Skip the Frontier

Countries outside the AI production chain cannot catch up by replicating it; they can only win by owning the deployment surface of the next substrate before it reaches scale somewhere else.

Half of the 2025 Nobel Prize in Economics was awarded to Joel Mokyr “for having identified the prerequisites for sustained growth through technological progress.” His framework rests on three components: useful knowledge (propositional knowledge of why something works and prescriptive knowledge of how to do it), mechanical competence (workforce that can apply that knowledge), and institutions conducive to innovation. The internet and open-weight models have partially democratized the first — propositional and prescriptive knowledge are now closer to a digital Republic of Letters than at any prior moment, an argument Mokyr himself has made. The other two have not democratized. Mechanical competence remains concentrated where industrial training pipelines exist, and institutional change takes decades . Countries outside the supply chain for scientific and industrial progress have been playing catch-up by trying to replicate other nations’ successes without significant headway. The advent of economically viable artificial intelligence threatens to widen the gap, and the stakes are no longer purely economic. AI’s growing capacity to act on behalf of nation-states in cyber operations and autonomous weapons means institutional lag now compounds into military and sovereignty risk. Mokyr put the concern directly : “If institutions don’t improve while technology improves, you’re giving more and more power to people who are likely to abuse and misuse it.”

The same historical period that anchors Mokyr’s framework was also the subject of Alexander Gerschenkron’s Economic Backwardness in Historical Perspective , written over a half-century earlier. Gerschenkron asked what late entrants do when they cannot replicate the leader’s path. His answer has held up across more than sixty years of empirical extension: institutional substitution. Britain industrialized through gradual private accumulation; Germany , lacking a comparable capital market, industrialized through universal banks like Deutsche Bank and Dresdner Bank that acted as long-horizon industrial financiers; Russia, lacking even those, industrialized through the state itself, with the Witte ministry buying two-thirds of all domestic metallurgical production by 1899. Each more-backward case required a more concentrated institutional substitute to compensate for missing prerequisites.

Crucially, the substitution wasn’t deployed against the contemporary leader’s existing frontier — it was deployed against the next one. In 1880, Britain held formidable control over textiles, iron, and steam, and the European powers that tried to compete on those terrains remained junior partners. Germany and the United States bet instead on chemicals and electricity , building the institutional capacity — Germany’s research universities and Technische Hochschulen, America’s corporate R&D labs and patent-protected industrial trusts — that those new substrates required. The bet paid off: by 1914, German firms produced 90% of the world’s synthetic dyes, and roughly half of the world’s electrical equipment; by 1913, U.S. industrial output already exceeded Britain, Germany, and France combined (~⅓ of the world’s industrial output). The lesson: late industrializers should spend on the substrate that has not yet been captured, with institutions sized to the task.

Gerschenkron’s case studies had something his modern analogues do not: baseline state capacity to commission and audit large projects. Germany’s banks could finance a chemical plant because their contracts were enforced ; Witte’s ministry could buy two-thirds of Russian metallurgical output because the procurement chain, however inefficient, delivered . Lant Pritchett’s 2009 diagnosis of the flailing state describes the modern failure mode: a capable head no longer reliably connected to the arms and legs of implementation. The human supply chain downstream of any commissioned project is the layer at which capital leaks and projects stall, and the layer that traditional executive oversight cannot scale to monitor. This is now technically solvable. Industry already runs AI-native logistics control towers for supply chain optimization and computer-vision construction monitoring that tracks worker presence and progress-against-plan. At the government layer, Ukraine’s Prozorro has saved $8.7 billion since 2016 and reduced wartime drone-procurement prices by roughly 30% . None of these systems individually closes the chain of custody between a politician commissioning a bridge and the bridge being built, but each is operational and produces measurable outcomes. With the right technical staff, governments can build out local compute inference harnesses that help the institutional efforts to close capital-allocation and accountability gaps .

Once administrative capacity is in place, the substrate issue becomes tractable. The number of patents , research papers , and capital allocations in Germany prior to the second industrial revolution were signposts for the direction its institutions were taking the country. The equivalent signposts today (research , venture capital , and patents ) point to physical AI, especially humanoid robotics, as the next substrate for industrialization. Nations with large scale production capabilities, such as India, should focus on developing capacity to produce the components of humanoid robotics most likely to cause bottlenecks . Most countries outside of the AI production line may not be able to establish industrial capacity fast enough to contribute to the production of humanoid robots. There are three moves that prevent nations in this position from being sidestepped. The primary framework is the scalable deployment of humanoid robotics.

1. Decentralized Energy Grids. Solar and battery storage along agricultural and industrial corridors can power embodied-AI rural deployment, and work as defense against the kind of grid attacks that have destroyed ~60% of Ukraine’s capacity since 2022.
2. Network and compute sovereignty. Physical fiber backbones and mandatory domestic internet exchanges keep embodied AI from exfiltrating operational data to foreign jurisdictions, or from malicious hackers. Even if robotic deployment comes later than expected, network sovereignty will still be a major issue due to the advances of LLM hacking . Compute sovereignty prevents foreign interference with government work and physical deployment.
3. Reward open weight/source development. The collective strength of the open source community is a force-multiplier that benefits all parties in competition with closed-source incumbents.

The pattern is consistent across Gerschenkron’s cases: the late entrant who succeeds identifies the next substrate before the leader’s institutional substitute calcifies around the current one. That window is still open. It will close.