling machines, and other machine tools that improved the precision of cutting and shaping metals. Rosenberg’s seminal study of “technological convergence” between the American machine tool industry and metal-using sectors highlighted how innovations in metalworking machines transformed production processes across a wide range of industries.50 Following Rosenberg’s interpretation, historians recognize the nexus of machine tools and mechanization as one of the key technological trajectories during this period.51
Sources of LS and GPT Trajectories
I aimed to include as many candidate technological drivers as possible, provided that the technological developments credibly met the criteria of a leading sector or GPT.52 All candidate leading sectors and GPTs I study in this period were flagged in multiple articles or books that explicitly identified leading sectors or GPTs in the IR-2 period, which helped provide an initial filter for selection. This allows for a good test of the GPT diffusion mechanism against the LS product cycles mechanism.53 This sorting process is an important initial step for evaluating the two mechanisms, though a deeper excavation of the historical evidence is required to determine whether the candidates actually made the cut.
table 4.1 Key Sources of Technological Trajectories in the IR-2
Candidate Leading Sectors
Candidate GPTs
Steel industry
Interchangeable manufacture
Electrical equipment industry
Electrification
Chemicals industry
Chemicalization
Automobile industry
Internal combustion engine
There is substantial overlap between the candidate GPTs and leading sectors in the IR-2, as reflected in table 4.1, but two key distinctions are worth emphasizing. First, one difference between the candidate GPTs and leading sectors is the inclusion of machine tools in the former category. The international relations scholarship on leading sectors overlooks the impact of machine tools in this period, possibly because the industry’s total output did not rank among the largest industries, and also because innovation in machine tools was relatively incremental.54 One survey of technical development in machine tools from 1850 to 1914 described the landscape as “essentially a series of minor adaptations and improvements.”55 Relatedly, the steel industry, commonly regarded as an LS, is not considered a candidate GPT. Under the GPT mechanism, innovations in steel are bound up in a GPT trajectory driven by advances in machine tools.
Second, even though some technological drivers, such as electricity, are considered both candidate leading sectors and candidate GPTs, there are different interpretations of how developments in these technological domains translated into an economic power transition. In the case of new electrical discoveries, control over market share and exports in the electrical equipment industry represents the LS trajectory, whereas the gradual spread of electrification across many industries stands in for the GPT trajectory. Two trajectories diverge in a yellow wood, and the case study evidence will show which one electricity traveled.56
GPT vs. LS Trajectories in the IR-2
Equipped with a better grasp of the possible technological drivers in the IR-2, I follow the same procedures used in the previous chapter to assess the validity of the GPT and LS mechanisms.
OBSERVABLE IMPLICATIONS RELATED TO THE IMPACT TIMEFRAME
GPT diffusion and LS product cycles present two competing interpretations of the IR-2’s impact timeframe. The LS mechanism expects growth associated with radical technological breakthroughs to be explosive in the initial stages. Under this view, off the back of major breakthroughs such as the first practical electric dynamo (1871), the modern internal combustion engine (1876), and the successful synthesis of indigo dye (1880), new leading sectors took off in the 1870s and 1880s.57 Then, according to the expected timeline of the LS mechanism, these new industries stimulated substantial growth in the early stages of their development, bringing about a pre–World War I upheaval in the industrial balance of power.58
The GPT trajectory gives a different timeline for when productivity benefits from major technological breakthroughs were realized on an economy-wide scale. Before stimulating economy-wide growth, the candidate GPTs that emerged in the 1880s—tied to advances in electricity, chemicals, and the internal combustion engine—required many decades of complementary innovations in application sectors and human capital upgrading. These candidate GPTs should have contributed only modestly to the industrial rise of the United States before World War I, with impacts, if any, materializing toward the very end of the period.
Critically, one candidate GPT should have produced substantial economic effects during this period. Unlike other GPT trajectories, interchangeable manufacture had been incubated by earlier advances in machine tools, such as the turret lathe (1845) and the universal milling machine (1861).59 Thus, by the late nineteenth century, interchangeable manufacturing methods should have diffused widely enough to make a significant impact on US industrial productivity.
OBSERVABLE IMPLICATIONS RELATED TO THE PHASE OF RELATIVE ADVANTAGE
When spelling out how the IR-2 produced an economic power transition, the two mechanisms also stress different phases of technological change. According to the LS mechanism, Britain’s industrial prominence waned because it lost its dominance of innovation in the IR-2’s new industries. The United States and Germany benefited from monopoly profits accrued from being lead innovators in electrical equipment, chemical production, automobiles, and steel. In particular, Germany’s industrial rise in this period draws a disproportionate share of attention. Many LS accounts attribute Germany’s rise to its dominance of innovations in the chemical industry, “the first science-based industry.”60 Others emphasize that the American global lead in the share of fundamental innovations after 1850 paved the way for the United States to dominate new industries and become the leading economy in the IR-2.61
The GPT mechanism has different expectations regarding the key determinant of productivity differentials. Where innovations are adopted more effectively has greater significance than where they are first introduced. According to this perspective, Britain lost its industrial preeminence because the United States was more effective at intensively adopting the IR-2’s GPTs.
OBSERVABLE IMPLICATIONS RELATED TO BREADTH OF GROWTH
Finally, regarding the breadth of growth, the third dimension on which the two mechanisms diverge, the LS trajectory expects that a narrow set of modernized industries drove productivity differentials, whereas the GPT trajectory holds that a broad range of industries contributed to productivity differentials. The US growth pattern serves as the best testing ground for these diverging predictions, since the United States overtook Britain as the economic leader in this period.
Table 4.2 Testable Predictions for the IR-2 Case Analysis
Prediction 1: LS (impact timeframe)
The steel, electrical equipment, chemicals, and/or* automobile industries made a significant impact on the rise of the United States to productivity leadership before 1914.
Prediction 1: GPT
Electrification, chemicalization, and/or the internal combustion engine made a significant impact on the rise of the United States to productivity leadership only after 1914.The extension of interchangeable manufacture made a significant impact on the rise of the United States to productivity leadership before 1914.
Prediction 2: LS (phase of relative advantage)
Innovations in the steel, electrical equipment, chemicals, and/or automobile industries were concentrated in the United States.German and American advantages in the production and export of electrical equipment, chemical products, automobiles, and/or steel were crucial to their industrial superiority.
Prediction 2: GPT
Innovations in machine tools, electricity, chemicals, and/or the internal combustion engine were not concentrated in the United States.
American advantages in the diffusion of interchangeable manufacture were crucial to its productivity leadership.
Hypothesis 3: LS (breadth of growth)
Productivity growth in the United States was limited to the steel, electrical, chemicals, and/or automotive industries.
Hypothesis 3: GPT
Productivity growth in the United States was spread across a broad range of industries linked to interchangeable manufacture.
*The operator “and/or” links all the candidate leading sectors and GPTs because it could be the case that only some of these technologies drove the trajectories of the period.
The two explanations hold different views about how technological disruptions produced an economic power transition, related to the impact timeframe of new advances, the phase of technological change that yields relative advantages, and the breadth of technology-fueled growth. Based on the differences between the LS and GPT mechanism across these dimensions, I derive three sets of diverging predictions for how technological changes contributed to relative shifts in economic productivity during this period. Table 4.2 collects these predictions, which structure the case analysis in the following sections.
Impact Timeframe: Gradual Gains vs. Immediate Effects from New Breakthroughs
The opening move in assessing the LS and GPT mechanisms is determining when the IR-2’s eye-catching technological advances actually made their mark on leading economies. Tracking the development timelines for all the candidate leading sectors and GPTs of the IR-2 produces two clear takeaways. First, innovations related to electricity, chemicals, and the internal combustion engine did not make a significant impact on US productivity leadership until after 1914. Second, advances in machine tools and steel—the remaining candidate GPT and leading sector, respectively—contributed substantially to US economic growth before World War I; thus, their impact timeframes fit better with when the United States overtook Britain as the preeminent economic power.
DELAYED TIMELINES: CHEMICALS, ELECTRICITY, AND THE INTERNAL COMBUSTION ENGINE
Developments in chemicals, electricity, and internal combustion provide evidence against the LS interpretation. If the LS mechanism was operational in the IR-2, advances in chemicals should have made a significant impact on US productivity leadership before World War I.62 Yet, in 1914, the United States was home to only seven dye-making firms.63 Major US chemicals firms did not establish industrial research laboratories like those of their German counterparts until the first decade of the twentieth century.64 Terry Reynolds, author of a history of the American Institute of Chemical Engineers, concludes, “Widespread use of chemists in American industrial research laboratories was largely a post–World War I phenomenon.”65 Thus, it is very unlikely that chemical innovations made a meaningful difference to growth differentials between the United States and Britain before 1914.
At first glance, the growth of the German chemical industry aligns with the LS model’s expectations. Germany was the first to incorporate scientific research into chemical production, resulting in the synthesis of many artificial dyes before 1880.66 Overtaking Britain in leadership of the chemical industry, Germany produced 140,000 tons of dyestuffs in 1913, more than 85 percent of the world total.67
While Germany’s rapid growth trajectory in synthetic dyes was impressive, the greater economic impacts of chemical advances materialized after 1914 through a different pathway: “chemicalization,” or the spread of chemical processes across ceramics, food-processing, glass, metallurgy, petroleum refining, and many other industries.68 Prior to key chemical engineering advances in the 1920s, industrial chemists devoted limited attention to unifying principles across the manufacture of different products. The rapid expansion of chemical-based industries in the twentieth century owed more to these later improvements in chemical engineering than earlier progress in synthetic dyes.69 Ultimately, these delayed spillovers from chemicalization were substantial, as evidenced by higher growth rates in the German chemical industry during the interwar period than in the two decades before World War I.70
Electrification’s impact timeframe with respect to US productivity growth mirrored that of chemicalization. Scholarly consensus attributes the US productivity upsurge after 1914 to the delayed impact of the electrification of manufacturing.71 From 1880 to 1930, power production and distribution systems gradually evolved from shaft and belt drive systems driven by a central steam engine or water wheel to electric unit drive, in which electric motors powered individual machines. Unit drive became the predominant method in the 1920s only after vigorous debates in technical associations over its relative merits, the emergence of large utilities that improved access to cheap electricity, and complementary innovations, like machine tools, that were compatible with electric motors.72
Quantitative indicators also verify the long interval between key electrical advances and electrification’s productivity boost. Economic geographer Sergio Petralia has investigated the causal relationship between adoption of electrical and electronic (E&E) technologies, operationalized as E&E patenting activity in individual American counties and the per-capita growth of those counties over time. One of his main findings is that the effects of E&E technology adoption on growth are not significant prior to 1914.73 This timeline is confirmed by a range of other metrics, including the energy efficiency of the American economy, electric motors’ share of horsepower in manufacturing, and estimates of electricity’s total contribution to economic growth.74
The diffusion of internal combustion engines across application sectors was also slow. Despite its initial promise, the internal combustion engine never accounted for more than 5 percent of the generation of total horsepower in US manufacturing from 1869 to 1939.75 In 1900, there were only eight thousand cars in the entire United States, and the U. motor vehicle industry did not overtake its French competitor as the world’s largest until 1904.76 Furthermore, the turning point for the mass production of automobiles, Ford’s installation of a moving assembly line for making Model Ts, did not occur until 1913.77
KEY TIMINGS: MACHINE TOOLS AND STEEL
When assigning credit to certain technologies for major upheavals in global affairs, awe of the new often overwhelms recognition of the old. Based on the previous analysis, it is unlikely that new breakthroughs in electricity, chemicals, and internal combustion fueled the economic power transition that transpired in this period. Instead, careful tracing reveals the persevering impact of earlier developments in machine tools.78 During the IR-2, technical advances in machine tools were incremental, continuous improvements that helped disseminate transformative breakthroughs from the mid-nineteenth century, such as the turret lathe and the universal milling machine.79
Profiles of key application sectors and quantitative indicators validate the GPT mechanism’s expected impact timeframe for machine tools. Marking 1880 as the date when “the proliferation of new machine tools in American industry had begun to reach torrential proportions,” Rosenberg outlines how three application sectors—sewing machines, bicycles, and automobiles—successively adopted improved metal-cutting techniques from 1880 to 1910.80 As the American system took hold, the number of potential machine tool users multiplied 15-fold, from just 95,000 workers in 1850 to almost 1.5 million in 1910.81 Patenting data identify the last third of the nineteenth century as the period when extensive technological convergence characterized the machine tool industry and application sectors.82
FIGURE 4.3. Technological Impact Timeframes in the IR-2. Note: US chemical production, horsepower from electric central stations, and machine intensity over time. Source: Murmann 2003; US Census Bureau 1975.
Figure 4.3 depicts the diverging impact timeframes of interchangeable manufacturing methods, electrification, and chemicalization. Machine intensity substantially increased from 1890 to 1910, as measured by horsepower installed per persons employed in manufacturing. By contrast, the United States did not experience significant increases in electrical and chemical production until after 1910.
Of all the candidate leading sectors, the steel industry best fits the expectations of the LS mechanism regarding when industries transformed by radical innovations stimulated growth in the rising powers. Just as the 1780s were a period when the technological conditions for cotton production were transformed, the mid-nineteenth century featured major breakthroughs in the steel industry that allowed for the mass production of steel, such as the Siemens-Martin open-hearth furnace (1867) and Bessemer converter (1856).83 Over the course of the IR-2 period, the United States and Germany quickly exploited these breakthroughs in steelmaking to massively boost steel production.
The overtaking of Britain by both Germany and the United States in total steel production by the early 1890s matches the timeline of Britain’s overall economic decline.84 Paul Kennedy cites Germany’s booming steel output as a key factor driving its industrial rise; by 1914, German steel output was larger than that of Britain, France, and Russia combined.85 Likewise, US steel output grew from one-fifth of British production in 1871 to almost five times more than British steel output in 1912.86 Given these impressive figures, the next section investigates the American and German advantages in steel production in further detail.
Phase of Relative Advantage: The American System’s Diffusion
The second dimension on which the GPT and LS trajectories differ relates to the phase of technological change that accounted for the relative success of the United States in the IR-2. Cross-country historical evidence on the IR-2’s technological drivers illustrates that the United States had true comparative advantages over other advanced economies that were rooted in its absorption and diffusion capabilities.
INNOVATION CLUSTERING IN STEEL, ELECTRICITY, CHEMICALS, AND/OR MOTOR VEHICLES?
In electricity, industrial powers fiercely contested innovation leadership as the United States, Germany, Great Britain, and France all built their first central power stations, electric trams, and alternating current power systems within a span of nine years.87 However, the United States clearly led in diffusing these systems: US electricity production per capita more than doubled that of Germany, the next closest competitor, in 1912. Along this metric of electrification, Britain’s level was just 20 percent of the US figure.88
To be clear, Britain fell behind in adopting electrification, even though it introduced some of the most significant electrical innovations.89 In 1884, for example, British inventor Charles Parsons demonstrated the first steam turbine for practical use, an essential step for commercializing electric power, but this technology was more rapidly and widely adopted in other countries.90 The British Institution of Electrical Engineers aptly captured this phenomenon in an 1892 resolution: “Notwithstanding that our countrymen have been among the first in inventive genius in electrical science, its development in the United Kingdom is in a backward condition, as compared with other countries, in respect of practical application to the industrial and social requirements of the nation.”91
In chemicals, the achievements of both the US and German chemical industries suggest that no single country monopolized innovation in this sector. Germany’s synthetic dye industry excelled not because it generated the initial breakthroughs in aniline-violet dye processes—in fact, those were first pioneered in Britain—but because it had perfected these processes for profitable exploitation.92 Similar dynamics characterized the US chemical industry.93
In most cases, the United States was not the first to introduce major innovations in leading sectors. Many countries introduced major innovations in chemicals, electricity, motor vehicles, and steel during this period (table 4.3).94 Across the four candidate leading sectors, American firms pioneered less than 30 percent of the innovations. Contrary to the propositions of the LS mechanism, innovations in steel, electricity, chemicals, and motor vehicles were spread across the leading economies.
table 4.3 Geographic Distribution of Major Innovations in Leading Sectors, 1850–1914
Chemicals
Electricity
Motor Vehicles
Steel
France
2
1
1
1
Germany
3
3
3
0
Great Britain
1
3
1
1
United States
2
3
1
0
Various other countries
0
1
0
2
Sole US share
25%
27%
17%
0%
Source: Van Duijn 1983, 176–79 (compilation of 160 innovations introduced during the nineteenth and twentieth centuries).
Moreover, the limited role of electrical and chemical exports in spurring American growth casts further doubt on the significance of monopoly profits from being the first to introduce new advances.95 The British share of global chemical exports almost doubled the US share in 1913.96 Overall, the United States derived only 8 percent of its national income from foreign trade in 1913, whereas the corresponding proportion for Britain was 26 percent.97 Even though the United States was the quickest to electrify its economy, Germany captured around half of the world’s exports in electrical products.98
If monopoly profits from innovation clustering in any leading sector propelled the industrial rise of the United States and Germany, it would be the steel industry. Both nations made remarkable gains in total steel output over this period, and scholars commonly employ crude steel production as a key indicator of British decline and the shifting balance of industrial power in the decades before World War I.99 Thus, having established the delayed impact of the electrical, chemical, and automobile industries in this period, the steel industry takes on an especially large burden for the LS mechanism’s explanatory power in this period.
Yet Britain capitalized on many major innovations in steelmaking, including the Talbot furnace, which became essential to producing open-hearth steel.100 Moreover, trade patterns reveal that Britain still held a comparative advantage in the export of steel between 1899 and 1913.101 How to square this with Germany’s dominance in total steel output?
The prevailing wisdom takes total steel output figures to stand for superior American and German technological know-how and productivity.102 In truth, new steelmaking processes created two separate steel industries. Britain shifted toward producing open-hearth steel, which was higher in quality and price. According to the British Iron Trade Association, Britain produced about four times more open-hearth steel than Germany in 1890.103 On the other hand, Germany produced cheap Thomas steel and exported a large amount at dumping prices. In fact, some of Germany’s steel exports went to Britain, where they were processed into higher-quality steel and re-exported.104 In sum, this evidence questions what one scholar deems “the myth of the technological superiority and outstanding productivity of the German steel industry before and after the First World War.”105
AMERICAN MACHINE TOOLS—GPT DIFFUSION ADVANTAGE
Though new industries like electricity and chemicals hog much of the spotlight, developments in machine tools underpin the most important channel between differential rates of technology adoption and the IR-2’s economic power transition. After noting the importance of the electrical and chemical industries during the period, British historian Eric Hobsbawm elevates the importance of machine tools: “Yet nowhere did foreign countries—and again chiefly the USA—leap ahead more decisively than in this field.”106
In line with the expectations of GPT diffusion theory, comparative estimates confirm a substantial US lead in mechanization in the early twentieth century. In 1907, machine intensity in the United States was more than two times higher than rates in Britain and Germany.107 In 1930, the earliest year for which data on installed machine tools per employee are available, Germany lagged behind the United States in installed machine tools per employee across manufacturing industries by 10 percent, with a significantly wider gap in the tools most crucial for mass production.108
This disparity in mechanization was not rooted in the exclusive access of the United States to special innovations in machine tools. In terms of quality, British machine tools were superior to their American counterparts throughout the IR-2 period.109 German firms also had advantages in certain fields like sophisticated power technology.110 Rather, the distinguishing feature of the US machine tool industry was excellence in diffusing innovations across industries.111 Reports by British and German study trips to the United States provide some of the most detailed, reliable accounts of transatlantic differences in manufacturing methods. German observers traveled to the United States to learn from their American competitors and eventually imitate American interchangeable manufacturing methods.112 British inspection teams reported that the US competitive edge came from the “adaptation of special apparatus to a single operation in almost all branches of industry”113 and “the eagerness with which they call in the aid of machinery in almost every department of industry.”114
Fittingly, one of the most colorful denunciations of American innovation capacity simultaneously underscored its strong diffusion capacity. In an 1883 address to the American Association for the Advancement of Science, Henry Rowland, the association’s vice president, denigrated the state of American science for its skew toward the commercialization of new advances. Rowland expressed his disgust with media representations that upheld the “obscure American who steals the ideas of some great mind of the past, and enriches himself by the application of the same to domestic uses” over “the great originator of the idea, who might have worked out hundreds of such applications, had his mind possessed the necessary element of vulgarity.”115 Yet, it was America’s diffusion capacity—in all its obscurity and vulgarity—that sustained its growth to economic preeminence.
Breadth of Growth: The Wide Reach of Interchangeable Manufacture
What were the sources of American productivity growth in the IR-2? The pattern of American economic growth is most pertinent to investigate because the United States overtook Britain in productivity leadership during the IR-2. Regarding the breadth of economic growth, the LS trajectory expects that American productivity growth was concentrated in a narrow set of modernized industries, whereas the GPT trajectory holds that American productivity growth was dispersed across a broad range of industries. Sector-level estimates of total factor productivity (TFP) growth provide useful evidence to assess these diverging propositions.
WIDESPREAD PRODUCTIVITY GROWTH
The historical data support GPT diffusion theory’s expectation of pervasive US productivity growth. John Kendrick’s detailed study of US productivity growth in this period depicts a relatively balanced distribution. Among the industries studied, nearly 60 percent averaged between 1 and 3 percent increases in output per labor-hour from 1899 to 1909.116 Broad swathes of the US economy, outside of the leading sectors, experienced technological change. For instance, the service sector, which included segments of the construction, transport, wholesale, and retail trade industries, played a key role in the US capacity to narrow the gap with Britain in productivity performance.117
R&D-centric sectors were not the primary engines of US growth. In a recent update to Kendrick’s estimates, a group of researchers estimated how much of US productivity growth was driven by “great inventions sectors,” a designation that roughly corresponds to this chapter’s candidate leading sectors.118 They found that these sectors accounted for only 29 percent of U.S. TFP growth from 1899–1909.119 Despite employing 40 percent of all research scientists in 1920, the chemical industry was responsible for only 7 percent of US TFP growth throughout the following decade.120
MACHINE TOOLS AND BROADLY DISTRIBUTED PRODUCTIVITY GROWTH
Broad-based productivity growth in the US economy does not necessarily mean that a GPT was at work. Macroeconomic factors or the accumulation of various, unconnected sources of TFP growth could produce this outcome. Therefore, if the GPT trajectory captures the breadth of growth in the IR-2, then the historical evidence should connect broadly distributed productivity growth in the United States to developments in machine tools.
The extension of the American system boosted productivity in a wide range of sectors. Applications of this system of special tools reshaped the processes of making firearms, furniture, sewing machines, bicycles, automobiles, cigarettes, clocks, boots and shoes, scientific instruments, typewriters, agricultural implements, locomotives, and naval ordnance.121 Its influence covered “almost every branch of industry where articles have to be repeated.”122 Per a 1930 inventory of American machine tools, the earliest complete survey, nearly 1.4 million metalworking machines were used across twenty industrial sectors.123 In his seminal study of American productivity growth during this period, Kendrick identifies progress in “certain types of new products developed by the machinery and other producer industries [that] have broad applications across industry lines” as a key source of the “broad, pervasive forces that promote efficiency throughout the economy.”124
The breadth of productivity spillovers from machine tools was not boundless. Machine-using industries constituted a minority of the manufacturing industries, which themselves accounted for less than one-quarter of national income.125 However, users of new machine tools extended beyond just manufacturing industries. Technologically intensive services, such as railroads and steam transportation, also benefited significantly from improved metalworking techniques.126 In agriculture, specialized machine tools helped advance the introduction of farm machinery such as the reaper, which revolutionized agricultural productivity.127
In describing how machine tools served as a transmission center in the US economy, Rosenberg describes the industry as a pool of skills and technical knowledge that replenishes the economy’s machine-using sectors—that is, an innovation that addresses one industry’s problem gets added to the pool and becomes available, with a few modifications, for all technologically related industries.128 As sales records from leading machine tool firms show, many application sectors purchased the same type of machine. In 1867, Brown and Sharpe Manufacturing Company sold the universal milling machine, just five years after its invention, not only to machinery firms that made tools for a diverse range of industries but also to twenty-seven other firms that produced everything from ammunition to jewelry.129 In this way, the machine tool industry functioned, in Rosenberg’s words, as “a center for the acquisition and diffusion of new skills and techniques in a machinofacture type of economy.”130
Indeed, advances in machine tools had economy-wide effects. The social savings method estimates how much a new technology contributed to economic growth, compared to a counterfactual situation in which the technology had not been invented.131 Referencing this method to differentiate between the impacts of new technologies in this period, economic historian Joel Mokyr puts forward the American system of manufacturing as the most important:
From a purely economic point of view, it could be argued that the most important invention was not another chemical dye, a better engine, or even electricity.… There is one innovation, however, for which “social savings” calculations from the vantage point of the twentieth century are certain to yield large gains. The so-called American System of manufacturing assembled complex products from mass-produced individual components. Modern manufacturing would be unthinkable without interchangeable parts.132
Institutional Complementarities: GPT Skill Infrastructure in the IR-2
With confirmation that the pattern of technological change in the IR-2 is better characterized by the GPT trajectory, the natural next step is to probe variation among leading economies in adapting to this trajectory. Why was the United States more successful than Britain and Germany in adapting to the demands of interchangeable manufacture? According to GPT diffusion theory, the historical evidence should reveal that the US edge was based on education and training systems that broadened and systematized mechanical engineering skills. These institutional adaptations would have resolved two key bottlenecks in the spread of interchangeable manufacture: a shortage of mechanical engineering talent and ineffective coordination between machine tool producers and users.
Widening the Base of Mechanical Engineers
Which institutions for skill formation were most central to the ability of the United States to take advantage of new advances in machine tools? Established accounts of economic rivalry among great powers in the IR-2 focus on skills linked to major innovations in new, science-based industries. Emphasizing Germany’s advantage in training scientific researchers, these studies attribute Germany’s technological success in this period to its investments in R&D facilities and advanced scientific and technical education.133 Such conclusions echo early-twentieth-century British accounts of Germany’s growing commercial prowess, which lauded German higher education for awarding doctorates in engineering and its qualitative superiority in scientific research.134
American leadership in the adoption of interchangeable manufacturing methods was beholden to a different set of institutions for skill formation. Progress in this domain did not depend on new scientific frontiers and industrial research laboratories.135 In fact, the United States trailed both Britain and Germany in scientific achievements and talent.136 Widespread mechanization in the United States rested instead on a broad base of mechanical engineering skills.
Alongside the development of more automatic and precise machine tools throughout the nineteenth century, this new trajectory of mechanization demanded more of machinists and mechanical engineers. Before 1870, US firms relied on informal apprenticeships at small workshops for training people who would design and use machine tools.137 At the same time, engineering education at independent technical schools and traditional colleges and universities did not prioritize mechanical engineers but were mostly oriented toward civil engineering.138 Yet craft-era methods and skills were no longer sufficient to handle advances that enhanced the sophistication of machine tools.139 Thus, in the mid-eighteenth century, the US potential for mechanization was significantly constrained by the need for more formal technical instruction in mechanical engineering.
Over the next few decades, advances on three main fronts met this need for a wider pool of mechanical engineering expertise: land-grant schools, technical institutes, and standardization efforts. In 1862, the US Congress passed the first Morrill Land-Grant Act, which financed the creation of land-grant colleges dedicated to the agricultural and mechanical arts. Although some of these schools offered low-quality instruction and initially restricted their mission to agricultural concerns, the land-grant funds also supported many important engineering schools, such as the Massachusetts Institute of Technology (MIT) and Cornell University.140 The number of US engineering schools multiplied from six in 1862, when the Morrill Act was passed, to 126 in 1917.141 These schools were especially significant in widening the base of professional mechanical engineers. In 1900, out of all students pursuing mechanical engineering at US higher education institutions, 88 percent were enrolled in land-grant colleges.142
The establishment of technical institutes also served demands for mechanical engineering training. Pure technical schools like the Worcester Polytechnic Institute, founded in 1868, and the Stevens Institute of Technology, founded in 1870, developed mechanical engineering curricula that would become templates for engineering programs at universities and colleges.143 Embedded with local and regional businesses, technical institutes developed laboratory exercises that familiarized students with real-world techniques and equipment. In this respect, these institutes and land-grant colleges “shared a common belief in the need to deliver a practice-oriented technical education.”144
Another significant development in the spread of mechanical engineering knowledge was the emergence of professional engineering societies that created industrial standards. The most prominent of these were the American Society of Mechanical Engineers (ASME), founded in 1880, the American Section of the International Association for Testing Materials, set up in 1898, and the Franklin Institute, which became America’s leading technical society around the start of the IR-2.145 As these associations coordinated to share best practices in mechanical engineering, they improved knowledge flows between the machine tool industry and application sectors.146 Standardization in various machine processes and components, such as screw threads, helped spread mechanization across disparate markets and communities.147
It should be emphasized that these efforts were effective in producing the skills and knowledge necessary for advancing mechanization because they broadened the field of mechanical engineering. Mechanical engineering instruction at land-grant schools and technical institutes and through professional associations allowed for more students to become “average engineers,” as opposed to “the perpetuation of a self-recognized elite.”148 Recent research finds that this diffused engineering capacity produced enduring benefits for American industrialization. By collecting granular data on engineering density for the United States at the county level, William Maloney and Felipe Caicedo capture the engineering talent spread across various US counties in 1880 and parse the effect of engineering capacity on industrial outcomes decades later. They find that there is a statistically significant, positive relationship between the level of engineering density in 1880 and the level of industrialization decades later.149
The Comparative Advantage of the United States over Britain and Germany in GPT Skill Infrastructure
Both Britain and Germany fell short of the US standard in GPT skill infrastructure. For Britain, the key gap was in the supply of mechanical engineering talent. British educational institutions and professional bodies fiercely guarded the apprenticeship tradition for training mechanical engineers.150 For instance, the University of Oxford did not establish an engineering professorship until 1908.151 Meanwhile, American engineers systematically experimented with machine redesigns, benefiting from their training at universities and technical institutes.
These diverging approaches resulted in stark differences in skill formation. In 1901, probably around 2,600 students were enrolled in full-time higher technical education in the United Kingdom.152 Limiting this population to those in their third or fourth year of full-time study—an important condition because many UK programs, unlike German and American institutions, did not progress beyond two years of study—leaves only about 400 students.153 By comparison, in mechanical engineering programs alone, the United States in 1900 had 4,459 students enrolled in higher technical education.154 Controlling for population differences, the United States substantially outpaced Britain in engineering density, as measured by the number of university-educated engineers per 100,000 male laborers.155
Germany developed a more practical and accessible form of higher technical education than Britain. From 1870 to 1900, enrollments in the German technische Hochschulen increased nearly fourfold, from 13,674 to 32,834 students.156 Alongside the technische Mittelschulen (technical intermediate schools comparable to American industrial trade schools and lower-level engineering colleges), the technische Hochschulen cultivated a broad base of mechanical engineers.157 Germany’s system of technical education attracted admirers from around the world. Some went there to study in the schools, and others went to study how the school system worked, with the aim of borrowing elements of the German model.158
Germany’s problems were with weak linkages between mechanical engineering education and industrial applications. Key German standards bodies and technical colleges prioritized scientific and theoretical education at the expense of practical skills—a trend “most pronounced in mechanical engineering.”159 According to an expert on German standard-setting in this period, “no national standards movement was inaugurated in [the machine industry] until after the outbreak of [World War I].”160 German experts on engineering education, intent on reforming technical instruction to get engineers more experience with factor organization and project management in the field, recommended, for example, that practical training courses be offered in partnerships with engineering associations.161 Articles in the Zeitschrift des Vereines Deutscher Ingenieure (Journal of the Association of German Engineers) lamented that the technische Hochschulen and technical universities were not equipping students with practical skills to operate in and manage factories and workshops.162 These issues slowed Germany’s incorporation of interchangeable parts and advanced machine tools.
A report by Professor Alois Riedler of the Technical University of Berlin, who was commissioned by the Prussian Ministry of Education to tour American engineering schools in the 1890s, illustrates the differences in engineering education between the United States and Germany. According to Riedler, extensive practical training and experience with shop and laboratory applications were distinctive features of an American engineering education. To substantiate differences in practical instruction between engineering departments in the two countries, Riedler analyzed the time allocated to theoretical and practical instruction across four-year courses of study.163 Compared to their German peers, American students spent far more time on exercises in mechanical technical laboratories and other types of practical training (figure 4.4). In the Technische Universität Berlin (Technical University Berlin), practical exercises in the laboratory and shop accounted for less than 6 percent of total instruction time over a four-year course of study.164 In contrast, engineering students at Cornell University spent more than one-third of their course engaged in laboratory study and shopwork. As a consequence of reports by Riedler and others, German institutions began establishing laboratories for mechanical engineering around 1900.165
FIGURE 4.4. Comparison of Curricula at German and American Engineering Schools (1893). Source: US Bureau of Education (BOE) 1895, 684–86. Note: In this BOE report, the German schools are labeled Technological University in Austria, Technological University in Prussia, and Technological University in South Germany. A reasonable assumption, informed by background research, is that these refer to Technical University Wien, Technical University Berlin, and Technical University of Munich, respectively. Though TU Wien is in Austria, it is used to illustrate trends in German engineering education because many German schools saw it as an influential model.
It should be made clear that, in the United States, institutional adaptations to new opportunities presented by interchangeable manufacture were not rooted in cultivating highly skilled scientific talent. The best and brightest American scientists furthered their education at European universities.166 Even proponents of American engineering education concluded that “strictly scientific and intellectual education in American technological schools” did not even match “the average of a secondary industrial school” in Germany.167 According to one study conducted by the National Association of German-American Technologists, an organization that regularly circulated ideas between the two countries, German technical institutes held an edge over their US peers in research on different technologies in mechanical engineering.168
The deeper roots of US institutions’ greater effectiveness at adapting to the skill formation needs of interchangeable manufacture cannot be fully explored here. The legacy of the Morrill Act certainly looms large, as do the contributions of a diverse set of institutional adaptations unrelated to that groundbreaking federal policy, including independent centers like the Franklin Institute, technical high schools, professional associations, and specialized engineering programs initiated at preexisting universities.169 Other potential sources for the US advantage in GPT skill infrastructure include its openness to foreign technicians and the unique challenges and culture of the American frontier.170
LS-Based Theories and Chemical Engineering
Analyzing the education and training systems for chemical advances provides a secondary test to determine which institutions are most apt to bring national success in technological revolutions.171 LS accounts typically point to Germany’s innovation capacity as the key determinant of its competitiveness in chemicals, especially in the key segment of synthetic dye production.172 To extend this lead in synthetic dyes, Germany profited from leading industrial research labs and scientific education institutions, which employed the world’s top academic chemists and produced about two-thirds of the world’s chemical research.173
By comparison, the US capacity to innovate in chemicals was weak. From 1901 to 1930, only one American researcher received a Nobel Prize in Chemistry, while German and British researchers captured almost three-fourths of the Nobel Prizes in Chemistry in that span.174 In 1899, German publications accounted for half of all citations in American chemical journals, essentially double the share credited to American publications.175 At the same time, American scholarship barely registered in Europe-based chemistry journals, where the best research was published. According to one analysis of references in Annual Reports on the Progress of Chemistry, an authoritative British review journal, American publications accounted for only 7 percent of the citations in 1904.176
As was the case with machine tools, effective adaptation to new chemical technologies in the United States rested on a different set of institutional competencies. Despite trailing Germany in chemical breakthroughs and top chemists, the United States pioneered a chemical engineering discipline that facilitated the gradual chemicalization of many industries. A crucial step in this process was the emergence of unit operations, which broke down chemical processes into a sequence of basic operations (for example, condensing, crystallizing, and electrolyzing) that were useful to many industries, including ceramics, food processing, glass, metallurgy, and petroleum refining.177 American institutions of higher education, most notably MIT, quickly adopted the unit operations model and helped cultivate a common language and professional community of chemical engineering.178 As Rosenberg and Steinmueller conclude, “American leadership in introducing a new engineering discipline into the university curriculum, even at a time when the country was far from the frontier of scientific research, was nowhere more conspicuous than in the discipline of chemical engineering early in the 20th century.”179
In contrast, Germany was slow to develop the infrastructure for supporting chemical engineers. Up through the interwar period, the chemical engineering profession “failed to coalesce in Germany.”180 Chemical engineering did not become a distinct academic subject area in Germany until after the Second World War.181 Because German universities did not equip chemists with engineering skills, the burden of training chemists was shifted to firms.182 Additionally, the German chemical industry maintained a strict division of labor between chemists and mechanical engineers. The lack of skill systematization resulted in more secrecy, less interfirm communication, and a failure to exploit externalities from common chemical processes.183
The United States reaped the spoils of technological convergence in chemicalization not just because it trained large numbers of chemical engineers but also because it strengthened university-industry linkages and standardized techniques in chemical engineering.184 Without the connective tissue that promotes information flows between the chemical sector and application sectors, a large base of chemical engineers was insufficient. Britain, for instance, was relatively successful at training chemical engineers during the interwar period; however, the weak links between British educational institutions and industrial actors limited the dissemination of technical knowledge, and the concept of unit operations did not take hold in Britain to the degree that it did in America.185 Additionally, professional engineering associations in the United States, including the American Institute of Chemical Engineers, advanced standardization in the chemical industry—an initiative not imitated in Britain until a decade later.186 Unlike their American peers, it was not until after World War II that British chemical engineers saw themselves as “members of a professional group that shared a broad commonality cutting across the boundary lines of a large number of industries.”187
Since substantial, economy-wide benefits from these chemical breakthroughs did not materialize until after the end of the IR-2 period, it is important to not overstate these points. Nonetheless, tracing which country best exploited chemical innovations through the interwar period can supplement the analysis of institutional complementarities for machine tools.188 Evidence from the coevolution of chemical technologies and skill formation institutions further illustrates how institutional adaptations suited to GPT trajectories differed from those suited to LS trajectories.
Alternative Factors
Like its predecessor, the IR-2 has been the subject of countless studies. Scholars have thoroughly investigated the decline of Britain and the rise of the United States and Germany, offering explanations ranging from immigration patterns and cultural and generational factors to natural resource endowments and labor relations.189 My aim is not to sort through all possible causes of British decline. Rather, I am tracing the mechanisms behind an established connection between the IR-2’s technological breakthroughs and an economic power transition. Thus, the contextual factors most likely to confound the GPT diffusion explanation are those that provide an alternative explanation of how significant technological changes translated into the United States supplanting Britain in economic leadership. Aside from the LS mechanism, which has been examined in detail, the influence of international security threats and varieties of capitalism deserve further examination.
Threat-Based Explanations
How did external threats influence technological leadership in the IR-2? Scholars have argued that US military investment, mobilized against the threat of a major war, was crucial to the development of many GPTs.190 US national armories’ subsidization of the production of small arms with interchangeable parts in the early nineteenth century was crucial, some studies argue, to the diffusion of the American system to other industries in the second half of the century.191
Though firearms production provided an important experimental ground for mechanized production, military support was not necessary to the development of the American system of manufacturing. Questioning the necessity of government funding and subsidies for the spread of the American system, one study credits the development of interchangeable manufacture to four civilian industries: clock manufacturing, ax manufacturing, typewriter manufacturing, and watch manufacturing.192 In particular, the clock industry played a crucial role in diffusing mechanized production practices. More attuned to the dynamics of the civilian economy than the small arms manufacturers, clockmakers demonstrated that interchangeable manufacture could drastically increase sales and cut costs.193 In his definitive study of the history of American interchangeable parts manufacture, David Hounshell concludes that “the sewing machine and other industries of the second half of the 19th century that borrowed small arms production techniques owed more to the clock industry than to firearms.”194
Military investment and government contracting did not provide long-term sources of demand for interchangeable manufacturing methods.195 Over the course of the IR-2, the small arms industry’s contribution to American manufacturing declined, totaling less than 0.3 percent of value added in American industry from 1850 to 1940.196 Thus, arguments centered on military investment neglect that the spread of the American system, not its initial incubation, is the focal point for understanding how the IR-2 catalyzed an economic power transition.197
Another threat-based argument posits that countries that face more external threats than internal rivalries will achieve more technological success.198 In the IR-2 case, however, the United States was relatively isolated from external conflicts, while the United Kingdom and Germany faced many more threats (including each other).199 Moreover, the United States was threatened more by internal rivalries than by external enemies at the beginning of the IR-2, as it had just experienced a civil war.200 This argument therefore provides limited leverage in the IR-2 case.
VoC Explanations
What about the connection between America’s particular type of capitalism and its economic rise? Rooted in the varieties of capitalism (VoC) tradition, one alternative explanation posits that the United States was especially suited to embrace the radical innovations of the IR-2 because it exhibited the characteristics of a liberal market economy (LME). GPT diffusion theory and VoC-based explanations clash most directly on points about skill formation. The latter’s expectation that LMEs like the United States should excel at cultivating general skills could account for US leadership in GPT diffusion during this period.201
The empirical evidence casts doubt on this explanation. In the early twentieth century, the leading nations had fairly similar levels of enrollment rates in elementary and post-elementary education. In 1910, enrollment rates for children ages five to nineteen in the United States was about 12 percent lower than Britain’s rate and only 3 percent higher than Germany’s.202 Years of education per worker increased by essentially the same proportion in both Britain (by a factor of 2.2) and the United States (by a factor of 2.3) between 1870 and 1929.203 In terms of higher education expenditures per capita, the two countries were essentially tied.204 Differences in the formation of general skills cannot account for the technological leadership of the United States in this period.205
Moreover, the degree to which the United States fully embraced the characteristics of LMEs is disputed. In the view of studies that position the United States as a model for managerial capitalism in this period, US industrial governance structures enabled the rise of giant managerialist firms.206 This approach primarily sees America’s rise to industrial preeminence through the most visible actors in the American system of political economy: oligopolies in the automobile, steel, and electrical industries. There was significant diversity, however, in firm structure. Though many giant corporations did grow to take advantage of economies of scale and capital requirements in some mass-produced goods (such as automobiles), networks of medium-sized firms still dominated important segments of these new industries, such as the production of electric motors. One-third of the fifty largest manufacturing plants in the United States made custom and specialty goods.207 From 1899 to 1909, sectors that relied on batch and custom production, including machine tools, accounted for one-third of value added in manufacturing.208 No specific brand of capitalism fulfilled the demands of production across all domains.209
Case-Specific Factors
Another traditional explanation highlights American natural resource abundance as a key factor in transatlantic differences in mechanization. Compared to its European competitors, the benefits to the United States derived from its endowment of natural resources, such as plentiful supplies of timber, biased its manufacturing processes toward standardized production.210 “The American turn in the direction of mass production was natural,” claims one influential study of diverging approaches to mechanization in this period.211
The extent to which natural resource endowments determined transatlantic differences in technological trajectories is disputed. Undeterred by natural resource differences, German engineers and industrialists frequently used American machine tools and imitated US production technology.212 In fact, around the early twentieth century, the level of machine intensity in German industries was catching up to the rate in American industries.213 The US-Germany gap in mechanization was more about Germany’s struggles in proficiently using advanced tools, not the choice of methods shaped by natural resource endowments. Crucially, skill formation and embedded knowledge about working with new machinery influenced the efficient utilization of American capital-intensive techniques.214
Summary
The standard version of the Second Industrial Revolution’s geopolitical aftershocks highlights Germany’s challenge to British power. Germany’s relative economic rise, according to this account, derived from its advantage in industrial research and scientific infrastructure, which enabled it to capture the gains from new industries such as electricity and chemicals. However, a range of indicators emphasize that it was the United States, not Germany, that surpassed Britain in productivity leadership during this period. The US industrial ascent in the IR-2 illustrates that dominating innovation in leading sectors is not the crucial mechanism in explaining the rise and fall of great powers. Britain’s decline was not a failure of innovation but of diffusion. As the renowned economist Sir William Arthur Lewis once mused, “Britain would have done well enough if she merely imitated German and American innovations.”215
Indeed, the IR-2 case further supports the conclusion that capacity to widely diffuse GPTs is the key driver of long-term growth differentials. The US success in broadening its talent base in mechanical engineering proved critical for its relative advantage in adapting machine tools across a broad range of industries. Like all GPT trajectories, this process was a protracted one, but it aligns better with when the United States surpassed Britain in productive leadership than the more dramatic breakthroughs in chemicals, electricity, and automobiles. To further investigate how the LS mechanism breaks down and why the GPT mechanism holds up, we turn to the high-tech competition in the twentieth century between the United States and Japan—or what some label the Third Industrial Revolution.
5 Japan’s Challenge in the Third Industrial Revolution
IN THE TWO previous cases, an industrial revolution preceded a shift in global leadership. Britain established its economic dominance in the early nineteenth century, and the United States took the mantle in the late nineteenth century. During the last third of the twentieth century (1960–2000), the technological environment underwent a transformation akin to the First and Second Industrial Revolutions. A cluster of information technologies, connected to fundamental breakthroughs in computers and semiconductors, disrupted the foundations of many industries. The terms “Third Industrial Revolution” (IR-3) and “Information Age” came to refer to an epochal shift from industrial systems to information-based and computerized systems.1 Amid this upheaval, many thought Japan would follow in the footsteps of Britain and the United States to become the “Number One” technological power.2
Of the countries racing to take advantage of the IR-3, Japan’s remarkable advances in electronics and information technology garnered a disproportionate share of the spotlight. “The more advanced economies, with Japan taking the lead in one industry after another, [were] restructuring their economies around the computer and other high tech industries of the third industrial revolution,” Gilpin wrote.3 In the late 1980s and early 1990s, a torrent of works bemoaned the loss of US technological leadership to Japan.4 In a best-selling book on US-Japan relations, Clyde Prestowitz, a former US trade negotiator, declared, “Japan has … become the undisputed world economic champion.”5
Japan’s dominance in the IR-3’s leading sectors was perceived as a threat to international security and to US overall leadership of the international system.6 Former secretary of state Henry Kissinger and other prominent thinkers warned that Japan would convert its economic strength into threatening military power.7 Per a 1990 New York Times poll, 58 percent of Americans believed that Japan’s economic power was more of a threat to American security than the Soviet Union’s military power.8
Historical precedents loomed over these worries. US policymakers feared that falling behind Japan in key technologies would, like relative declines experienced by previous leading powers, culminate in an economic power transition. Paul Kennedy and other historically minded thinkers likened the US position in the 1980s to Britain’s backwardness a century earlier: two industrial hegemons on the brink of losing their supremacy.9 Often alluding to the LS mechanism, these comparisons highlighted Japan’s lead in specific industries that were experiencing significant technological disruption, such as consumer electronics and semiconductors. As David Mowery and Nathan Rosenberg wrote in 1991, “Rapidly growing German domination of dyestuffs helped to propel that country into the position of the strongest continental industrial power. The parallels to the Japanese strategy in electronics in recent decades are striking.”10
Many voices called for the United States to mimic Japan’s keiretsu system of industrial organization and proactive industrial policy, which they viewed as crucial to the rising power’s success with leading sectors.11 Kennedy’s The Rise and Fall of the Great Powers attributed Japan’s surge in global market shares of high-tech industries to R&D investments and the organizing role of the Ministry of International Trade and Industry (MITI).12 These claims about the basis of Japan’s leadership in the information revolution relied on LS product cycles as the filter for the most important institutional factors.
The feared economic power transition, however, never occurred. To be sure, Japanese firms did take dominant positions in key segments of high-growth industries like semiconductors and consumer electronics. Additionally, the Japanese economy did grow at a remarkable pace, averaging an annual 2.4 percent increase in total factor productivity (TFP) between 1983 and 1991. However, Japan’s TFP growth stalled at an average of 0.2 percent per year in the 1990s—a period known as its “lost decade.” By 2002, the per capita GDP gap between Japan and the United States was larger than it had been in 1980.13 Becoming the world’s leading producer in high-tech industries did not catalyze Japan’s overtaking of the United States as the leading economy.
The IR-3 case is particularly damaging for LS-based explanations. Japan took advantage of the IR-3’s opportunities by cornering the market in new, technologically progressive industries, fulfilling the conditions posited by the LS mechanism for Japan to become the foremost economic power. Yet, as the case study evidence will reveal, an economic power transition did not occur, even though all these conditions were present. The Japanese challenge to American technological leadership in the last third of the twentieth century therefore primarily functions as a deviant, or falsifying, case for the LS mechanism.14
By contrast, the IR-3 case does not undermine the GPT mechanism. Since Japan did not lead the United States in the diffusion of general-purpose information technologies, the conditions for an economic power transition under the GPT mechanism were absent in the IR-3. Since there could be many reasons why an economic power transition does not occur, the absence of a mechanism in a negative case provides limited leverage for explaining how technology-driven economic power transitions occur. Still, the IR-3 case evidence will show that LS theory expects an outcome that does not occur—a US-Japan economic power transition—in part because it fails to account for the relative success of the United States in GPT diffusion. This advantage stemmed from its superior ability to cultivate the computer engineering talent necessary to advance computerization. In that regard, this deviant case can help form better mechanism-based explanations.15
Surprisingly, few scholars have revisited claims that Japan’s leadership in leading sectors meant that it was on its way to economic preeminence.16 Decades of hindsight bring not just perspective but also more sources to pore over. Revised estimates and the greater availability of data help paint a more granular picture of how the US-Japan productivity gap evolved in this period. To narrow down the crucial technological trajectories, I pieced together histories of semiconductors and other key technologies, comparative histories of technological development in the United States and Japan, and general economic histories of the IR-3. In addition, I leveraged bibliometric techniques to estimate the number of universities in both countries that could supply a baseline quality of software engineering education. Surveys on computer utilization by Japanese agencies, presentations on computer science education by Japanese and American analysts at international meetings, documents from the Edward A. Feigenbaum Papers collection, and back issues of Nikkei Computer (日経コンピュータ) at the Stanford University East Asia Library all helped flesh out the state of GPT skill infrastructure in the IR-3.
The evaluation of the GPT and LS mechanisms against historical evidence from the IR-3 proceeds as follows. The chapter first makes clear that a US-Japan economic power transition did not take place. Subsequently, it reviews and organizes the technological breakthroughs of the IR-3 into candidate leading sectors and GPTs. It then examines whether all the components of the GPT or LS mechanism were present. Since the outcome did not occur in this case, it is important to trace where the mechanisms break down. All the aspects of the LS mechanism were present in the IR-3, but the GPT mechanism was not operational because Japan fell behind the United States in diffusing information technologies across a broad range of sectors. Based on this evidence, the next section explains why institutional explanations rooted in LS trajectories are unconvincing. Before turning to alternative factors and explanations, the chapter analyzes whether GPT skill infrastructure was a factor in sustained US technological leadership.
A Power Transition Unfulfilled: Japan’s Rise Stagnates
In a 1983 article for Parade, the Pulitzer Prize–winning journalist David Halberstam described Japan’s industrial ascent as America’s “most difficult challenge for the rest of the century” and “a more intense competition than the previous political-military competition with the Soviet Union.”17 By the end of the century, however, the possibility of Japan displacing the United States as the technological hegemon was barely considered, let alone feared.18 The economic power transition that accompanied the IR-1 and IR-2 did not materialize in this case. Indeed, most indicators presented a clear trend: Japan’s economy catches up in the 1980s, stagnates in the 1990s, and ultimately fails to overtake the US economy in productivity leadership.
GDP PER-CAPITA INDICATORS
During the three decades after 1960, Japan’s economy experienced remarkable growth, reaching a GDP per capita in 1990 that was 81 percent of the US mark that year. In the following ten years, known as Japan’s “lost decade,” Japan’s growth in GDP per capita stalled. By 2007, Japan’s GDP per capita had dropped back down to 73 percent of that of the United States (figure 5.1).
INDUSTRIALIZATION INDICATORS
Comparative industrialization statistics tell a similar story. In terms of global manufacturing output, Japan gained on the United States through the 1970s and 1980s and nearly matched the United States, at 20 percent of global manufacturing output, in the early 1990s. Japan’s share of global manufacturing output subsequently declined to around 10 percent in 2010, while the US share increased in the 1990s and held at 20 percent until 2010.19 In manufacturing industries, Japan’s labor productivity growth from 1995 to 2004 averaged only 3.3 percent, whereas the United States averaged 6.1 percent in the same metric.20
FIGURE 5.1. Japan’s Catch-up to the United States in GDP per Capita Stalls in the 1990s. Note: Real GDP per capita in 2011$ prices. Source: Maddison Project Database, version 2020 (Bolt and van Zanden 2020).
PRODUCTIVITY INDICATORS
Productivity statistics also reveal a general trend of convergence without overtaking. From a total factor productivity of just half that of the United States in 1955, Japan’s TFP grew steadily. By 1991, the productivity gap between the United States and Japan was only 5 percent. As was the case with the GDP per capita and industrialization figures, Japan’s productivity growth then slowed, and the gap between the United States and Japan widened during the 1990s (figure 5.2). Throughout this decade, Japan averaged just 0.2 percent annual TFP growth.21 By 2009, Japan’s TFP dropped to only 83 percent of the US figure. The US-Japan labor productivity gap followed a similar course.22
FIGURE 5.2. Japan’s Catch-up to the United States in Productivity Stalls in the 1990s. Source: Jorgenson, Nomura, and Samuels 2018, 18.
Key Technological Changes in the IR-3
Parsing through the different trajectories of technological change is a necessary first step to determine whether the LS and GPT mechanisms were operative in this period. This task is complicated by the tremendous technological changes that emerged in the IR-3, such as the first microprocessor (1971), the production of recombinant DNA (1972), the VHS format for video recording (1976), and the first personal computer (1981). Guided by past scholars’ efforts to map the key nodes of the information revolution as well as analytic measures for leading sectors and GPTs, this section takes stock of key technological drivers that affected the US-Japan economic power balance.
Candidate Leading Sectors
Amid a shifting technological landscape, the most likely sources of LS trajectories were information and communications technologies (ICTs). Certainly, scholars have highlighted technological developments in a wide range of industries as possible leading sectors, including lasers and robotics.23 Nonetheless, Japan’s success in the computer, consumer electronics, and semiconductor industries was most relevant for its prospects of overtaking the United States as the foremost economic power. In each of these three leading sectors, Japan dominated the production of key components.24
All three relatively new industries achieved extraordinary growth off the back of technological breakthroughs, fulfilling the established criteria for leading sectors. In the US economy during the 1980s, computer and data-processing services ranked as the fastest-growing industries in terms of jobs added.25 After Japan’s MITI identified semiconductors and computers as strategic industries in 1971, both industries experienced extremely high growth rates in the next two decades.26 The US electronics industry also experienced a remarkable surge during the late twentieth century, growing thirty times faster than other manufacturing industries by some estimates.27 These trends in computers and electronics held across advanced industrialized countries.28
Candidate GPTs
Given that the IR-3 is often known as the “information revolution,” clusters of ICT innovations naturally serve as the most likely sources of GPT trajectories. Efforts to map this era’s GPTs specifically highlight computers,29 semiconductors,30 and the internet.31 Each of these technological domains exhibited great scope for improvement and complementarities with other technologies.
Because advances in computers, semiconductors, and the internet were all closely connected, I group technological developments in these three domains under “computerization,” the process in which computers take over tasks such as the storage and management of information. This is consistent with other studies that identify the general category of ICT as the GPT operative in this period.32 The growing prevalence of software-intensive systems enabled computers to become more general-purpose. Computerization also benefited from advances in both semiconductors, which reduced costs for investments in IT equipment, and the internet, which connected computers in efficient networks.33
Sources of LS and GPT Trajectories
In sum, the IR-3’s candidate leading sectors and GPTs all revolved around ICTs (table 5.1). Other technical advances in lasers, new sources of energy, and biotechnology also drew attention as possible sources of LS and GPT trajectories. I do not trace developments in these technologies because their potential was largely unrealized, at least within the context of US-Japan economic competition during the IR-3.
Though candidate leading sectors and GPTs converge on ICTs, they diverge on the key trajectories. The GPT perspective emphasizes the process by which firms transfer tasks and activities to computers. In contrast, LS accounts spotlight the growth of key industry verticals. For example, the consumer electronics industry fits the mold of previous candidate leading sectors like automobiles and cotton textiles, which were large and fast-growing but limited in their linkages to other industries. Taking stock of candidate leading sectors and GPTs merely functions as a preliminary filter. The rest of the chapter will further flesh out the differences between the LS- and GPT-based explanations of how these technologies affected the US-Japan economic rivalry.
table 5.1 Key Sources of Technological Trajectories in the IR-3
Candidate Leading Sectors
Candidate GPTs
Computer industry
Computerization
Consumer electronics industry
Semiconductor industry
GPT vs. LS Mechanisms: The (Non) Spread of ICTs across Japan’s Economy
In both the IR-1 and IR-2 cases, a technological revolution sparked a shift in global economic leadership. The case analyses aimed to determine whether the historical evidence fit better with observable implications derived from the GPT or LS mechanism. Revolutionary technological breakthroughs also occurred in the IR-3, but an economic power transition never occurred. Thus, if the case study reveals that Japan dominated innovation in the computer, consumer electronics, and semiconductor industries, then this would provide disconfirming evidence against the LS mechanism. Likewise, if the historical evidence shows that Japan led the way in computerization during this period, this would undermine GPT diffusion theory.
LS Mechanism Present: Japan Dominates the Production of Key ICTs
A bevy of evidence establishes that the LS mechanism was operative during the IR-3. From the mid-twentieth century through the 1980s, Japan captured a growing global market in new industries tied to major technological discoveries in semiconductors, consumer electronics, and computers. In dynamic random-access memory (DRAM) chips, one of the highest-volume verticals in the semiconductor industry, Japanese firms controlled 76 percent of the global market share.34 By one US federal interagency working group estimate, from 1980 to 1987 the United States lost the lead to Japan in more than 75 percent of critical semiconductor technologies.35
Japanese industry also gained competitive advantages in consumer electronics. From 1984 to 1990, US firms lost global market share in thirty-five of thirty-seven electronics categories as Japanese firms took over the production of many electronic products.36 Japan occupied dominant shares of global production of color televisions and DVDs.37 It was also the first economy to commercialize high-definition television (HDTV) systems, a highly touted part of the consumer electronics market.38
A similar trend held in computers, especially in computer hardware components like flat panel displays.39 The US trade balance in computers with Japan turned from a surplus in 1980 into a $6 billion deficit by 1988.40According to the Yearbook of World Electronics Data, in 1990 Japan’s share of global computer production eclipsed the share held by the United States, which had previously led the world.41
Comparisons of LS growth rates also indicate that Japan was poised to overtake the United States as the economic leader. In an International Organization article published in 1990, Thompson posited that average annual growth rates in leading sectors across major economies heralded shifts in economic leadership. Over the nineteenth century, Britain’s growth rate in leading sectors peaked in the 1830s before flattening between 1860 and 1890, a period when the United States and Germany outstripped Britain in LS growth rates.42 Crucially, Thompson’s data showed that Japan outpaced the United States in growth rates within leading sectors from 1960 to 1990.43 Linking these historical trends, Thompson identified Japan as America’s main competitor for “systemic leadership.”44
Comprehensive assessments of Japan’s relative industrial strength support this account of LS growth rates. US government, academic, and industry entities issued a plethora of reports warning of Japan’s growing global market share and exports in key technologies. One review of six such reports, all published between 1987 and 1991, found a growing consensus that US capabilities in many of these technologies were declining relative to Japan’s.45 A 1990 US Department of Commerce report on trends in twelve emerging technologies, including supercomputers, advanced semiconductor devices, and digital imaging technology, projected that the United States would lag behind Japan in most of these technologies before 2000.46 The 1989 MIT Commission on Industrial Productivity’s Made in America: Regaining the Productive Edge serves as a particularly useful barometer of Japan’s position in leading sectors.47Made in America argued that the United States was losing out to Japan in eight manufacturing sectors, including consumer electronics, semiconductors, and computers. As business historian Richard Langlois summarizes, “By the mid-1980s, by most accounts, America had ‘lost’ consumer electronics and was in imminent danger of losing semiconductors and computers.”48
Some argued that Japan’s advantage in these leading sectors was rooted in certain institutional arrangements. Observers regularly pointed to Japan’s keiretsu system, which was structured around large, integrated business groups, as the key institutional factor in its success in high-tech industries. The MIT Commission’s Made in America report, for instance, questioned whether the US system of industrial organization could match up against “much stronger and better organized Japanese competition.”49 This aligned with a common narrative in the mid-1980s that “American firms should become more like Japanese firms.”50
Others pointed to Japan’s industrial policy, coordinated by MITI, as the key institutional competency that explained Japan’s success in leading sectors. Academics and policymakers pushed for the United States to imitate Japan’s industrial policy approach, which they perceived as effective because of MITI’s ability to strategically coordinate R&D investments in key technologies.51 For instance, scholars regarded the “Fifth Generation Project,” a national initiative launched by MITI in 1982, as a stepping-stone to Japan’s building of the world’s most advanced computers.52 The American aversion to industrial policy and a decentralized economic policymaking apparatus, by comparison, was alleged to be detrimental to innovation in the IR-3’s new technologies.
By the turn of the millennium, such arguments were no longer being put forward. Despite capturing key LS industries and rapidly catching up to the United States in the 1980s, Japan did not ultimately overtake the United States as the lead economy. Contrary to the expected outcome of the LS mechanism, Japan’s control of critical sectors in semiconductors and consumer electronics did not translate into strong, sustained economic growth. This outcome challenges the LS mechanism’s validity in the IR-3 case.
An Absent GPT Mechanism: The United States Leads Japan in ICT Diffusion
Does evidence from the IR-3 also discredit the GPT mechanism? If the components of the GPT mechanism, like the LS mechanism, were present during this period, this would weaken the explanatory power of GPT diffusion theory. However, in contrast to its success in key leading sectors, Japan lagged in adopting computerized technologies. Thus, GPT diffusion theory would not expect Japan to have overtaken the United States in the IR-3.
To account for sustained US economic leadership in the IR-3, this section traces the developments of the IR-3 in the United States and Japan across the three dimensions that differentiate GPT from LS trajectories. First, relative to the LS mechanism, the impact timeframes of the IR-3’s technological breakthroughs are more elongated. Developments in ICTs did not spread to a wide range of economic applications until the 1990s. Second, though Japan excelled in the production of computers and electronics, it fell behind in the general pace of computerization across the economy. Lastly, Japan’s advantages were concentrated in a narrow range of ICT-producing industries, whereas the United States benefited from broad-based productivity growth.
IMPACT TIMEFRAME
The advance of computerization, like past GPT trajectories, demanded a prolonged period of organizational adaptation and complementary innovations. It is reasonable to date the computerization GPT’s emergence as the year 1971—the year when Intel introduced the microprocessor, which greatly expanded the functionalities of computers.53 It was also the year when the share of information technology equipment and software reached 1 percent in the net capital stocks of the median sector in the US economy.54 Before then, during the 1960s, mainframe computers powered by integrated circuits serviced only a limited range of commercial purposes, such as producing bank statements and managing airline reservations. With the internet’s rise in the 1990s, new information and communication networks further spread computerization to different business models, such as e-commerce.55 Alongside this stream of complementary technical advances, firms needed time to build up their computer capital stock and reorganize their business processes to match the needs of information technology.56
Computers traveled a gradual and slow road to widespread use. By the late 1980s, many observers bemoaned the computer revolution’s failure to induce a surge of productivity growth. In 1987, the renowned economist Robert Solow distilled this “productivity paradox” in a famous quip: “We see the computers everywhere but in the productivity statistics.”57 A decade later, however, the growing adoption of information technology triggered a remarkable surge in US productivity growth.58 It took some time, but the American economy did eventually see the computers in the productivity statistics.
The landscape of US-Japan technology competition looks very different when accounting for the elongated lag from computerization’s arrival to its widespread diffusion. Japan’s control over key segments of ICT production in the 1970s and 1980s did not correspond to an advantage in GPT diffusion. A more patient outlook illuminates that the delayed impact of computerization aligns with when the United States extended its productivity lead over Japan. After 1995, while Japan’s economic rise stalled, labor and total factor productivity grew rapidly for a decade in the United States. The difference was that the United States benefited greatly from an ICT-driven productivity acceleration.59
FIGURE 5.3. The US-Japan Computerization Gap Widens in the 1990s. Source: Milner and Solstad 2021; Comin 2010, 381.
PHASE OF RELATIVE ADVANTAGE
If the GPT mechanism was operative for Japan’s rise in the IR-3, Japan should have led the United States in computerization. Though Japan continued to contest US leadership in the production of certain computer architectures and devices, it failed to keep up with the United States in the adoption of computers across industries. As figure 5.3 reveals, the gap between the United States and Japan in computerization widened in the 1990s. In fact, South Korea, which lagged behind Japan in generating new innovations in computer systems, surpassed Japan in its computer usage rate during the 1990s. Taken together, these indicators suggest that Japan’s problem was with GPT diffusion, not innovation.
The pathway by which ICTs drove the US-Japan productivity gap is particularly revealing. In sectors that produced ICTs, Japan’s TFP acceleration was similar to the US trajectory; however, in sectors that intensively used IT, Japan’s TFP growth lagged far behind that of its rival.60 In particular, US ICT-using service industries adapted better to computerization. In terms of labor productivity growth in these industries, the United States experienced the strongest improvement out of all OECD countries from the first half of the 1990s to the second half of the decade.61 The contribution of ICT-using services to Japan’s labor productivity growth, by contrast, declined from the first half to the second half of the decade.62
Japan eventually adopted the GPT trajectory associated with ICT. Like all advanced economies at the frontier, Japan could draw from the same technology pool as America. Using a growth accounting framework that accounts for cyclical factors, Takuji Fueki and Takuji Kawamoto trace Japan’s post-2000 resurgence in TFP growth to the extension of the ICT revolution to a broader range of IT-using sectors.63 By then, however, Japan was at least five years behind the United States in taking advantage of computerization.
BREADTH OF GROWTH
Alongside the dispersion of ICTs throughout the US economy, the sources of productivity growth also spread out. In the United States, spillovers from ICT advances, especially in services, contributed to economy-wide TFP growth. Industry-level patterns of TFP growth reveal that US productivity growth became noticeably more broad-based after 1995, a trend that accelerated further after 2000.64
In contrast, Japan’s advantages in the IR-3 were concentrated in a narrow range of sectors. After 1995, Japanese productivity growth remained localized, only transitioning to a more broad-based pattern after 2000.65 Michael Porter’s exhaustive account of national competitiveness across leading economies describes Japan as a “study in contrasts,” with some of the most internationally competitive industries found side by side with some of the most uncompetitive.66 The hype over Japan’s success in leading sectors induced some analysts to generalize from a few exceptional industrial sectors while overlooking developments in struggling sectors.67
LS Mechanisms and “Wintelism”
Some international political economy scholars divide trends in US competitiveness into a phase of relative decline in the leading sectors of the 1980s (consumer electronics, computer hardware, and parts of the semiconductor industry) and a period of resurgence in the new leading sector of the 1990s (software electronics).68 This explanation for why Japan’s LS advantage did not convert into an economic power transition could conceivably restore credibility to the LS mechanism. However, even this generous interpretation fails to capture the dynamics of the IR-3.69
Consider one prominent line of LS-based thinking that emphasizes US advantages in adapting to “Wintelism,” a type of industrial structure best suited for new advances in software electronics. Wintelism, a portmanteau of “Windows” and “Intel,” refers to the transformation of the computer industry from a vertically integrated oligopoly into a horizontally segmented structure dominated by components providers with controlling architectural standards, such as Intel and Microsoft.70 Compared to Japan, the US institutional environment was more supportive of horizontal, specialized value-chains in software electronics. It is put forward that Japan’s inability to adapt to the Wintelist industrial paradigm explains why it was unable to overtake the United States as the leading economy.71
The Wintelism argument falls prey to general issues with the LS mechanism.72 Like LS accounts of Japan’s advantage in semiconductors and consumer electronics in the 1980s, the Wintelism argument still places too much value on ICT-producing industries. Capturing profit shares in the global software electronics industry is not the same as translating new advances in software electronics into economy-wide growth.73 The pathway from new technologies to overall productivity growth involves much more than just the success of companies like Microsoft and Intel.
Indeed, as a GPT diffuses more widely, large monopolists may hinder coordination between the GPT and application sectors. Both Microsoft and Intel, for instance, often restricted sharing information about their technology roadmaps, thereby hindering complementary innovations and adoption of microelectronics in applications sectors such as automobiles. In fact, regulatory and technological forces were necessary to limit the influence of dominant firms and encourage the development of complementary technologies, which were crucial to widening the GPT trajectory of computerization.74 It is conceivable that the United States could have achieved a greater rate of computerization if its computer industry had not been dominated by two firms that held key architectural standards.
Overall, hindsight is not kind to LS-based accounts of Japan’s institutional advantages in the IR-3. While matching institutional competencies to particularly significant technological trajectories is a sound approach, the LS trajectory fails to capture how technological changes opened opportunities for an economic power transition. Japan’s industrial structure and sectoral targeting policies reaped economic gains that were temporary and limited to specific industries.75 To understand why the United States gained lasting and broad-based advantages from computerization, a different set of institutional complementarities must be explored.
Institutional Complementarities: GPT Skill Infrastructure in the IR-3
In line with GPT diffusion theory, institutional adaptations that widened the base of computer engineering skills and knowledge proved crucial to the enduring technological leadership of the United States in the IR-3.76 Computerization required not just innovators who created new software architectures but also programmers who undertook more routine software engineering tasks. It was the US ability to tap into a more expansive pool of the latter that fueled a more intensive pace of computerization than in Japan.
Widening the Base of Computer Engineers
GPTs generate an imbalance between the possibility of sweeping changes across many domains and the constraints of existing skills. Historically, engineering disciplines have developed procedures for adapting new GPT-linked knowledge across localized applications and expanded access to such knowledge to a broader set of individuals. Similarly, the key element of the American GPT skill infrastructure in the IR-3 was the development of a computer science discipline—the latest in a line of engineering disciplines that have emerged in the wake of GPTs.77
US education effectively adapted to changes in the computerization trajectory. The recognition of computer science as an independent discipline, as evidenced by the early and rapid growth of computer science departments in the United States, helped systematize the knowledge necessary for the spread of computerization.78 Led by top universities and the Association of Computing Machinery (ACM), US institutions piloted new training programs in computer science. In 1968, the ACM published an influential and forward-looking curriculum that helped colleges organize their computing education.79 These adaptations converted the practical experience from developers of new computer and software breakthroughs into general, accessible knowledge.80
The development of this GPT skill infrastructure met significant hurdles. In the 1970s, the US computer science discipline struggled to meet demands for software engineering education in the midst of conflicts over the right balance between theory-oriented learning and hands-on programming work.81 Reacting to this situation, the ACM’s 1978 curriculum revision directed computer science toward application-based work, stating that programming exercises “provide a philosophy of discipline which pervades all of the course work.”82 Industry pressure and the US Department of Defense’s Software Engineering Institute, a partnership established with Carnegie Mellon University in 1984, expanded the number of software engineering specializations at universities.83 Teaching capacity was another inhibiting factor. Computer science departments had to limit course enrollments because they could not fill the faculty positions to meet exploding student demand, resulting in a decline in computer science enrollments in the mid-1980s.84
Though the process was not seamless, overall trends reflect the US success in widening the pool of engineering skills and knowledge for advancing computerization. From 1966 to 1996, the number of undergraduate computer science degrees awarded annually in the United States grew from 89 to about 24,500.85 In 1986, at its peak during this period, computer science accounted for 12.5 percent of all science and engineering degrees awarded in the United States.86 According to benchmarking efforts by the Working Group for Software Engineering Education and Training, US institutions accounted for about one-third of the world’s bachelor’s degree programs in software engineering in 2003.87 Throughout this period, the United States also benefited from a system open to tapping foreign software engineering talent.88
The US Comparative Advantage over Japan in GPT Skill Infrastructure
Did GPT skill infrastructure factor decisively in Japan’s inability to keep pace with the United States in the diffusion of ICTs? Varying data reporting conventions, especially the fact that Japanese universities subsumed computer science under broader engineering fields, make it difficult to precisely quantify the US-Japan gap in software engineering skills.89 One narrative that gained momentum in the late 1980s accredited Japan’s success in high-tech manufacturing industries to its quantitative advantage in engineers. A National Science Foundation (NSF) report and two State of the Union addresses by President Ronald Reagan endorsed this belief.90 For example, the NSF’s 1997 special report on Japan’s scientific and technological capabilities declared, “By 1994, with roughly one-half the population, Japan produced more engineering and computer science degrees at the undergraduate level than the United States.”91 Such statements blended computer science degrees with all categories of engineering education.
Computer science–specific data reveal the US edge in ICT human resources. According to data from Japan’s Information Technology Promotion Agency, Japan awarded about 16,300 computer science and mathematics bachelor’s degrees in 2009, while the United States awarded 63,300 of these types of degrees that same year.92 One survey by the Japanese Information Technology Services Industry Association found that only 3.6 percent of college graduates who entered Japan’s information service industry in April 1990 received their degree from a computer science department.93 By counting immigrants entering computer-related professions and total university graduates in ICT software and hardware fields, one study estimated annual inflows into the US and Japanese ICT labor pools. In 1995, these inflows into the US ICT labor pool outpaced those in Japan by 68 percent. By 2001, this gap between the two countries’ annual inflows of ICT talent had reached almost 300 percent.94 Therefore, in the years when the US advantage in ICT diffusion was most pertinent, the skill gap between the United States and Japan in computer and software engineering talent grew even wider.95
Moreover, a computer science degree in Japan did not provide the same training as one in America. First, Japanese universities were slow to adapt to emerging trends in computer science. In both 1997 and 2007, the Information Processing Society of Japan modeled its computing curriculum revisions on American efforts that had been made six years earlier.96 The University of Tokyo, Japan’s leading university, did not establish a separate department of computer science until 1991, which was twenty-six years later than Stanford.97 Overly centralized governance of universities also inhibited the development of computer science as an independent discipline in Japan.98 As Jeffrey Hart and Sangbae Kim concluded in 2002, “The organizational and disciplinary flexibility of US universities in computer science has not been matched in any of the competing economies.”99
Software engineering presented a particular challenge for Japan. In 1988, the Japanese-language industry journal Nikkei Computer surveyed six thousand Japanese firms that used office computers. Situated outside the computer industry, these firms were involved in a broad range of fields, including materials manufacturing, finance, services, government, and education. More than 80 percent of the responding companies disclosed shortages of software programmers and designers.100 On average, outside firms provided one-quarter of their information technology personnel, and their reliance on outsourcing was magnified for programmers, system designers, and application managers.101 A nonprofit foundation for ICT development in Japan reported similar barriers to computer utilization in 1991. The survey found that companies relied heavily on computer personnel, especially software engineers, dispatched temporarily from other organizations.102 Small and medium-sized software departments, which could not afford to invest in on-the-job training, were especially disadvantaged by the lack of formal software engineering education in Japan.103
Bibliometric techniques can help substantiate the gap between the United States and Japan in skill infrastructure for software engineering. I analyzed around seven thousand publications from 1995 in the Web of Science Core Collection’s “Computer Science, Software Engineering” category.104 To gauge the breadth of institutional training in software engineering, I counted the number of Japanese and American universities that employed at least one researcher represented in this dataset. According to my estimates, the United States boasted 1.59 universities per million people that met this baseline quality of software engineering education, while Japan only had 1.17 universities per million people. This amounts to a gap of around 40 percent.
Lastly, weak industry-university linkages in computer science hampered Japan’s development of GPT skill infrastructure. Imposing centralized control over universities, the Japanese Ministry of Education, Science, and Culture (MESC) inhibited cooperation between new departments of information science and the corporate labs where much of the computing talent was concentrated.105 Japanese researchers regularly complained about the size of MESC grants, as well as the ministry’s restrictions on their ability to seek alternative funding sources. Japan’s overall budget level for university facilities in 1992 remained the same as it was in 1975. Additional government funds went instead to independent centers of excellence, diverting resources away from efforts to broaden the pool of training institutions in software engineering.106
Alternative Factors
How do alternative explanations perform in this case? A range of factors beyond GPT skill infrastructure could have influenced diverging rates of ICT diffusion in the United States and Japan. I focus on the role of external threats and varieties of capitalism because they present alternative mechanisms for how states adapted differently to the technological revolution that occurred in the IR-3.
Threat-Based Explanations
Threat-based theories struggle to account for differences in the US and Japanese technological performance in this period. A “cult of vulnerability” permeated Japan’s leaders over this period as they coped with tensions in East Asia and the oil crises of the 1970s.107 Likewise, the growth of the US “national security state,” fueled by the dangers of the Cold War, functioned as “the secret to American innovation.”108 Under his creative insecurity framework, Taylor holds up both Japan and the United States as exemplars of the IR-3 period, reasoning that they both partly owed their technological success to the galvanizing effects of a threatening international environment.109 General threat-based explanations therefore cannot explain differences in technological outcomes between the United States and Japan, namely, why the United States was more successful in ICTs than Japan.
A related argument points to the significance of US military procurement for computerization. As was the case with its influence on the American system of manufacturing in the IR-2, the US military provided the demand for initial investments in computers and semiconductors. In the 1940s and 1950s the US military was a key patron behind computing breakthroughs.110 Assured by large military procurements, innovative firms undertook risky, fundamental research that produced spillovers to many other industries. For instance, the first all-purpose electronic digital computer, the University of Pennsylvania’s electronic numerical integrator and calculator (ENIAC), was developed during World War II. The ENIAC was supported by funding from the Army Ballistics Research Laboratory, and the first program run on the computer was a simulation of the ignition of the hydrogen bomb.111
In place of the military, could other entities have served as a large demand source for ICTs? Commercial entities like Bell Labs and IBM also developed fundamental breakthroughs in ICTs. According to Timothy Bresnahan and Manuel Trajtenberg, it was “only a coincidence” that US government demand played such a pivotal role in semiconductor development.112 Others argue that while commercial advances in semiconductors and computers would likely still have occurred absent the impetus of military funding, they would have emerged after substantial delay.113
Resolving this debate depends on one’s view of the key stage in computerization. Those who emphasize the importance of military procurement often hold up the importance of first-mover advantages in the American computer industry.114 However, decades after the military helped develop the first computers and transistors, Japan had cornered the market in many related industries. The significance of military procurement is diminished when a GPT’s dissemination, as opposed to its emergence, is taken as the starting point. By 1960, the start of the IR-3 period, ICT development in the United States was already much less reliant on military support. In 1955, the demand for Bell’s transistors from two large telephone networks alone was nearly ten times more than from all military projects.115 In fact, as the commercial sector increasingly drove momentum in ICTs, military involvement arguably hindered continued advances in the commercial sector, as there was tension between the different technical cultures.116
On balance, the most significant aspect of the military’s involvement in the advance of computerization in the United States was its role in building up GPT skill infrastructure. The US military played a key role in cultivating the computer science discipline in its early years. Beginning in the 1960s, defense agencies supported academic research in computer science, such as the aforementioned Software Engineering Institute, which created centers of excellence and broadened the base of computer science education.117 From 1977 through the mid-1980s, defense funding supported more than half of academic computer science R&D.118 At the same time, military investment in computer science did not come without downsides. Defense funding was concentrated in elite research universities at the cutting edge of the field, such as Carnegie Mellon and Stanford, whereas nondefense government funding supported computer science education across a wider range of US universities.119 On the effects of military computer science funding, Stanford professor Terry Winograd wrote, “It has resulted in a highly unequal situation in which a few schools have received almost all the resources. Although this may have led to more effective research in the short run, it has also been a factor contributing to the significant long-term shortage of trained computer researchers.”120
VoC Explanations
The varieties-of-capitalism (VoC) approach provides another possible explanation for why the US economy benefited more than Japan’s from the innovations of the IR-3.121 According to the VoC framework, firms in coordinated market economies (CMEs) provide industry-specific training that is more conducive to incremental innovation, whereas worker training in more general skills in liberal market economies (LMEs) proves more favorable for radical innovations. VoC scholars point to some evidence from the international pattern of innovation during the IR-3 that supports these expectations. Based on patent data from 1983–1984 and 1993–1994, Peter Hall and David Soskice find that Germany, a CME, specialized in technology classes characterized by incremental innovation, whereas the United States, an LME, specialized in domains characterized by radical innovation.122 Therefore, the VoC perspective expects that Japan, a CME like Germany, was unable to keep up with the United States in the IR-3 because high-tech sectors such as computer software and biotechnology demanded radical innovations.123
This VoC-derived explanation provides an incomplete account of the IR-3 case. First, comprehensive empirical investigations into the innovative performance of CMEs and LMEs, especially the success of Japan as a radical innovator, undermine the explanatory power of VoC theory for the IR-3 period. Hall and Soskice’s initial analysis was based on four years of data on patent counts from only two countries, the United States and Germany. Taylor’s more extensive analysis, which covered thirty-six years (1963–1999) of patent counts and forward citations for all LME and CME countries, found that the predictions of VoC theory are not supported by the empirical data.124 In fact, contrary to the expectations of the VoC explanation, Japan was a leader in radical innovation, ranking second only to the United States in patent counts weighted by forward citations, which are a strong proxy for the radicalness of innovations.125
Second, VoC theory does not make distinctions between different types of general skills, which varied in their significance to national success in the IR-3. Regarding general training in terms of foundational schooling, Japan was making substantial improvements in average years of schooling, enrollment ratio, and access to higher education.126 GPT diffusion theory specifies the key general skills as those that best suited the advance of computerization. Consistent with these expectations, the case study evidence points to the US-Japan gap in software engineering, a set of general skills that permeated sectoral boundaries, as the crucial factor in US success with widespread computerization.
Case-Specific Factors
Other factors unique to the IR-3 case deserve further consideration. Among these alternative explanations, one popular theory was that Japan’s kanji system (Chinese-based characters) contributed to its slow adoption of computers.127 Marshall Unger, a specialist in the Japanese writing system, highlighted difficulties with representing kanji in computerized formats, which resulted in higher costs for storing data and for word-processing functions.128 American computers had to handle only ninety-five printable characters, whereas Japanese personal computers needed to store six thousand Japanese characters.129 Not only did language differences increase the cost of Japanese computers, but they also prevented Japanese adopters from using off-the-shelf computers from the United States, as these did not support Japanese language functions.
While particularities of the Japanese language may have initially hindered Japan’s computerization, it is important not to overstate the impact of this language barrier. In a review of this theory, another expert on computational linguistics argued that Unger overemphasized the additional overhead and speed costs associated with Japanese writing systems.130 Moreover, users and companies adapted over time. By the end of the 1980s, advances in processor technology allowed computers to support the greater word-processing demands of Japanese language systems.131 Therefore, during the critical years when the US-Japan computerization gap widened, the impact of the kanji system was less pronounced.
Summary
Through much of the late twentieth century, it was only a matter of time until Japan achieved economic preeminence—at least in the eyes of many analysts and scholars. Invoking the assumptions of the LS mechanism, they expected that this economic power transition would be brought about by Japan’s advantages in new sectors such as consumer electronics, semiconductor components, and computer hardware. Today, after Japan’s decade-long slowdown in productivity growth, there is virtually no discussion of it overtaking the United States as the leading economic power.
Looking back, one might be tempted to conclude that history has vindicated past critics who labeled the claims of Japan’s imminent ascension to technological hegemony as “impressionistic,”132 as well as the retrospective analyses that called out such projections for being “premature.”133
This chapter’s conclusions suggest a more nuanced interpretation. It is not that the prognoses of LS-based accounts were overeager or overly subjective. The real issue is that they were based on faulty assumptions about the pathway by which technological advances make economic power transitions possible. Indeed, the IR-3 case provides strong negative evidence against the LS mechanism, revealing that the expected outcome of an economic power transition failed to materialize in part because of the US advantage in GPT diffusion. The relative success of the United States in diffusing the trajectory of computerization across many ICT-using sectors, in line with GPT diffusion theory, was due to institutional adaptations to widen the skill base in computer engineering. In sum, the US advantage in GPT diffusion accounts for why the economic power transition expected by the LS account failed to transpire.
6 A Statistical Analysis of Software Engineering Skill Infrastructure and Computerization
I HAVE ARGUED that the shape of technological change is an overlooked dimension of the rise and fall of great powers. Most researchers point to various institutions to explain why some countries experience more scientific and technological progress than others. A central insight of this book is that the institutional factors most relevant for technological leadership depend on whether key technological trajectories map onto GPT diffusion or LS product cycles. GPT diffusion theory posits that great powers with better GPT skill infrastructure, defined as the ability to broaden the engineering skills and knowledge linked to a GPT, will more effectively adapt to technological revolutions.
This chapter evaluates a key observable implication of GPT diffusion theory. The expectation is that where there is a wider pool of institutions that can train engineering talent related to a GPT, there will be more intensive rates of GPT diffusion. Using data on computerization and a novel approach to estimate the number of universities that provide quality software engineering education in a country, this chapter first tests the theorized connection between GPT skill infrastructure and GPT adoption on time-series cross-sectional data for 19 advanced and emerging economies from 1995 to 2020. I supplement this panel analysis with two additional tests: a duration model of the speed by which 76 countries achieved a certain computerization threshold, as well as a cross-sectional regression of data on 127 countries averaged over the 1995–2020 period. While the historical case studies verified this relationship among great powers, large-n quantitative analysis allows us to explore how GPT diffusion applies beyond the chosen case studies.
To preview the findings, the evidence in this chapter backs GPT diffusion theory. Countries better positioned to widen their base of software engineering skills preside over higher rates of computer adoption. This relationship holds even when accounting for other factors that could affect computerization and different specifications of the independent and dependent variables. This chapter proceeds by operationalizing computerization rates and skill infrastructure in software engineering and then statistically testing the relationship between the two variables.
Operationalizing the Independent Variable: GPT Skill Infrastructure in Software Engineering
My key independent variable is skill infrastructure connected to computerization. The computer, a prototypical GPT, represents a natural choice for this type of inquiry, as engineering education data for many past GPTs is nonexistent for many countries.1 Plus, enough time has passed for us to see the effects of computerization. The statistical analysis focuses on the effects of skill formation institutions in software engineering, the computer science discipline tasked with training generalists in computing technology.2 Concretely, this chapter’s measure of GPT skill infrastructure captures the breadth of a country’s pool of software engineering skills and knowledge.
Efforts to measure the GPT skill infrastructure in software engineering face three challenges. First, standardized measures of human capital in computer science across countries are not available. The UNESCO Institute for Statistics (UIS) collects internationally comparable data on technicians and researchers in various fields, but this dataset does not include information specific to computer science and has limited temporal coverage.3 Second, variation across countries in the format of computer science education undercuts some potential benchmarks, such as undergraduate enrollments in computer science programs. In some countries, computer science education is subsumed under a broad engineering category, not recognized as a separate degree course.4 Lastly, comparisons of computer science education struggle to account for the quality of such training. International rankings of universities for computer science garner media coverage, but they rely on subjective survey responses about reputation and largely concentrate on elite programs.5
To the extent possible, my measure of GPT skill infrastructure addresses these obstacles. The goal is to operationalize engineering-oriented computer science education in a way that can be standardized across countries and accounts for differences in the format and quality of computer science education. My novel approach estimates the number of universities in each country that can be reasonably expected to provide a baseline quality of software engineering education. To establish this baseline in each country, I count the number of universities that employ at least one researcher who has published in a venue indexed by the Web of Science (WoS) Core Collection’s category on “Computer Science, Software Engineering.” In this category, the WoS citation database extends back to 1954 and allows for reliable cross-country comparisons based on institutional affiliations for published papers and conference proceedings.6 This approach is also insulated from distinctions related to whether certain degrees count as “computer science” programs or as “general engineering” courses. A particular university’s naming scheme has no bearing; as long as an institution retains at least one researcher who has published in the software engineering field, it counts in the GPT skill infrastructure measure.
To gather data on the number of universities that contribute to software engineering skill formation around the world, I analyze 467,198 papers from the WoS Core Collection’s “Computer Science, Software Engineering” category published between the years 1995 and 2020. I use the Bibliometrix open-source software to derive institutional and country affiliations from this corpus.7 Specifically, I collect the university and country affiliations for the corresponding authors of all 467,198 publications. For each country, I count the number of distinct university affiliations. Hypothetically, if country X’s researchers were all concentrated at a single center of excellence, it could boast more researchers represented in the corpus than country Y but still score lower on my metric. For making comparisons across countries, the number of distinct university affiliations is a better indicator of a country’s access to a broad pool of training institutions for software engineering, which is central to GPT skill infrastructure.
FIGURE 6.1. Software Engineering Skill Infrastructure by Country (2007). Source: Author’s calculations based on Web of Science Core Collection database.
I estimate a country’s GPT skill infrastructure in a particular year by averaging its score on this metric in that year along with its scores in the two previous years. This step provides checks against the risk that developments specific to a particular year—a conference cancellation, for example—may muddle the measure. To illustrate country distributions on this metric, figure 6.1 depicts the number of universities that meet my baseline for software engineering skill formation in G20 countries for one of the middle years in the dataset.
As with all bibliometric research, given the bias toward English-language papers in publication datasets, this method could undercount publications from non-English-speaking countries.8 Fortunately, this linguistic bias found in social science papers is less pronounced in engineering and mathematics papers, which comprise my dataset.9 Another factor that mitigates this bias is the very low standard for quality engineering education. Even if an institution’s researchers publish a substantial portion of their writings in a non-English language, as long as just one publication landed in the WoS “Computer Science, Software Engineering” category, that institution would still count in my definition of GPT skill infrastructure.
I considered other measures of software engineering skills, but none were suitable for this type of analysis. Data on the number of full-time equivalent telecommunication employees, collected by the ITU, shed some light on the distribution of ICT skills in various economies. Concretely, this indicator captures the total number of people employed by telecommunication operators for the provision of fixed-telephone, mobile-cellular, and internet and data services.10 The rationale is that the number of employees in this critical ICT services industry represents the broader pool of software engineering talent in a country. However, since this measure is biased toward specialists who develop and install computer technologies and other ICTs, it overlooks many engineers who intensively use ICTs in their work, even if they are not involved in developing software and computing tools.11
The time coverage of other measures was limited. For instance, the International Telecommunication Union (ITU) database on ICT skills like programming or coding in digital environments only goes back to 2019.12 In its Global Competitiveness Index, the World Economic Forum surveys business executives on the digital skills of their country’s population, but this data series starts in 2017.13
Operationalizing the Dependent Variable: Computerization Rates
Divergences among nations in scientific and technological capabilities have attracted a wide range of scholarship. While my focus is on the overall adoption rate of GPTs within economies, many scholars and government bodies have made significant contributions to quantifying national rates of innovation, often based on patenting activity, publications, and R&D investments.14 Cross-national data on the diffusion of specific technologies, by comparison, has been sparse.15 The Cross-country Historical Adoption of Technology (CHAT) dataset, which documents the intensity with which countries around the world use fifteen historically significant technologies, has helped address this deficiency.16 Other studies of cross-national technology adoption gaps quantify the diffusion of the internet and government uses of information technology.17
My primary measure of computerization is the proportion of households with a computer. These data are sourced from the International Telecommunication Union’s World Telecommunication/ICT Indicators (WTI) database.18 In this dataset, access to a computer includes use of both desktop and portable computers but excludes devices with some computing ability, such as TV sets and mobile phones.19 By estimating the number of households in a country with access to a computer, this measure elucidates cross-country differences in the intensity of computerization. Though observations for some countries start in 1984, there is limited coverage before 1995, which serves as the initial year for the data collection effort detailed in this chapter.
After the ITU was tasked with supplying indicators for access to ICTs around the world—a crucial target of the United Nations’ Millennium Development Goals (MDG) adopted in 2000—the agency started to track the number of personal computers by country.20 The ITU produces computer usage figures through two methods. First, when available, survey data from national and supranational statistical offices (such as Eurostat) are used. Though the MDG initiative has encouraged national statistical offices to help the ITU in monitoring ICT access, data coverage is still incomplete. If data on the number of households with a computer are unavailable for a country in one year, the ITU makes an estimate based on computer sales and import figures, adjusted to incorporate the average life of a computer as well as other related indicators, such as the number of computer users. For example, the computer usage indicator for Latvia comes from Eurostat in 2013, an ITU estimate in 2014, and the Central Statistical Bureau of Latvia in 2015.
Despite its limitations, I prefer the ITU’s computerization indicator over alternative measures. Francesco Caselli and Wilbur John Coleman examine the determinants of computer adoption across a large sample of countries between 1970 and 1990. To estimate the intensive margin of diffusion, they use the value of a country’s computing equipment imports per worker as a proxy for its computer investment per worker.21 However, imports do not account for a country’s computer investments sourced from a domestic computer industry; this is an issue that becomes more salient in the later years of the dataset.22
A more optimal indicator would estimate computer access and usage among businesses, since such economic activity is more likely to produce productivity improvements than household use. I examined a few alternatives. The CHAT dataset employs the number of personal computers per capita, which is one of three measures highlighted by the authors as conveying information on GPTs.23 However, this indicator still does not capture the degree of computerization in productive processes, as opposed to personal use, and has limited temporal coverage compared to the ITU’s household computerization measure.24 The OECD collects some data on ICT access and usage by businesses, but this effort did not start until 2005 and covers only OECD countries.25
Fortunately, it stands to reason that a country’s household computer adoption can serve as a proxy for its computerization rates in business activities. In the appendix, I provide further support for this claim. Comparing the available data on household and business computerization for twenty-six countries between 2005 and 2014, I find a strong correlation between these two variables (correlation coefficient = 0.8).26
Summary of Main Model Specifications
To review, this chapter tests whether GPTs diffuse more intensively and quickly in countries that have institutional advantages in widening the pool of relevant engineering skills and knowledge. The first hypothesis reads as follows:
H1: Countries with higher levels of GPT skill infrastructure in software engineering will sustain more intensive computerization rates.
With country-years as the unit of analysis, I estimate time-series cross-sectional (TSCS) models of nineteen countries over twenty-six years. Quantitative analysis permits an expansion of scope beyond the great powers covered in the case studies. As outlined in the theory chapter, differences in GPT skill infrastructure are most relevant for economies that possess the absorptive capacity to assimilate new breakthroughs from the global technological frontier.27 Since less-developed economies are often still striving to build the baseline physical infrastructure and knowledge context to access the technological frontier, variation in GPT skill infrastructure among these countries is less salient. Thus, I limit the sample to nineteen G20 countries (the excluded member is the European Union), which represent most of the world’s major industrialized and emerging economies.
Before constructing TSCS regressions, I first probe the relationship between GPT skill infrastructure in software engineering and computerization rates. Prior to the inclusion of any control variables, I plot the independent and dependent variable in aggregate to gauge whether the hypothesized effect of GPT skill infrastructure is plausible. The resulting bivariate plot suggests that there could be a strong, positive relationship between these two variables (figure 6.2).28
The basic trend in figure 6.2 provides evidence for the contention that countries better equipped with computer science skill infrastructure experience higher rates of computerization. Although these preliminary results point to a strong relationship between these two variables, further tests are needed to rule out unseen confounders that could create spurious correlations and influence the strength of this relationship. TSCS regression analysis facilitates a deeper investigation of the relationship between GPT skill infrastructure and computerization.
To control for factors that could distort the relationship between computer-related skill infrastructure and computerization, I incorporate a number of control variables in the baseline model. Rich countries may be able to spend more on computer science education; additionally, they also more easily bear the expenses of adopting new technologies, as exemplified by large disparities between developed and developing countries in information technology investment levels.29 The inclusion of GDP per capita in the model accounts for economic development as a possible confounder. I use expenditure-side real GDP at current purchasing power parities (PPPs), which is best suited to comparing relative living standards across countries. A country’s total population, another control variable, addresses the possibility that larger countries may be more likely to benefit from network effects and economies of scale, which have been positively linked to technology adoption.30 I also include the polity score for regime type. Research suggests that democratic governments provide more favorable environments for technology diffusion, and studies have confirmed this connection in the specific context of internet technologies.31
Finally, the baseline model includes two control variables that represent alternative theories of how technological changes differentially advantage advanced economies. First, I include military spending as a proportion of GDP in the regressions. The case studies have interrogated the influential view that military procurement is an essential stimulus for GPT adoption.32 By examining the relationship between military spending and computerization across a large sample of countries, the statistical analysis provides another test of this argument. Moreover, the varieties of capitalism (VoC) scholarship suggests that liberal market economies (LMEs) are especially suited to form the general skills that could aid GPT adoption across sectors. Therefore, the baseline model also controls for whether a country is designated as an LME according to the VoC typology.33
In terms of model specification, I employ panel-corrected standard errors with a correction for autocorrelation, a typical method for analyzing TSCS data.34 Given the presence of both autocorrelation and heteroskedasticity, I estimate linear models on panel data structures using a two-step Prais-Winsten feasible generalized least squares procedure.35
Time-Series Cross-Sectional Results
Table 6.1 gives the results of the three initial models, which provide further support for the theoretical expectations.36 Model 1 incorporates controls that relate to economic size and level of development. Model 2 adds a control variable for regime type. Lastly, model 3 includes a variable that represents a prominent alternative theory for how technological breakthroughs can differentially advantage certain economies. This also functions as the baseline model. In all three models, the coefficient on the GPT skill infrastructure measure is positive and highly statistically significant (p < .05).
Table 6.1 Results of Time-Series Cross-Sectional Models
Dependent Variable
Computerization
(1)
(2)
(3)
GPT skill infrastructure
3.760**
4.064***
4.227**
(1.643)
(1.676)
(1.666)
GDP per capita
29.754***
29.319***
29.435***
(3.760)
(3.737)
(3.789)
Total population
6.969***
7.046***
6.781***
(1.625)
(1.654)
(1.549)
Polity score
−0.456
−0.472*
(0.295)
(0.277)
Military spending
−0.940
(2.413)
Liberal market economy
−2.194
(3.961)
Constant
−374.599***
−368.051***
−361.885***
(60.452)
(61.173)
(58.374)
Observations
383
370
370
Note: Standard errors in parentheses.
*p < .10; **p < .05; ***p < .01
The effect of GPT skill infrastructure on GPT adoption is also substantively significant. Given the coefficient of the GPT skill infrastructure measure in the baseline model,37 a 1 percent increase in the universities per 100,000 people that provide software engineering education results in an increase of the computerization rate by 0.042 percentage points.38 Though the substantive effect seems small at first glance, its magnitude becomes clear when contextualized by differences in GPT skill infrastructure across the sample. For example, in China over this time period, the average number of universities per 100,000 people that met my baseline for GPT skill infrastructure was 0.040. The corresponding figure for the United States was 0.248. According to the coefficient estimate for GPT skill infrastructure, this difference of 520 percent corresponds to a difference of nearly 22 percentage point units in the computerization rate.
It should be noted that only two control variables, economic development and population, came in as statistically significant in the baseline model. As expected, wealthier countries and more populous countries presided over more intensive adoption of computing technologies. The null result for regime type is worth highlighting, as the effects of democracy on technology adoption are disputed.39 Finally, contrary to the expectations of competing explanations to GPT diffusion theory, the effects of military spending and VoC are insignificant. This is consistent with the findings from the historical case studies.
Quantitative appendix table 1 displays the results after incorporation of three additional controls. First, trade linkages expose countries to advanced techniques and new ideas, opening the door to technology diffusion. A high level of trade openness has been associated with more intensive adoption of information technologies.40 Relatedly, there is evidence that a country’s openness to international trade has a positive and sizable effect on various measures of innovation, including high-technology exports, scientific publications, and patents.41 Second, higher urban density has been linked to faster diffusion of technologies such as the television and the internet.42 Model 8 incorporates a trade openness variable and an urbanization variable.
Third, patterns at the regional level could shape how computerization spreads around the world. Scholars have identified such regional effects on the diffusion of ideas, policies, and technologies.43 In model 9, I assess spatial dynamics with dummy variables for the following regions: East Asia and Pacific; Europe and Central Asia; Latin America and Caribbean; the Middle East and North Africa; North America; South Asia; and sub-Saharan Africa.44 The positive effect of GPT skill infrastructure on computerization stays strong and highly statistically significant across these two models.
To ensure that the results were not determined by my choice of independent variable, I constructed an alternative specification of GPT skill infrastructure. Reanalyzing data on 467,198 software engineering publications, I counted the number of distinct authors for each country, as a proxy for the breadth of researchers who could foster human capital in software engineering. Though I still maintain that the primary specification best captures software engineering skill infrastructure, this alternative construction guards against possible issues with institution-based indicators, such as problems with institutional disambiguation and nonstandardized author affiliations.45 With this alternative specification, the estimated effect of GPT skill infrastructure on computerization remains positive and significant for the baseline model as well as for models with additional controls.46
Duration Analysis
When it comes to whether great powers can harness the potential of GPTs for productivity growth, the speed of adoption—not just the intensity of adoption—is pertinent. In the historical case studies, technological leaders adapted more quickly to industrial revolutions because of their investments in widening the base of engineering knowledge and skills associated with GPTs. This leads to the second hypothesis.
H2: Countries with higher levels of GPT skill infrastructure in software engineering will experience faster levels of computer adoption.
In testing this hypothesis, the dependent variable shifts to the amount of time it takes for a country to reach a particular computerization rate. A critical step is to establish both the specific computerization rate that constitutes successful “adoption” of computers as well as when the process of diffusion begins. Regarding the former, I count the “first adoption” of computerization as when the proportion of households with a computer in a country reaches 25 percent. This approach is in line with Everett Rogers’s seminal work on the S-shaped curve for successful diffusion of an innovation, which typically takes off once the innovation reaches a 10 to 25 percent adoption rate.47 For the duration analysis, since many of the countries enter the dataset with levels of computer adoption higher than 10 percent, the 25 percent level threshold is more suitable.48
I take 1995 as the starting point for the diffusion of computers as a GPT. Though an earlier date may be more historically precise, the 1995 date is more appropriate for modeling purposes, as the data on computerization rates for countries before this time are sparse. In a few cases, a country clearly achieved the 25 percent computerization threshold before 1995.49 As a practical measure to estimate the duration models, I assume that the time it took for these countries to adopt computers was one year. Right-censoring occurs with the last year of data, 2020, as many countries still had not reached the 25 percent computerization rate.
Using these data, I employ a Cox proportional hazards model to estimate the time it takes for countries to reach a 25 percent computerization rate based on the start date of 1995. Often used by political scientists to study conflict duration or the survival of peace agreements, duration models are also commonly used to investigate the diffusion of new technologies and to determine why some firms take longer to adopt a certain technology than others.50 Freed from the demands of TSCS analysis for yearly data, the duration analysis expands the county coverage, incorporating all upper-middle-income economies or high-income economies, based on the World Bank’s income group classifications.51 The resulting sample, which excludes countries that never attained the 25 percent computerization threshold, includes seventy-six countries.
Table 6.2 reports the estimated coefficients from the duration analysis. Positive coefficients match with a greater likelihood of reaching the computerization threshold. I use the same explanatory variable and controls as the baseline model from the TSCS analysis. These variables all enter the model with their measures in 1995. Model 4a takes the 25 percent computerization rate as the adoption threshold, while model 4b adjusts it to 20 percent to ensure that this designation is not driving results.
Table 6.2 Time to Computerization by Country
Dependent Variable
25% Threshold(4a)
20% Threshold(4b)
GPT skill infrastructure
0.673***
0.517***
(0.137)
(0.119)
GDP per capita
1.186***
1.110***
(0.335)
(0.288)
Total population
0.127
0.062
(0.085)
(0.074)
Polity score
0.022
0.024
(0.025)
(0.023)
Military spending
0.017
−0.042
(0.218)
(0.198)
Liberal market economy
0.760
0.785
(0.503)
(0.493)
N (number of events)
76 (74)
83 (83)
Likelihood ratio test (df = 6)
112.9***
111.2***
Note: Standard errors in parentheses.
*p < .10; **p < .05; ***p < .01
As the models demonstrate, the effect of GPT skill infrastructure on the speed by which countries achieve computerization is positive and highly statistically significant, providing support for hypothesis 2. Based on model 4a’s hazard ratio for the independent variable (1.96) for a given year, a tenfold increase in software engineering university density doubles the chances of a country reaching the computerization threshold.52 These results hold up after introducing additional control variables (quantitative appendix table 3).53
A Cross-Sectional Approach: Averages across the 1995–2020 Period
As an additional check, I collapse the panel dataset into cross-sectional averages of GPT skill infrastructure and computerization over the 1995–2020 period in a large sample of countries. In certain aspects, cross-sectional evidence could be more appropriate for comprehending the impact of features, like skill formation institutions, that are difficult to capture in yearly intervals because they change gradually.54 This approach allows for more countries to be included, as the yearly data necessary for TSCS analysis were unavailable for many countries. Limiting the sample based on the same scope conditions as the duration analysis leaves 127 countries.
I also include the same set of controls used in the previous analyses of the panel data: GDP per capita, total population, regime type, military spending, and liberal market economies. I employ ordinary least squares (OLS) regression to estimate the model. Since both a scale-location plot and a Breausch-Pagan test demonstrate that heteroskedasticity is not present in the data, it is appropriate to use an OLS regression estimator with normal standard errors.
The results of the regression analysis provide further support for the theoretical expectations. To recap, the independent variable is the estimated average skill infrastructure for software engineering between 1995 and 2020, and the dependent variable is the average computerization rate during the same period. Since analyzing bibliographic information on yearly software engineering publications for 127 countries is a demanding exercise, I estimated the average number of universities that nurture software engineering skills based on publications for the two middle years (2007, 2008) in the dataset, instead of deriving this average based on publication data across the entire time range.55 Table 6.3 displays the results, with the incremental inclusion of control variables, also averaged over the period 1995–2020, across three models.56 The coefficient on the GPT skill infrastructure measure remains positive and highly statistically significant across all three models (p < .01).
Table 6.3 GPT Skill Infrastructure Predicts More Computerization
Dependent Variable
Computerization
(5)
(6)
(7)
GPT skill infrastructure
3.211***
3.737***
3.761***
(0.528)
(0.609)
(0.649)
GDP per capita
16.617***
15.536***
14.977***
(1.564)
(1.723)
(1.812)
Total population
−1.647***
−0.739
−0.831*
(0.440)
(0.485)
(0.500)
Polity score
−0.066
−0.070
(0.147)
(0.180)
Military spending
0.577
(1.604)
Liberal market economy
5.017
(4.269)
Constant
−83.099***
−86.165***
−79.773***
(20.577)
(22.316)
(23.636)
Observations
127
112
110
R2
0.812
0.833
0.834
Note: Standard errors in parentheses.
*p < .10; **p < .05; ***p < .01
I perform several additional tests to confirm the robustness of the results. I first include the same additional controls used in the preceding TSCS analysis. Quantitative appendix table 4 shows that the main findings are supported. One limitation of models that rely on cross-sectional averages is endogeneity arising from reverse causality. In other words, if greater diffusion of computers throughout the economy spurs more investment in institutions that broaden the pool of software engineers, then this could confound the baseline model’s estimates. To account for this possibility, in model 16 in quantitative appendix table 5, I operationalize GPT skill infrastructure using the estimate for the year 1995, the start of the period, instead of its average level over the 1995–2020 period.57 Thus, this model captures the impact of GPT skill infrastructure in 1995 on how computerization progressed over the remaining sample years.58 The effect remains positive and statistically significant.
While this chapter’s primary aim is to investigate empirical patterns expected by GPT diffusion theory, the quantitative analysis can also probe whether computerization is positively influenced by institutions that complement LS product cycles. To that end, I add a control variable that stands in for the institutional competencies prioritized by the LS model.59 Measures of computer exports and ICT patents serve as two ways to capture a country’s ability to generate novel innovations in the computer industry.60 In the resulting analysis, the LS-linked variables do not register as statistically significant.61
These results should be interpreted with care. In many cases, measures of institutional capacity to build a strong, innovative computer sector may be highly correlated with measures of GPT skill infrastructure. Because the statistical approach struggles to differentiate between the causal processes that connect these two features and computerization, the historical case studies take on the main burden of comparing the GPT diffusion and LS mechanisms against each other. Still, the inclusion of variables linked with the LS mechanism does suggest that, in the context of computerization, there is ineffectual evidence that the presence of a strong leading sector spills over into other sectors and generates multiplier effects—a key observable implication of LS accounts.62 Additionally, these models drive home the importance of differentiating between institutions linked to innovative activities (in the sense of introducing new products and processes) and engineering-oriented institutions, like GPT skill infrastructure, which are more connected to adoptive activities.63
Summary
Using a variety of statistical methods, this chapter tested the expectation that countries better equipped to widen the base of engineering talent in a GPT will be more successful at diffusing that GPT throughout their economies. The combination of TSCS models, duration analysis, and cross-sectional regressions lends credence to the strength of the relationship between GPT skill infrastructure and computerization. The results hold across a range of additional tests and robustness checks.
There are two major limitations to this chapter’s approach. First, the statistical analysis should be interpreted mainly as an independent evaluation of GPT diffusion theory, not as an additional comparison between GPT diffusion theory and the causal pathway linked to LS product cycles. In the large-scale statistical analysis, indicators linked with the LS mechanism can also be associated with higher computerization rates, making it difficult to weigh the two explanations against each other. The rich historical detail in the case studies therefore provides the prime ground for tracing causal mechanisms.
Second, this chapter evaluates only one aspect of GPT skill infrastructure. A more comprehensive assessment would include not just the capacity to widen the pool of software engineering talent, which was the independent variable in this analysis, but also the strength of information flows between the GPT sector and application sectors. For instance, in the IR-2 case, both the United States and Germany trained large numbers of mechanical engineers, but American technological institutes placed more emphasis on practical training and shop experience, which strengthened connections between the US mechanical engineering education system and industrial application sectors. Building on data collection efforts that are starting to measure these types of linkages, such as the proportion of publications in a technological domain that involve industry-academia collaborations, future research should conduct a more complete assessment of GPT skill infrastructure.64
Notwithstanding these limitations, the quantitative analysis backs a key observable implication of GPT diffusion theory: advanced economies’ level of GPT skill infrastructure is strongly linked to GPT adoption rates. Not only does this provide some initial support for the generalizability of this book’s central argument beyond just great powers, but it also gives further credibility to GPT diffusion theory’s relevance to US-China competition today.
7 US-China Competition in AI and the Fourth Industrial Revolution
THE FIRST MACHINE to defeat a human Go champion; powerful language models that can understand and generate humanlike text; a computer program that can predict protein structures and enable faster drug discovery—these are just a few of the newest discoveries in AI that have led some to declare the arrival of a Fourth Industrial Revolution (IR-4).1 As for the latest geopolitical trends, China’s rise has been the dominant story of this century as national security establishments grapple with the return of great power competition. Located squarely at the intersection of these two currents, the US-China technological rivalry has become an inescapable topic of debate among those interested in the future of power—and the future itself.
Who will lead the way in the Fourth Industrial Revolution? To answer this question, leading thinkers and policymakers in the United States and China are both drawing lessons from past technology-driven power transitions to grapple with the present landscape. Unfortunately, as this chapter argues, they have learned the wrong lessons. Specifically, the leading-sector perspective holds undue influence over thinking about the relationship between technological change and the possibility of a US-China power transition. Yet careful tracing of historical cases and statistical analysis have revealed that the GPT mechanism provides a better model for how industrial revolutions generate the potential for a power transition. When applied to the IR-4 and the evolving US-China power relationship, GPT diffusion theory produces different insights into the effects of the technological breakthroughs of today on the US-China power balance, as well as on the optimal strategies for the United States and China to pursue.
This chapter sketches out the potential impacts of today’s emerging technologies on the US-China power balance. I first describe the current productivity gap between the United States and China, with particular attention to concerns that the size of this gap invalidates analogies to previous rising powers. Next, I review the array of technologies that have drawn consideration as the next GPT or next LS. Acknowledging the speculative nature of technological forecasting, I narrow my focus to developments in AI because of its potential to revitalize growth in ICT industries and transform the development trajectories of other enabling technologies.
The essence of this chapter is a comparison of the implications of the LS and GPT mechanisms for how advances in AI will affect a possible US-China economic power transition. In contrast to prevailing opinion, which hews closely to the LS template, GPT diffusion theory suggests that the effects of AI on China’s rise will materialize through the widespread adoption of AI across many sectors in a decades-long process. The institutional factors most pertinent to whether the United States or China will more successfully benefit from AI advances are related to widening the base of AI-related engineering skills and knowledge. I also spell out how the implications of GPT diffusion theory for the US-China power balance differ from those derived from alternative explanations.
The objective here is not to debate whether China will or will not catch up to the United States, or whether technological capabilities are more significant than all other considerations that could affect China’s long-term economic growth. Rather, this chapter probes a more limited set of questions: If emerging technologies were to significantly influence the US-China economic power balance, how would this occur? Which country is better positioned to take advantage of the Fourth Industrial Revolution? What would be the key institutional adaptations to track?2
A Power Transition in Progress?
Over the past four decades, there has been no greater shift in the global economic balance than China’s rise. China is either already the world’s largest economy, if measured by purchasing power parity (PPP) exchange rates, or is projected to soon overtake the United States, based on nominal exchange rates.3 China’s impressive economic growth has led many to proclaim that the era of US hegemony is over.4
Economic size puts China in contention with the United States as a great power competitor, but China’s economic efficiency will determine whether a power transition occurs. Countries like Switzerland outpace the United States on some measures of economic efficiency, but they lack the economic size to contend. Other rising powers, such as India, boast large economies but lag far behind in economic efficiency. Mike Beckley concludes: “If the United States faces a peer competitor in the twenty-first century … it will surely be China.”5 For this conditional to be true, China’s productivity growth is critical.6 This is not the first time in history that China has boasted the world’s largest economy; after all, it held that distinction even as Britain was taking over economic leadership on the back of the First Industrial Revolution.
Where does China currently stand in comparison to the productivity frontier? Based on 2018 figures, China’s real GDP per capita (at 2010 PPPs) is about 30 percent that of the United States.7 From 2000 to 2017, China’s total factor productivity (TFP) never surpassed 43 percent of US TFP (figure 7.1).8 In 2015, labor productivity in China remained at only 30 percent of that in the United States, though this figure had doubled over the past two decades.9
These numbers suggest that China sits much further from the productivity frontier than past rising powers. If the US-China power relationship is fundamentally different from those in previous eras, the relevance of conclusions from previous cases could be limited.10 For instance, in the early years of the IR-1, Britain was only slightly behind the Netherlands, the productivity leader at the time. The United Kingdom’s GDP per capita was 80 percent of Dutch GDP per capita in 1800.11 Similarly, at the beginning of the IR-2, the United States trailed Britain in productivity by a small margin. In 1870, labor productivity and TFP in the United States were 90 percent and 95 percent, respectively, of labor productivity and TFP in Britain.12 During the 1870s, average GDP per capita in the United Kingdom was about 15 percent higher than average GDP per capita in the States.13
FIGURE 7.1. US-China Productivity Gap (2000–2017). Source: Penn World Table version 9.1; Feenstra, Inklaar, and Timmer 2015.
Still, that China could surpass the United States in productivity leadership is not outside the realm of possibility.14 The IR-3 case is a better comparison point for the current productivity gap between the United States and China. In 1960, the start of the IR-3 case, Japanese GDP per capita was 35 percent of US GDP.15 At the time, Japan’s TFP was 63 percent of US TFP, and Japan’s labor productivity was only 23 percent of US labor productivity, a lower proportion than the China-US ratio at present.16 Despite the initial chasm, the TFP gap between the United States and Japan narrowed to only 5 percent in 1991.17
Indeed, productivity growth is crucial for China to sustain its economic rise in the long term. For the 1978–2007 period, Xiaodong Zhu decomposed the sources of China’s economic growth into labor deepening, human capital, capital deepening, and total factor productivity growth. He found that growth in TFP accounted for 78 percent of China’s growth in GDP per capita.18 The burden on TFP improvements will only increase, given the diminishing impact of other drivers of China’s growth miracle, such as urbanization and demographic dividends.19
Whether China can sustain productivity growth is an open question. Plagued by inefficient infrastructure outlays, China’s aggregate TFP growth declined from 2.8 percent in the decade before the global financial crisis to 0.7 percent in the decade after (2009–2018).20 If calculated using alternative estimates of GDP growth, China’s TFP growth was actually negative from 2010 to 2017, averaging −0.5 percent.21 This productivity slowdown is not unique to China. Even before the 2008 global financial crisis, there was a slowdown in TFP growth among advanced economies due to waning effects from the information and communications technologies (ICTs) boom.22 China’s labor productivity growth also decelerated from 8.1 percent in 2000–2007 to 4.2 percent in 2011–2019, but the later period’s slowed growth rate was still six times greater than the US labor productivity growth rate in the same period.23
Adaptation to technological advances will be central to China’s prospects of maintaining high rates of productivity growth. China’s leaders worry about getting stuck in the “middle-income trap,” a situation in which an economy is unable to advance to high-income status after it exhausts export-driven, low-cost manufacturing advantages. Many studies have stressed the linkage between China’s capacity to develop and absorb emerging technologies and its prospects for escaping the middle-income trap.24 The Chinese government also increasingly pushes the development and adoption of information technology and other cutting-edge technologies as a way to increase TFP.25 Thus, tracking China’s future productivity growth necessitates a better understanding of specific technological trajectories in the current period.
Key Technological Changes in the IR-4
ICTs, the key technological drivers of the IR-3, are still at the heart of the IR-4. From visionaries and daydreamers to economists and technology forecasters, there is a wide-ranging consensus that AI will breathe new life into the spread of digitization. The World Economic Forum calls AI the “engine that drives the Fourth Industrial Revolution.”26 Kai-Fu Lee, the former head of Google China, boldly asserts, “The AI revolution will be on the scale of the Industrial Revolution but probably larger and definitely faster.”27 To further explore AI’s role in the IR-4, I examine this technological domain as a source of both GPT and LS trajectories.
Candidate Leading Sectors
LS accounts forecast that in future waves of technological change ICTs will continue to drive economic transformation. According to one analysis of US-China technological rivalry in the twenty-first century, ICTs are “widely regarded as the current leading sector.”28 I reviewed five key texts that predicted future leading sectors, all written by scholars who study historical cycles of technological change and global leadership transitions.29 These forecasts also highlight other candidate leading sectors, including lasers and new sources of energy, but they converge on ICTs as the leading sector of the next wave of technological disruption.
Informed by the LS model, AI’s effects on global technological competition are often framed through its potential to open up new opportunities for latecomers to catch up and leapfrog advanced countries in key segments like AI chips. China’s national AI development plan outlines its ambition to become the world’s leading center of innovation in AI by 2030.30 Scholars analyze China’s capacity to develop global intellectual monopolies in certain AI applications and enhance independent innovation in AI so as to guard against other countries leveraging weaponized interdependence.31 Descriptions of China’s AI strategy as aimed toward seizing “the commanding heights” of next-generation technologies reflect the belief that competition in AI will be over global market shares in strategic sectors.32
Candidate GPTs
Among the possible GPTs that could significantly impact a US-China economic power transition, AI stands out. Like the literature on leading sectors, the GPT literature also converges on ICTs as a continued driver of technological revolution. Kenneth Carlaw, Richard Lipsey, and Ryan Webb, three pioneers of GPT-based analysis, identify programmable computing networks as the basic GPT that is driving the modern ICT revolution.33 Crucially, AI could open up a new trajectory for this ICT revolution. Recent breakthroughs in deep learning have improved the ability of machines to learn from data in fundamental ways that can apply across hundreds of domains, including medicine, transportation, and other candidate GPTs like biotechnology and robotics. This is why AI is often called the “new electricity”—a comparison to the prototypical GPT. Economists regard it as the “next GPT”34 and “the most important general-purpose technology of our era.”35
Several studies have found evidence for a GPT trajectory in AI. One study, using a novel dataset of preprint papers, finds that articles on deep learning conform with a GPT trajectory.36 Using patent data from 2005 to 2010 to construct a three-dimensional indicator for the GPT-ness of a technology, Petralia ranks technological classes based on their GPT potential.37 His analysis finds that image analysis, a field that is closely tied to recent advances in deep learning and AI, ranks among the top technological classes in terms of GPT-ness.38 Another effort employs online job posting data to differentiate among the GPT-ness of various technological domains, finding that machine learning technologies are more likely to be GPTs than other technologies such as blockchain, nanotechnology, and 3D printing.39
To be sure, forecasts of future GPTs call attention to other technological trends as well. Other studies have verified the GPT potential of biotechnology.40 Robotics, another candidate GPT for the IR-4, could underpin “the next production system” that will boost economy-wide productivity, succeeding the previous one driven by information technology.41 While my primary reasons for limiting my analysis to AI are based on space constraints as well as on its prominence in the surrounding literature, it is also important to note that developments in both biotechnology and robotics are becoming increasingly dependent on advances in deep learning and big data.42
The Limits of Technological Foresight
Unlike exercises to pinpoint key technologies of previous industrial revolutions, which benefited from hindsight, identifying technological drivers in the IR-4 is a more speculative exercise. It is difficult to find true promise amid the hype. The task is made harder by the fact that even experts and technological forecasting bodies regularly miss the next big thing. In 1945, a team led by Dr. Theodore von Kármán, an eminent aerospace engineer, published Toward New Horizons, a thirty-two-volume text about the future of aviation. The study failed to foresee major new horizons such as the first human in space, intercontinental ballistic missiles (ICBMs), and solid-state electronics—all of which emerged within fifteen years.43 In the early 1990s, the US Army conducted a technology forecast assessment to identify the technologies most likely to transform ground warfare in the next century. When the forecast was evaluated in 2008 by the Army’s senior scientists and engineers, it graded out at a “C.” Among its most significant misses was the development of the internet.44
I am no Cassandra. It is very possible that if I were writing this book in 2000, this chapter would focus on the promise of nanotechnology, not AI. At that time, President Bill Clinton had just unveiled the National Nanotechnology Initiative. In a 2003 speech, Philip J. Bond, the undersecretary for technology at the Department of Commerce at the time, declared:
Nano’s potential rises to near Biblical proportions. It is not inconceivable that these technologies could eventually achieve the truly miraculous: enabling the blind to see, the lame to walk, and the deaf to hear; curing AIDS, cancer, diabetes and other afflictions; ending hunger; and even supplementing the power of our minds, enabling us to think great thoughts, create new knowledge, and gain new insights.45
Decades later, there is a collective exhaustion around the hype surrounding nanotechnology, a phenomenon one scientist calls “nanofatigue.”46
One lesson that stands out from mapping the technological landscape in past industrial revolutions is that the most significant GPTs of an era often have humble origins. In the IR-2, new innovations in electricity and chemicals garnered the most attention, but America’s economic rise owed more to advances in machine tools that were first introduced decades earlier. In the same way, “old” GPTs like electricity could still shock the world.47 Today there is still a lot of potential for expanded industrial electrification, which could have a major impact on productivity.48 Similarly, high-capacity battery technologies could transform productivity on a broad scale.49 Interestingly, patent data also demonstrate the continued importance of electrical technologies. Among the top ten GPT candidates as ranked by Petralia’s indicator of “GPT-ness,” there were as many technological classes in the electrical and electronic category as there were in the computers and communications category.50
For my purposes, it is reassuring that developments in AI also draw on a long history. In the United States, the legitimization of AI as an important field of research dates back to the 1960s.51 Thus, though the rest of the chapter takes AI as the most important GPT for the near future, it does so with a humble mindset, acknowledging that looking forward to the future often starts with digging deeper into the past.
The GPT vs. LS Mechanisms in the IR-4
There has been no shortage of speculation about whether the United States or China is better fit for the new AI revolution. Each week it seems there is a new development in the “AI arms race” between the two nations.52 Many believe that China is an AI superpower on the verge of overtaking the United States in the key driver of the IR-4.53 As the following sections will show, these discussions tend to follow the LS template in their assumptions about the trajectory of AI development and key institutional adjustments.
Conversely, GPT diffusion theory provides an alternative model for how AI could affect the US-China power balance, with implications for the optimal institutional adaptations to the AI revolution. I conclude that, if the lessons of past industrial revolutions hold, the key driver of a possible US-China economic power transition will be the relative success of these nations in diffusing AI throughout their economies over many decades. This technological pathway demands institutional adaptations to widen the base of AI engineering skills and knowledge. While recognizing that international competition over AI is still in the early innings, this chapter outlines a preliminary framework for assessing which country’s roster is better equipped for success.
Impact Timeframe: The Decisive Years in the US-China AI Competition
If guided by the LS mechanism, one would expect the impacts of AI on US-China power competition to be very significant in the early stages of the technology’s trajectory. Indeed, many prominent voices have articulated this perspective. Consider, for example, a report titled “Is China Beating the US to AI Supremacy?,” authored by Professor Graham Allison, the director of Harvard Kennedy School’s Belfer Center for Science and International Affairs and Eric Schmidt, former CEO of Google and cochair of the National Security Commission on Artificial Intelligence (NSCAI). For Allison and Schmidt, the decisive years in US-China AI competition are just around the corner. Assuming that AI advances will be rapidly adopted across many economic domains, their aim is to “sound an alarm over China’s rapid progress and the current prospect of it overtaking the United States in applying AI in the decade ahead.”54 Shaped by a similar framework, other influential texts also predict that China’s productivity boost from AI will come to fruition in the 2020s.55
If GPT diffusion theory serves as the basis for analysis, these influential texts severely underestimate the time needed for economic payoffs from AI. Historical patterns of GPT advance have borne out that, even in early adopter countries, it takes at least three or four decades for these fundamental technologies to produce a significant productivity boost.56
Using this pattern as guidance, we can roughly project AI’s impact timeframe, after establishing an initial emergence date for this GPT. In 2012, the AlexNet submission to ImageNet, a competition that evaluates algorithms on large-scale image classification, is widely recognized as spurring this current, deep learning–based paradigm of AI development.57 If using the metric of when a GPT achieves a 1 percent adoption rate in the median sector, the AI era probably began in the late 2010s.58 As of 2018, according to the most recent census survey on the extent of AI adoption in the US economy, only 2.75 percent of firms in the median sector reported using AI technologies.59 Thus, regardless of which arrival date is used, if AI, like previous GPTs, requires a prolonged period of gestation, substantial productivity payoffs should not materialize until the 2040s and 2050s.60
Of course, other factors could affect AI’s expected impact timeframe, including the possibility that the general process of technological adoption is accelerating. Some evidence indicates that the waiting time for a significant productivity boost from a new GPT has decreased over time.61 Lee argues that the AI revolution will be faster than previous GPT trajectories owing to the increasingly frictionless distribution of digital algorithms and more mature venture-capital industry.62 Nevertheless, preliminary evidence suggests that AI will face similar implementation lags as previous GPTs, including bottlenecks in access to computing resources, human capital training, and business process transformations.63
Phase of Relative Advantage: Innovation-centrism and China’s AI Capabilities
Debates about China’s scientific and technological power reduce complex dynamics to one magic word—“innovation.”64 Whether China can generate novel technologies is often the crux of debates over China’s growing scientific and technological capabilities and a potential US-China power transition.65 For David Rapkin and William Thompson, the prospect of China overtaking the United States as the leading power is dependent on “China’s capacity to innovate”—specifically as it relates to revolutionary technological changes that allow challengers to leapfrog the leader in economic competition.66 “If … China’s innovativeness continues to lag a considerable distance behind that of the US, then China overtaking the US might wait until the twenty-second century,” they posit.67 China’s innovation imperative, as Andrew Kennedy and Darren Lim describe it in language common to LS analysis, is motivated by “monopoly rents generated by new discoveries.”68
Innovation-centric views of China’s AI capabilities paint an overly optimistic picture of China’s challenge to US technological leadership. Allison and Schmidt’s Belfer Center paper, for instance, emphasizes China’s growing strength in AI-related R&D investments, leading AI start-ups, and valuable internet companies.69 Likewise, the NSCAI’s final report suggests that China is poised to overtake the United States in the capacity to generate new-to-the-world advances in AI, citing shares of top-cited, breakthrough papers in AI and investments in start-ups.70 These evaluations match up with viewpoints that are bullish on China’s overall technological capabilities, which also point to its impressive performance along indicators of innovation capacity, such as R&D expenditures, scientific publications, and patents.71
Some other comparisons of US and Chinese AI capabilities arrive at the opposite conclusion but still rely on the LS template. For instance, two Oxford scholars, Carl Frey and Michael Osborne, have likened claims that China is on the verge of overtaking the United States in AI to overestimates of Japan’s technological leadership in computers in the 1980s. In their view, just like Japan, China will fail to overtake the United States as the world’s technological leader because of its inability to produce radical innovations in AI. In fact, they claim, the prospects are even bleaker this time around: “China, if anything, looks less likely to overtake the United States in artificial intelligence than Japan looked to dominate in computers in the 1980s.”72
If analysis of US-China competition in AI was centered on GPT diffusion theory, it would focus more on China’s capacity to widely adopt AI advances. In this scenario, it is neither surprising nor particularly alarming that China, like other great power contenders such as Japan in the IR-3, Germany in the IR-2, and France in the IR-1, contributes to fundamental innovations. No one country will corner all breakthroughs in a GPT like AI. The key point of differentiation will be the ability to adapt and spread AI innovations across a wide array of sectors.
A diffusion-centric perspective suggests that China is far from being an AI superpower. Trends in ICT adoption reveal a large gap between the United States and China. China ranks eighty-third in the world on the International Telecommunication Union’s ICT development index, a composite measure of a country’s level of networked infrastructure, access to ICTs, and adoption of ICTs.73 By comparison, the United States sits among the world’s leaders at fifteenth. Though China has demonstrated a strong diffusion capacity in consumer-facing ICT applications, such as mobile payments and food delivery, Chinese businesses have been slow to embrace digital transformation.74
In fact, it is often Chinese scholars and think tanks that acknowledge these deficiencies. According to an Alibaba Research Institute report, China significantly trails the United States in penetration rates of many digital technologies across industrial applications, including digital factories, industrial robots, smart sensors, key industrial software, and cloud computing.75 China also significantly trails the United States in an influential index for adoption of cloud computing, which is essential to implementing AI applications.76 In 2018, US firms averaged a cloud adoption rate of over 85 percent, more than double the comparable rate for Chinese firms.77
To be fair, China has achieved some success in adopting robots, a key application sector of AI. China leads the world in total installations of industrial robots. Aided by favorable industry composition and demographic conditions, China added 154,000 industrial robots in 2018, which was more than were installed by the United States and Japan combined.78 Based on 2021 data from the International Federation of Robotics, China outpaces the United States in robot density as measured by the number of industrial robots per 10,000 manufacturing employees.79
However, China’s reputed success in robot adoption warrants further scrutiny. The IFR’s figures for employees in China’s manufacturing sector significantly underestimate China’s actual manufacturing workforce. If these figures are revised to be more in line with those from the International Labor Organization (ILO), China’s robot density would fall to less than 100 robots per 10,000 manufacturing employees, which would be around one-third of the US figure.80 On top of that, talent bottlenecks hamper robot diffusion in China, since skilled technicians are required to reprogram robots for specific applications.81 An unused or ineffective robot counts toward robot density statistics but not toward productivity growth.
Breadth of Growth: Picking Winners vs. Horizontal Approaches to AI Development
Divergent perspectives on the breadth of growth in technological revolutions also separate LS-based and GPT-based views of the US-China case. If technological competition in the IR-4 is limited to which country gets a bigger share of the market in new leading industries like AI, then direct sectoral interventions in the mold of China’s AI strategy could be successful. However, if the breadth of growth in the IR-4 follows the GPT trajectory of the three previous industrial revolutions, another approach will be more effective.
China’s AI strategy has hewed closely to the LS model. This approach builds off a series of directives that prioritize indigenous innovation in select frontier technologies, an emphasis that first appeared in the 2006 “National Medium- and Long-Term Plan for the Development of Science and Technology” and extends through the controversial “Made in China 2025” plan.82 Since the mid-2000s, the number of sectoral industrial policies issued by the State Council, China’s cabinet-level body, has significantly increased.83 Appropriately, the State Council’s 2017 AI development plan outlined China’s ambitions to become the world’s primary innovation center for AI technology.84
On the breadth of growth dimension, tension between GPT diffusion theory and China’s application of the LS template is rooted in differing expectations for how the economic boost from revolutionary technologies will unfold. Take, for example, China’s 2010 “Strategic Emerging Industries” (SEI) initiative, which targets seven technological sectors based on opportunities for China to leapfrog ahead in new industries.85 Oriented around assumptions that a limited number of technologically progressive industries will drive China’s future growth, the SEI defines success based on the resultant size of these industries, as measured by their value added as a share of GDP.86
In contrast, GPT diffusion theory expects that, in the country that best capitalizes on the IR-4, productivity growth will be more dispersed. In this view, the AI industry never needs to be one of the largest, provided that AI techniques trigger complementary innovations across a broad range of industries. Relatedly, some Chinese thinkers have pushed back against industrial policies that favor narrow technology sectors. A research center under China’s own State Council, in a joint analysis with the World Bank, concluded in 2012: “A better innovation policy in China will begin with a redefinition of government’s role in the national innovation system, shifting away from targeted attempts at developing specific new technologies and moving toward institutional development and an enabling environment that supports economy-wide innovation efforts within a competitive market system.”87 The economy-wide transformation enabled by AI, if it lives up to its potential as a GPT, demands a more broad-based response.
When it comes to technology policy, there is always a push and pull between two ends of a spectrum. Vertical industrial policy, or “picking winners,” targets certain technologies, often leading to top-down intervention to ensure that the nation’s firms are competitive in specific industries. Horizontal industrial policy promotes across-the-board technological development and avoids labeling certain technologies as more significant than others. This book argues that both camps have it partly right, at least when it comes to ensuring long-term economic growth in times of technological revolution. Picking technological winners is needed in that some technologies do matter more than others; however, the “winners” are GPTs, which require horizontal industrial policies to diffuse across many application sectors. Institutions for skill formation in AI engineering, the subject of the next section, split the difference between these two approaches.
Institutional Complementarities: GPT Skill Infrastructure in the IR-4
In 2014, Baidu, one of China’s leading tech giants, hired Andrew Ng away from Google, poaching the cofounder of Google’s deep learning team. Three years later, Baidu lured Qi Lu away from Microsoft, where he had served as the architect of the company’s AI strategy. Their departures were headline news and spurred broader discussions about China’s growing AI talent.88
When Alibaba, another one of China’s tech giants, celebrated its listing on the Hong Kong stock exchange in November 2019, it showcased a different form of AI talent. In one picture of the gong-ringing celebration, Yuan Wenkai, who works for an Alibaba-owned logistics warehouse, stood third from the right. A former tally clerk who graduated from a run-of-the-mill Guangdong vocational school, Yuan holds substantial expertise in automation management. His success with boosting the sorting capacity of a logistics warehouse by twenty thousand orders per hour—responding to elevated demand during the shopping frenzy of Single’s Day (November 11)—merited an invite to the ceremony.89
Even as AI systems exceed human-level performance at tasks ranging from playing Go to translating news articles, human talent will remain crucial for designing and implementing such systems.90 According to one global survey of more than three thousand business executives, landing the “right AI talent” ranked as the top barrier to AI adoption for companies at the frontier of incorporating AI into their products, services, and internal processes.91 But what makes up the “right AI talent”? In its distilled form, GPT diffusion theory suggests that China’s chance of leading the AI revolution rests more on the Yuan Wenkais of the world than the Andrew Ngs. The most important institutional adjustments to the IR-4 are those that widen the pool of AI engineering skills and knowledge.
Indeed, alongside the maturation of the AI field, recent reports have emphasized skills linked to implementing theoretical algorithms in practice and in ways suited for large-scale deployment. In early 2022, the China Academy of Information and Communications Technology (CAICT), an influential research institute under the Ministry of Industry and Information Technology, published two reports that identified AI’s “engineering-ization” (工程化) as a significant trend that involves addressing challenges in transforming AI-based projects from prototypes to large-scale production.92 Relatedly, per Burning Glass job postings from 2010 to 2019, the overall increase in demand for “AI-adjacent” positions in the United States far exceeded that for “core AI” positions.93 Covering skills needed to implement AI throughout many sectors and legacy systems, this pool of AI-adjacent jobs includes positions for systems engineers and software development engineers.
GPT Skill Infrastructure for AI: A US-China Comparison
At present, the United States is better positioned than China to develop the skill infrastructure suitable for AI. First, the United States has more favorable conditions for expanding the number of AI engineers. According to three separate projects that mapped out the global AI talent landscape, many more AI engineers work in the United States than in any other country.94 In 2017, Tencent Research Institute and BOSS Zhipin (a Chinese online job search platform) found that the number of AI “practitioners” (从业者) in the United States far surpassed the corresponding Chinese figure. Figure 7.2 captures this gap across four key AI subdomains: natural language processing (by three times), chips and processors (by fourteen times), machine learning applications (by two times), and computer vision (by three times).95 Overall, the total number of AI practitioners in the United States was two times greater than the corresponding figure for China.96 Furthermore, data from two separate reports by LinkedIn and SCMP Research confirm that the United States leads the world in AI engineers.97
In addition to statistics on the AI workforce, the quantity and quality of AI education is another consideration for which country is better positioned to develop GPT skill infrastructure for AI. Again, the United States leads China by a significant margin in terms of universities with faculty who are proficient in AI. In 2017, the United States was home to nearly half of the world’s 367 universities that provide AI education, operationalized by universities that have at least one faculty member who has published at least one paper in a top AI conference.98 In comparison, China had only 20 universities that met this standard. After replicating this methodology for the years 2020–2021, the US advantage is still pronounced, with 159 universities to China’s 29 universities.99
FIGURE 7.2. A US-China Comparison of AI Practitioners in Key Subdomains. Source: Tencent Research Institute and BOSS Zhipin 2017.
These findings contradict prominent views on the present global distribution of AI engineering talent. In his best-selling book AI Superpowers: China, Silicon Valley, and the New World Order, Kai-Fu Lee argues that the current AI landscape is shifting from an age of discovery, when the country with the highest-quality AI experts wins out, to an age of implementation, when the country with the largest number of sound AI engineers is advantaged.100 In an age of implementation, Lee concludes, “China will soon match or even overtake the United States in developing and deploying artificial intelligence.”101 Pitted against the statistics from the previous passages, Lee’s evidence for China’s lead in AI implementers is meager. His attention is concentrated on anecdotes about the voracious appetite for learning about AI by Chinese entrepreneurs in Beijing.102 While this analysis benefits from Lee’s experience as CEO of Sinovation Ventures, a venture capital fund that invests in many Chinese AI start-ups, it could also be colored by his personal stake in hyping China’s AI capabilities.
Drawing on Lee’s book, Allison and Schmidt also assert that China is cultivating a broader pool of AI talent than the United States today. Specifically, they point out that China graduates three times as many computer science students as the United States on an annual basis.103 Yet the study on which this figure is based finds that computer science graduates in the United States have much higher levels of computer science skills than their Chinese peers. In fact, the average fourth-year computer science undergraduate in the United States scores higher than seniors from the top programs in China.104 Therefore, estimates of China’s pool of AI engineering talent will be misleading if they do not establish some baseline level of education quality. This is another reason to favor the indicators that support an enduring US advantage in AI engineers.105
Second, as previous industrial revolutions have demonstrated, strong linkages between entrepreneurs and scientists that systematize the engineering knowledge related to a GPT are essential to GPT skill infrastructure. In the AI domain, an initial evaluation suggests that this connective tissue is especially strong in the United States. Based on 2015–2019 data, it led the world with the highest number of academic-corporate hybrid AI publications—defined as those coauthored by at least one researcher from both industry and academia—more than doubling the number of such publications from China.106 Xinhua News Agency, China’s most influential media organization, has pinpointed the lack of technical exchanges between academia and industry as one of five main shortcomings in China’s AI talent ecosystem.107
These preliminary indicators align with assessments of the overall state of industry-academia exchanges in China. Barriers to stronger industry-academia linkages include low mobility between institutions, aimless government-sponsored research collaborations, and misguided evaluation incentives for academic researchers.108 One indicator of this shortcoming is the share of R&D outsourced by Chinese firms to domestic research institutes: this figure declined from 2.4 percent in 2010 to 1.9 percent in 2020. Over the same time period, the share of Chinese firms’ R&D expenditures performed by domestic higher education institutions also decreased from 1.2 percent to 0.4 percent.109
Moreover, the US approach to AI standard-setting could prove more optimal for coordinating information flows between labs working on fundamental AI advances and specific application sectors. Market-mediated, decentralized standardization systems are particularly suited for advancing technological domains characterized by significant uncertainty about future trajectories, which clearly applies to AI.110 In such fields, governments confront a “blind giant’s quandary” when attempting to influence technological development through standards-setting.111 The period when government involvement can exert the most influence over the trajectory of an emerging technology coincides with the time when the government possesses the least technical knowledge about the technology. Government intervention therefore could lock in inferior AI standards compared with market-driven standardization efforts.
In that light, China’s state-led approach to technical standards development could hinder the sustainable penetration of AI throughout its economy. For example, the Chinese central government plays a dominant role in China’s AI Industry Alliance, which has pushed to wrest leadership of standards setting in some AI applications away from industry-led standardization efforts.112 Excessive government intervention has been a long-standing weakness of China’s standardization system, producing standards not attuned to market demands and bureaucratic rivalries that undermine the convergence of standards.113 Wang Ping, a leading authority on this topic, has argued that China needs to reform its standardization system to allow private standards development organizations more space to operate, like the Institute of Electrical and Electronics Engineers in the United States and the European Committee for Electrotechnical Standardization.114
In sum, the United States is better positioned than China to not only broaden its pool of AI engineering skills but also benefit from academia-industry linkages in AI engineering. In previous industrial revolutions, these types of institutional adaptations proved crucial to technological leadership. Still, much uncertainty remains in forecasts about GPT skill infrastructure for AI, especially with regard to determining the right measures for the right AI talent. Recent studies of market demand for AI- and ICT-related jobs suggest that employers are softening their demands for a four-year degree in computer science as a requirement for such positions.115 Certificate programs in data science and machine learning that operate under the bachelor-degree level could play an important role in expanding the pool of AI engineering talent.116 Taking into account these caveats, this section’s evaluation of GPT skill infrastructure at the very least calls into question sweeping claims that China is best placed to capitalize on the IR-4.
Reframing National AI Strategies
The preceding conclusions offer a marked contrast with how American and Chinese policymakers devise national AI strategies. Policy proposals for pursuing US leadership in AI consistently call for more AI R&D as the highest priority. For example, the report “Meeting the China Challenge: A New American Strategy for Technology Competition,” published in 2020 by a working group of twenty-eight China specialists and experts, provided sixteen policy recommendations for how the United States should ensure its leadership in AI and three other key technological domains. The very first recommendation was for the United States to significantly expand investment in basic research, raising total R&D funding to at least 3 percent of GDP.117 The Trump administration’s “American AI Initiative,” launched to maintain US leadership in AI “in a time of global power competition,” also listed AI R&D spending as its very first policy recommendation.118
The Chinese government also prioritizes investments in R&D, sometimes at the expense of other routes to productivity growth oriented around technology adoption and education.119 China’s five-year plan (2021–2025) aims to raise basic research spending by over 10 percent in 2021, targeting AI and six other key technological domains.120 China consistently sets and meets ambitious targets for R&D spending, but that same commitment has not extended to education funding. While China’s R&D spending as a percentage of GDP in 2018 was higher than that of Brazil, Malaysia, Mexico, or South Africa (other middle-income countries that industrialized on a similar timeline), China’s public expenditure on education as a percentage of GDP was lower than the figure in those countries.121 According to a group of experts on China’s science and technology policy, one possible explanation for this disparity between attention to R&D versus education is the longer time required for efforts in the latter to yield tangible progress in technological development.122
As both the United States and China transition from initiating a new GPT trajectory to diffusing one across varied application sectors, investing in the broader AI-adjacent skill base will become more crucial than cornering the best and the brightest AI experts. Policies directed at widening the AI talent base, such as enhancing the role of community colleges in developing the AI workforce, deserve more attention.123 Applied technology centers, dedicated field services, and other technology diffusion institutions can incentivize and aid adoption of AI techniques by small and medium-sized enterprises.124 Reorienting engineering education toward maintaining and overseeing AI systems, not solely inventing new ones, also fits this frame.125
A strategy oriented around GPT diffusion does not necessarily exclude support for the exciting research progress in a country’s leading labs and universities. R&D spending undoubtedly will not just help cultivate novel AI breakthroughs but also contribute to widening the GPT skill infrastructure in AI. All too often, however, boosting R&D spending seems to be the boilerplate recommendation for any strategic technology.126 GPTs like AI are not like other technologies, and they demand a different toolkit of strategies.
Alternative Factors
In exploring how the IR-4 could bring about an economic power transition, it is important to compare the implications derived from the GPT diffusion mechanism to those that follow from explanations stressing other factors. Consistent with the previous chapters, I first consider threat-based explanations and the varieties of capitalism (VoC) approach. I then address how US-China competition over emerging technologies could be shaped by differences in regime type, a factor that is particularly salient in this case.
Threat-Based Explanations
One potentially dangerous implication of threat-based explanations is that war, or manufacturing the threat of war, is necessary for economic leadership in the IR-4. Crediting the US military’s key role in spurring advances in GPTs during the twentieth century, Ruttan doubts that the United States could initiate the development of GPTs “in the absence of at least a threat of major war.”127 Extending Ruttan’s line of thinking to the US strategic context in 2014, Linda Weiss expressed concerns that the end of the Cold War, along with the lack of an existential threat, removed the impetus for continued scientific and technological innovation. Specifically, she questioned “why China has not yet metamorphosed into a rival that spurs innovation like the Soviet Union and Japan.”128 Weiss only needed a little more patience. A few years later, the narrative of a US-China “Tech Cold War” gained momentum as both sides of the bilateral relationship trumped up threats to push national scientific and technological priorities.129
GPT diffusion theory strongly refutes the notion that manufacturing external threats is necessary for the United States or China to prevail in the IR-4. An external menace did not drive the rise of the United States in the IR-2. Across all cases, military actors were involved in but not indispensable to spurring the development of GPTs, as many civilian entities also fulfilled the purported role of military investment in providing a large initial demand for incubating GPTs. Furthermore, threat-based interpretations extend only to the moment when one country stimulates the first breakthroughs in a GPT. Even if stoking fears can galvanize support for grand moonshot projects, these do not determine which country is able to benefit most from the widespread adoption of advances in GPTs like AI. That hinges on the more low-key toil of broadening the engineering skill base and advancing interoperability standards in GPTs—not fearmongering.
VoC Explanations
Applying the VoC framework to US-China competition in AI gives more ambiguous results. The VoC approach states that liberal market economies (LME)—prototypically represented by the United States—are more conducive to radical innovation than coordinated market economies (CME).130 It is unclear, however, whether China fits into the VoC framework as a CME or LME. While some label China a CME, others characterize it as an LME.131 This disputed status speaks to the substantial hybridity of China’s economy.132 China has been treated as a “white space on the map” of VoC scholarship, which was originally developed to classify different forms of advanced capitalist economies.133 This makes it difficult to derive strong conclusions from VoC scholarship about China’s ability to adapt to the IR-4’s radical innovations.
The same holds if we focus on the skill formation aspect of the VoC framework. China’s education system emphasizes training for general skills over vocational skills.134 In this respect, it is similar to LMEs like the United States, which means VoC theory provides limited leverage for explaining how the IR-4 could differentially advantage the United States or China. On this topic, GPT diffusion theory points to differences in AI engineering education as more significant than distinctions based on the CME-LME models.
Case-Specific Factors: Regime Type
What is the effect of regime type on technological leadership in the IR-4? The distinction between authoritarian China and the democratic United States takes center stage in arguments about the future of great power rivalry.135 Regime type could also influence the specific aspect of great power competition that GPT diffusion tackles—whether one great power is able to sustain productivity growth at greater rates than its rivals by taking advantage of emerging technologies. Some evidence suggests that, owing to investments in inclusive economic institutions, democracies produce more favorable conditions for growth than autocracies.136 Additionally, empirical work shows that democracies outgrow autocracies in the long run because they are more open to absorbing and diffusing new techniques.137 Other studies find, more specifically, that internet technologies diffuse faster in democracies, possibly because nondemocracies are threatened by the internet’s potential to empower antigovernment movements.138
On the other hand, the impact of democracy on technological progress and economic growth is disputed. Drawing on data from fifty countries over the 1970–2010 period, Taylor finds that regime type does not have a strong relationship with national innovation rates, as measured by patenting rates.139 One review of the econometric evidence on democracy and growth concludes that “the net effect of democracy on growth performance cross-nationally over the last five decades is negative or null.”140 Moreover, China’s rapid economic growth and adoption of internet technologies stands out as an exception to general claims about a democratic advantage when it comes to leveraging new technologies as sources of productivity growth. Contrary to initial expectations, incentives to control online spaces have made some autocratic regimes like China more inclined to spread internet technologies.141 Other scholars point out that the stability of China’s authoritarian regime has encouraged substantial contributions to R&D and technical education, the type of investments in sustained productivity growth typically associated with democracies.142
It is not within the scope of this chapter to settle these debates.143 Still, the juxtaposition of the GPT and LS mechanisms does speak to how regime type could shape US-China competition in the IR-4. Though the conventional wisdom links democracy to freewheeling thought and capacity for innovation, the most important effects of regime type in US-China technological competition during the IR-4, under the GPT diffusion framework, may materialize through changes in GPT skill infrastructure. Democracies tend to be more politically decentralized than autocracies, and decentralized states could be more responsive to the new demands for engineering skills and knowledge in a particular GPT. This accords with evidence that new technologies consistently diffuse more quickly in decentralized states.144
Summary
How far can we take GPT diffusion theory’s implications for the US-China case? I have presented support for the GPT mechanism across a range of historical case studies, each of which covers at least four decades and two countries. At the same time, it is necessary to acknowledge limitations in translating lessons from past industrial revolutions and power transitions to the present.
To begin, it is important to clarify that my findings directly address the mechanism by which technological breakthroughs enable China to surpass the United States in economic productivity.145 The scenario in which China overtakes the United States as the most powerful economy is different from one in which the US-China power gap narrows but does not fully disappear. Scholars have rightly noted that the latter scenario—“posing problems without catching up,” in the words of Thomas Christensen—still significantly bears on issues such as Taiwan’s sovereignty.146 Even so, the possibility of China fully closing the power gap with the United States is especially crucial to study. When rising and established powers are close to parity, according to power transition theory, the risk of hegemonic war is the greatest.147 China’s ability to sustain economic growth also affects its willingness and ability to exert influence in the international arena.
Next, GPT diffusion theory speaks to only one pathway to productivity leadership. The historical case studies have demonstrated that institutional responses to disruptive technological breakthroughs play a key part in economic power transitions. However, China’s prospects for long-term economic growth could also hinge on demographic and geographic drivers.148
A number of factors could affect whether lessons from the past three industrial revolutions extend to the implications of present-day technological advances for a US-China power transition. The most plausible transferability issues can be grouped into those that relate to the nature of great power competition and those that relate to the nature of technological change.
First, as put forward by Stephen Brooks and William Wohlforth, China’s rise in the twenty-first century could face structural barriers that did not exist in previous eras.149 Relying in part on data from 2005–2006, they argue that the present gap between the United States and China in military capabilities, as captured in long-term investments in military R&D, is much larger than past gaps between rising powers and established powers.150 Arguably, the US-China gap in military expenditures has narrowed to the extent that comparisons to historical distributions of military capabilities are more viable. In 2021, China accounted for 14 percent of global military expenditures. This updated figure, albeit still much lower than the US share (38 percent), reflects China’s military modernization efforts over the past two decades during a time of declining US military spending.151 This ratio is more comparable to distributions of military capabilities in the historical periods analyzed in the case studies.152
Another structural barrier China faces is that the growing complexity of developing and deploying advanced military systems now makes it more difficult for rising powers to convert economic capacity into military capacity than it was in the past.153 There are a few reasons why it is still relevant to study China’s ability to convert the technological breakthroughs of the IR-4 into sustained productivity growth. To start, rising states could still benefit from the steady diffusion of some complex military technologies connected to advances in the commercial domain, such as armed uninhabited vehicles.154 In addition, military effectiveness does not solely derive from extremely complex systems like the F-22 stealth fighter. Converting production capacity to military strength could be more relevant for China’s investments in asymmetric capabilities and those suited for specific regional conflicts, such as ground-based air defense systems and the rapid replacement of naval forces.155 Lastly, there remains a strong connection between economic development and countries’ capabilities to “produce, maintain, and coordinate complex military systems.”156
As for the second set of transferability issues, the technological landscape itself is changing. Accelerated globalization of scientific and technological activities may reduce the likelihood of adoption gaps between advanced economies when it comes to emerging technologies.157 Despite these considerations, there are also compelling reasons to think that the nature of technological change in this current period only magnifies the importance of GPT diffusion theory. Cross-country studies indicate that while new technologies are spreading between countries faster than ever, they are spreading to all firms within a country at increasingly slower rates. Networks of multinational firms at the global technology frontiers have reduced cross-national lags in the initial adoption of new technologies, but the cross-national lags in the “intensive adoption” of new technologies, as measured by the time between the technologies’ initial adoption to intensive penetration throughout a country, has only grown.158 These trends give more weight to the GPT mechanism.
Finally, even if the rise and fall of great technologies and powers is fundamentally different in the twenty-first century, previous industrial revolutions still exert substantial influence in the minds of academics and policymakers.159 To justify and sustain their agendas, influential figures in both the United States and China still draw on these historical episodes. At the very least, this chapter submits different lessons to be learned from these guiding precedents.
When some of the leading thinkers of our era declare that the AI revolution will be more significant than the industrial revolution, it is difficult to not get caught up in their excitement. Somehow, every generation winds up believing that their lives coincide with a uniquely important period in history. But our present moment might not be so unprecedented. Unpacking how AI could influence a possible US-China power transition in the twenty-first century requires first learning the lessons of GPT diffusion from past industrial revolutions.
8 Conclusion
STUDIES OF HOW TECHNOLOGY interacts with the international landscape often fixate on the most dramatic aspect of technological change—the eureka moment. Consistent with this frame, the standard explanation for the technological causes of economic power transitions emphasizes a rising power’s ability to dominate profits in leading sectors by generating the first implementation of radical inventions. This book draws attention, in contrast, to the often unassuming process by which an innovation spreads throughout an economy. The rate and scope of diffusion is particularly relevant for GPTs—fundamental advances like electricity or AI that have the potential to drive pervasive transformation across many economic sectors.
Based on the process of GPT diffusion, this book puts forward an alternative theory of how and when significant technological breakthroughs generate differential rates of economic growth among great powers. When evaluating how technological revolutions affect economic power transitions, GPTs stand out as historical engines of growth that can provide major boosts to national productivity. Though each is different, GPTs tend to follow a common pattern: after multiple decades of complementary innovations and institutional adaptations, they gradually diffuse across a broad range of industries. Everything, everywhere, but not all at once.
This impact pathway markedly diverges from existing theories based on leading sectors. Akin to a sprint on a narrow lane, great power competition over leading sectors is framed as a race to dominate initial breakthroughs in the early growth periods of new industries. In contrast, GPT diffusion theory proposes that by more effectively adopting GPTs across many application sectors, some great powers can sustain higher levels of productivity growth than their competitors. Like a marathon on a wide road, great power competition over GPTs is a test of endurance.
Disruptive technological advances can bring about economic power transitions because some countries are more successful at GPT diffusion than others. A nation’s effectiveness at adapting to emerging technologies is determined by the fit between its institutions and the demands of those technologies. Thus, if economic power transitions are driven by the GPT trajectory, as opposed to LS product cycles, the institutional adaptations that matter most are those that facilitate information exchanges between the GPT sector and application sectors, in particular the ability of nations to widen the engineering skill base linked to a new GPT.
Three historical case studies, designed and conducted in a way to assess the explanatory power of the GPT mechanism against the LS mechanism, provide support for GPT diffusion theory. The case studies cover periods characterized by both remarkable technological change—the “three great industrial revolutions” in the eyes of some scholars—and major fluctuations in the global balance of economic power.1 Overall, the case study evidence underscores the significance of GPT diffusion as the key pathway by which the technological changes associated with each industrial revolution translated into differential rates of economic growth among the great powers.
In the case of Britain’s rise to economic preeminence during the First Industrial Revolution, expanded uses of iron in machine-making spurred mechanization, the key GPT trajectory. The gradual progression of mechanization aligned with the period when Britain’s productivity growth outpaced that of France and the Netherlands. Britain’s proficiency in adopting iron machinery across a wide range of economic activities, rather than export advantages from dominating innovation in leading sectors such as cotton textiles, proved more central to its industrial ascent. Though its industrial rivals boasted superior systems of higher technical education for training expert scientists and top-flight engineers, Britain benefited from mechanics’ institutes, educational centers like the Manchester College of Arts and Sciences, and other associations that expanded access to technical literacy and applied mechanics knowledge.
The Second Industrial Revolution case also favors the GPT mechanism’s explanation of why certain great powers better adapt to periods of remarkable technological change. The LS mechanism focuses on Germany’s discoveries in new science-based industries, such as chemicals, as the driving force behind its catching up to Britain before World War I. However, the United States, emerging as the preeminent economic power during this period, was more successful than Germany in exploiting the technological opportunities of the Second Industrial Revolution. Enabled by innovations in machine tools, the extension of interchangeable manufacturing techniques across many American industries functioned as the key GPT trajectory that fueled the rise of the United States. Scientific infrastructure or industrial R&D capabilities, areas in which the United States lagged behind its industrial rivals, cannot account for its advantage in adopting special-purpose machinery across nearly all branches of industry. Rather, the United States gained from institutional adaptations to widen the base of mechanical engineering talent, including through the expansion of technical higher education schools and the professionalization of mechanical engineering.
Evidence from the US-Japan rivalry amid the information revolution exposes more gaps in the LS account. During the late twentieth century Japan captured global market shares in new fast-growing sectors such as consumer electronics and semiconductor components, prompting many to predict that it would overtake the United States as the leading economic power. Yet such an economic power transition, an inevitability based on the expectations of the LS mechanism, never occurred. Instead, the United States sustained higher rates of economic growth than Japan owing, in part, to greater spread of computerization across many economic sectors. Japan’s productivity growth kept up with the US rate in sectors that produced information technology but lagged far behind in sectors that intensively used information technology. Once again, differences in institutional adaptations to widen the GPT skill base turned out to be significant. While Japanese universities were very slow to adapt their training to the demand for more software engineers, US institutions effectively broadened the pool of such skills by cultivating a separate discipline of computer science.
As a supplement to the case studies, I conducted a large-n statistical analysis to test whether countries with superior GPT skill infrastructure preside over higher rates of GPT diffusion. Leveraging time-series cross-sectional data on software engineering education and computerization rates in nineteen countries (the G20 economies) across twenty-five years, the quantitative analysis confirmed this crucial expectation derived from GPT diffusion theory. I found less support for other factors often assumed to have a positive effect on an economy’s broader technological transformation, including institutional factors linked to securing LS product cycles. This empirical test validates a core component of GPT diffusion theory across a sample of the world’s major emerging and developed economies.
Main Contributions
First, at its core, Technology and the Rise of Great Powers introduces and defends GPT diffusion theory as a novel explanation for how and when technological change can lead to a power transition. Historical case studies and statistical analysis substantiate the explanatory power of GPT diffusion theory over the standard explanation of technology-driven power transitions based on leading sectors, which exerts enduring influence in policy and academic circles.2 In doing so, the book answers the call by scholars such as Michael Beckley and Matthew Kroenig for the international relations field to devote more attention to the causes of power transitions, not just their consequences.3
By expounding on the significance of GPT skill infrastructure, the book points toward next steps to better understanding the politics behind some of the most significant technological advances in human history. To investigate why some countries are more successful in cultivating GPT skill infrastructure, promising avenues of research could tap into existing work that concentrates on centralization, inclusiveness of political institutions, government capacity to adopt long time horizons, and industrial organization.4 In these efforts, being careful to differentiate between various pathways by which technological changes make their mark, as this book does with the GPT and LS mechanisms, will be especially important when underlying political factors that satisfy the demands of one type of technological trajectory run counter to the demands of another.
Future work should also probe other institutional factors beyond GPT skill infrastructure that contribute to cross-national differences in GPT adoption. This opens up a universe of institutions that are often ignored in innovation-centered accounts of technological leadership, including gender gaps in engineering education,5 transnational ethnic networks that facilitate technology transfer,6 and “technology diffusion institutions,” such as standard-setting organizations and applied technology centers.7
In positioning the LS mechanism as the main foil to the GPT mechanism, my intention is to use this clash between theories to productively advance our understanding of the rise and fall of great technologies and powers. Contestation should not be misread as disparagement. In one sense, GPT diffusion theory builds on previous scholarship about leading sectors, which first identified the need to flesh out more specific linkages between certain technological advances and more highly aggregated economic changes in the context of power transitions.8 Testing, revising, and improving upon established theories is essential to gradual yet impactful scientific progress—not so unlike the incremental, protracted advance of a GPT.
Second, the book’s central argument also suggests revisions to assessments of power in international politics. Recognizing that scientific and technological capabilities are becoming increasingly central to a nation’s overall power, researchers tend to equate technological leadership with a country’s ability to initiate “key ‘leading sectors’ that are most likely to dominate the world economy into the twenty-first century.”9 For instance, an influential RAND report, “Measuring National Power in the Postindustrial Age,” proposes a template for measuring national power based on a country’s capacity to dominate innovation cycles in “leading sectors.”10 In this effort, the authors draw directly from LS-based scholarship: “The conceptual underpinnings of this template are inspired by the work of Schumpeter, Rostow, Gilpin, Kennedy, and Modelski and Thompson.”11 This study has gained considerable traction in academic and policymaking circles, inspired further workshops on national power, and has been called “the definitive US study on CNP [Comprehensive National Power].”12
Contrary to these approaches, this book submits that evaluations of scientific and technological power should take diffusion seriously. Assessments that solely rely on indicators of innovation capacity in leading sectors will be misleading, especially if a state lags behind in its ability to spread and embed innovations across productive processes. A more balanced judgement of a state’s potential for technological leadership requires looking beyond multinational corporations, innovation clusters like Silicon Valley, and eye-popping R&D numbers to the humble undertaking of diffusion. It shines the spotlight on a different cast of characters: medium-sized firms in small towns, engineers who tweak and implement new methods, and channels that connect the technological frontier with the rest of the economy.
In an article published in the Review of International Political Economy journal, I illustrated the value of this diffusion-oriented approach in gauging China’s scientific and technological capabilities.13 Preoccupied with China’s growing strength in developing new-to-the-world advances, existing scholarship warns that China is poised to overtake the United States in technological leadership. This is mistaken. There is still a large gap between the United States and China when it comes to the countries’ readiness to effectively spread and utilize cutting-edge technologies, as measured by penetration rates of digital technologies such as cloud computing, smart sensors, and key industrial software. When the focus shifts away from impressive and flashy R&D achievements and highly cited publications, China’s “diffusion deficit” comes to light. Indeed, a diffusion-centric assessment indicates that China is much less likely to become a scientific and technological superpower than innovation-centric assessments predict.
Relatedly, the GPT diffusion framework can be fruitfully applied to debates about the effects of emerging technologies on military power. Major theories of military innovation focus on relatively narrow technological developments, such as aircraft carriers, but the most consequential military implications of technological change might come from more fundamental advances like GPTs. In an article that employs evidence from electricity’s impact on military effectiveness to analyze how AI could affect the future of warfare, Allan Dafoe and I challenge studies that predict AI will rapidly spread to militaries around the world and narrow gaps in capabilities.14
Third, as chapter 7 spells out in detail, GPT diffusion theory provides an alternative model for how revolutionary technologies, in particular AI, could affect the US-China power balance. This, in turn, implies different optimal policies for securing technological advantage. Drawing on the LS template, influential thinkers and policymakers in both the United States and China place undue emphasis on three points: the rapid timeframe of economic payoffs from AI and other emerging technologies; where the initial, fundamental innovations in such technologies cluster; and growth driven by a narrow range of economic sectors.
GPT diffusion theory suggests diverging conclusions on all three dimensions. The key technological trajectory is the relative success of the United States and China in adopting AI advances across many industries in a gradual process that will play out over multiple decades. It will be infeasible for one side to cut the other off from foundational innovations in GPTs. The most important institutional factors, therefore, are not R&D infrastructure or training grounds for elite AI scientists but rather those factors that widen the skill base in AI and enmesh AI engineers in cross-cutting networks with entrepreneurs and scientists.15
Yet, the United States is fixated on dominating innovation cycles in leading sectors. When it comes to their grand AI strategy, US policymakers are engrossed in ensuring that leading-edge innovations do not leak to China, whether by restricting the immigration of Chinese graduate students in advanced technical fields or by imposing export controls on high-end chips for training large models like GPT-3 and ChatGPT.16 A strategy informed by GPT diffusion theory would, instead, prioritize improving and sustaining the rate at which AI becomes embedded in a wide range of productive processes. For instance, in their analysis of almost 900,000 associate’s degree programs, Center for Security and Emerging Technology researchers Diana Gehlhaus and Luke Koslosky identified investment in community and technical colleges as a way to unlock “latent potential” in the US AI talent pipeline.17 This recommendation accords with an OECD working paper on the beneficial effects of a wider ICT skills pool on digital adoption rates across twenty-five European countries. The study finds that “the marginal benefit of training for adoption is found to be twice as large for low-skilled than for high-skilled workers, suggesting that measures that encourage the training of low-skilled workers are likely to entail a double dividend for productivity and inclusiveness.”18
At the broadest level, this book demonstrates a method to unpack the causal effects of technological change on international politics. International relations scholars persistently appeal for the discipline to better anticipate the consequences of scientific and technological change, yet these demands remain unmet. By one measure, between 1990 and 2007, only 0.7 percent of the twenty-one thousand articles published in major international relations journals explicitly dealt with the topic of science and technology.19 One bottleneck to researching this topic, which Harold Sprout articulated back in 1963, is that most theories either grossly underestimate the implications of technological advances or assume that technological advance is the “master variable” of international politics.20
This book shows that the middle ground can be a place for fruitful inquiry. Technology does not determine the rise and fall of great powers, but some technological trends, like the diffusion of GPTs, do seem to possess momentum of their own. Social and political factors, as represented by GPT skill infrastructure, shape the pace and direction of these technological trajectories. This approach is particularly useful for understanding the effects of technological change across larger scales of time and space.21
《伯尔尼公约》将署名表述为“表明作者身份的权利”（The right to identify as author, the right to claim authorship），此后多数国家在本国著作权法中将署名行为与表明作者身份行为画上了等号，但也有少数国家将署名与作者身份权分设，将署名作为表明作者身份的下属概念。因此，署名行为体现的究竟是作者身份还是创作行为本身就值得商榷。从历史沿革来看，署名最初表明的仅是创作行为，是无需意思表示的事实行为，后随着人权意识的日益发展而最终被冠以权利之名，署名权中的人格属性是在权利化过程中被后来赋予的。在法国18世纪末《表演权法》之前，署名行为的主体在世界范围内并不以具备人格精神为前提。本文认为，在目前对如何将人工智能纳入“以人为中心”的法律体系的讨论甚嚣尘上之际，对其署名的讨论可以回归到署名权利化之前，以署名行为为中心进行评判。参与创作过程的人工智能可以基于创作事实进行署名标注，这是基于未来作品流通的市场秩序考量，更是诚实信用原则的体现。在人工智能署名行为的具体展开上，应当充分考量人工智能的风险属性，其署名应受到严格限制。自然人创作作品后，仅有以何种方式署名或不署名的权利，署名权不可转让、不可放弃。对于人工智能而言，由于人工智能本身应受监督与管理，其对署名自主选择的空间应当更为狭窄。署名的目的在于避免混淆，而人工智能生成服务的标注方式又相对固定。因此，人工智能的署名不仅是不可选择、不可放弃的，而且应当是强制的，且署名人或单位要对署名的真实性与可视化承担责任。这种强制标识义务与知识产权中的商标权较为相似。尽管商标权包括利用与排他权能，但根据《商标法》第6条规定，“法律、行政法规规定必须使用注册商标的商品，必须申请商标注册，未经核准注册的，不得在市场销售”。与此相对应，我国《烟草专卖法》第19条规定“卷烟、雪茄烟和有包装的烟丝必须申请商标注册，未经核准注册的，不得生产、销售”。事实上，尽管法律提出了如果不实际使用商标有可能面临商标被撤销的风险，但是原则上，法律对商标权人是否在商品或者服务上使用注册商标并未提出强制性要求。在国家严格管理的领域中，商标权人的商标利用权能受到限制，必须在该类商品或服务中使用注册商标以建立标识、形成品牌、避免混淆。人工智能领域同样如此，不同于普通商品或服务，人工智能的技术、设备、系统和应用具有多样性、复杂性和不透明性，其对社会、经济和个人隐私会产生潜在影响，故人工智能领域的标注行为同样应当被强制。这既是人工智能的特点所决定的，也是构建技术信任与科技伦理的前提。
其次，企业数据之上承载个人信息利益和财产利益，需要接受个人信息保护和竞争法的法律规制。海量的用户数据经过企业等数据主体的收集与汇聚即形成规模化的企业数据。由于此类数据集合可以反映出市场客观规律，预测未来趋势，故其构成人工智能训练数据的重要来源。从人工智能训练数据的实例来看，OpenAI在训练其人工智能产品ChatGPT时，就将Raw Story Media和Alter Net Media等新闻机构的一系列新闻稿件作为人工智能训的练数据来源，并因相关数据使用行为未经机构授权许可而面临著作权侵权纠纷。企业数据承载着包括个人信息权益、财产利益等在内的多元利益形态。一方面，企业数据来源于不同的用户数据，在一定程度上可以视为对个人信息的集合。如果人工智能训练数据具备直接或间接可识别性，可被识别定位为特定自然人主体，则该数据集合之上依然承载着用户的个人信息利益。此时，人工智能训练数据需要通过个人信息保护的法律规制，以消除数据集合中潜在的对个人信息权益的侵权风险。另一方面，企业数据产生方式凝结了数据主体的劳动成果及其利益诉求。企业数据通常是企业等数据主体收集、分析、加工数据后所获得的数据集合，凝结着企业等数据主体财力、物力与人力等劳动投入，由此产生了值得产权制度保护的财产利益。目前，不同客体形态下企业数据的财产利益已经获得司法的保护与认可。在谷米诉元米案、淘宝诉美景等案中，法院即认为企业开发的数据集合能够为权利人带来现实或潜在的经济利益，具备无形财产属性，企业应当对该数据集合享有独立的财产性权益。尤其在企业数据的作品属性受到广泛质疑且企业数据财产权立法缺位的现状下，更多法院选择以《反不正当竞争法》一般条款作为规制范式，强化对企业数据中财产利益的保护。
