The Limits of Scale: Why China masters some technologies and struggles in others

China has built the most formidable manufacturing system in modern history. Yet some technologies resist it because the knowledge they require does not scale. In the clean rooms of Veldhoven, a city in the southern Netherlands that most Europeans could not locate on a map, sits a mirror the size of a dinner plate. Its surface appears unremarkable. Under a microscope, however, it reveals a peculiarity that borders on the absurd: across its entire face, the deviation from perfect flatness measures only a few atoms. Enlarged to the size of Germany, its tallest hill would rise no higher than a few centimetres.

This mirror forms the optical heart of the world's most advanced semiconductor lithography machines. Without it, the circuits of modern computing cannot be printed. With it, the entire semiconductor industry runs. The mirror took 30 years to make possible.

9,000 kilometers east, a different kind of precision is taking shape. In the industrial districts of Guangdong, factories manufacturing electric vehicle batteries stretch across cavernous buildings whose production lines fold back on themselves like rivers. Engineers with laptops adjust coating parameters by fractions of a micron. Every doubling of output teaches the factory something new: a refinement here, an adjustment there, each small, their cumulative effect measured in a cost-per-kilowatt-hour that has fallen by more than 90% since 2010.

China now produces roughly a third of all manufactured goods on Earth, more than the United States, Germany, Japan and South Korea combined. Its manufacturing profit margins hover around 4.5%, less than half American levels. This is not a failure of execution but the structural signature of a civilisation that excels at scale: volume without pricing power, dominance without depth. The mirror and battery factory are not simply different products. They are the products of different industrial civilisations, and the distance between them cannot be bridged by money alone.

Some technologies improve by being made more. The recipe is well understood: large markets accelerate learning curves, dense supply chains multiply iteration speed, and state capital funds losses that private investors won't absorb. The photovoltaic effect behind solar panels was discovered by a French teenager in 1839. The chemistry of lithium-ion cells was established in Japanese and American laboratories decades ago. What China supplied was scale and that, in technologies with modular architecture and long production runs, is the innovation.

The critical feature of these technologies is that their progress can be measured. Improvements are visible, incremental and cumulative. A government can decree more batteries, more panels, more vehicles, and observe in real time whether the decree is being honoured. Industrial policy, at its most effective, is the coordination of measurable objectives at speed. China’s political economy was built for this. What it was not built for is knowledge that cannot be stated.

In 1958, the Hungarian-British chemist Michael Polanyi, himself a man displaced from one civilization to another, offered an observation that reorganises much of what we think we understand about technology: we know more than we can tell. The phrase sounds modest but its implications are severe. Much of what makes a technology work resides not in patents or specifications but the accumulated judgement of practitioners: the machinist who hears when a tolerance is drifting, the metallurgist who reads a fracture pattern before the instruments register it, the semiconductor technician who notices a barely perceptible shift in plasma colour that signals a process about to fail. This knowledge is real, precise and indispensable. Yet it cannot be taught by instruction. Instead it must be lived given it is transmitted through apprenticeship, deepened through failure and accumulated across institutions over generations.

Technologies governed by this kind of knowledge do not improve primarily through scale. They improve through time. Each failed experiment deposits information about what does not work. Each investigation of an unexpected failure produces understanding that no simulation could have generated in advance. Over decades, these incremental discoveries sediment into a body of practical knowledge that no individual fully possesses but that an institution, taken as a whole, embodies. These technologies have memory and this cannot be manufactured overnight.

This is where the problem becomes structural rather than incidental. Industrial policy funds laboratories, recruits scientists and sets targets. What it cannot do is specify in advance what will be learned because the most important knowledge in precision industries is precisely the knowledge that reveals itself only through failure. China now spends more than 2% of its GDP on research and development, a level comparable to Germany. The spending is real. The knowledge it is trying to buy is structurally resistant to purchase. The jet engine makes this concrete. A modern turbofan operates at the edge of what matter can survive: inside the combustion chamber, temperatures exceed the melting point of the metals from which the engine is made. Turbine blades endure this by being cast as single crystals, a technique decades in refinement, and threaded with internal cooling channels of extraordinary precision. General Electric's GE90 engine required nearly 20 years of development. Rolls-Royce's Trent family took similar time. Each generation incorporated thousands of improvements discovered through testing cycles whose most productive moments were failures: blades cracking in unexpected locations, vibrations propagating through the engine core in ways dynamic models had not predicted, coatings degrading under conditions the laboratory had not replicated. Engineers remember these failures with the clarity of personal anecdotes. China's CJ-1000 programme is serious and progressing. The bottleneck is the sheer density of accumulated failure, and the learning it produced, embedded in organisations that have been living inside this problem since before China's aerospace industry existed. The lithography mirror tells a version of the same story in a different register. EUV lithography requires light at 13.5 nanometres — shorter than the diameter of a virus, absorbed by almost every material it contacts, including air. The machines operate in near-perfect vacuum, reflecting light from mirrors polished to tolerances measurable in atoms. Producing those mirrors required decades of work at the German optics firm Zeiss, beginning in the early 1990s when most experts believed the technology would never become practical. Prototypes failed repeatedly. Entire optical assemblies were redesigned when minute imperfections distorted the light path in ways that theory had not anticipated. What emerged from that process was not a set of instructions but a body of tacit knowledge about how materials behave at extreme tolerances, how polishing must be adapted when conventional techniques fail, how vibration and thermal fluctuation interact at the nanometre scale. ASML, which assembles the finished machines in Veldhoven, required a parallel 30-year trajectory. Its EUV programme stretched across three decades and consumed billions in prototypes that never reached production. The resulting machines, costing upwards of $200 million each and drawing on hundreds of specialised suppliers across Europe and North America, now underpin the entire semiconductor industry. The knowledge required to build them cannot be reconstructed by reverse-engineering the finished machine. It resides in the engineers who learned, the hard way, how the system behaves. Behind both jet engines and lithography lies a deeper structure that rarely receives its due: the precision ecosystem — a network of small, often family-owned firms producing the components and tools on which advanced manufacturing depends. This network is governed by a regress problem. To build a turbine blade at the necessary accuracy, you need machine tools capable of cuts measured in microns. To build those machine tools, you need grinders more precise still. You cannot manufacture a machine more accurate than the machine that built it. Progress requires an entire ecology of mutually reinforcing precision — bearings enabling spindles enabling grinding machines enabling components enabling the end product. This ecology concentrates in regions whose industrial cultures formed over generations: the Black Forest firms whose lineages trace to 19th-century clockmakers, the Nagano spindle-makers who grew from watchmaking, the Swiss Jura positioning systems transmitted through apprenticeship programmes that predate the Second World War. These firms are not merely suppliers. They are repositories of knowledge so old it has become virtually invisible even to the people who hold it. It is embedded in how engineers describe problems, in the informal standards by which work is judged, and in the craft sense that tells a machinist whether a spindle is running true before any instrument has been consulted. China can acquire German grinding machines, though the most advanced are not available for export, but even acquisition of the best would not transfer the knowledge required to use them at their limits, still less to improve upon them. Entering the precision ecosystem at any single point is insufficient as the whole ecology must be grown. The transition from scale to precision has been done before but it's never been done quickly.

Japan in the immediate postwar decades was associated with cheap imitation in the realm of transistor radios, inexpensive optics. By the 1980s the same country dominated precision machine tools and advanced semiconductor equipment. The transformation required three to four decades of industrial culture maturing gradually, firm by firm. FANUC's founder Seiuemon Inaba spent his career insisting that precision came only through obsessive focus on a narrow range of products, absorbing the lessons of each generation into the next. The Japanese machine tool industry did not become world-class by scaling foreign technology. It became world-class by spending decades pushing against the limits of what its own engineers could produce.

Samsung entered DRAM memory in 1983 to widespread scepticism. The early years were defined by fabrication failures, such as contamination, yield problems and unreliability, whose investigation consumed enormous resources. By the early 1990s Samsung was competitive with the best Japanese manufacturers. The elapsed time was roughly a decade of learning from failure at an intensity few programmes have matched. Morris Chang founded TSMC in 1987 on a thesis that proved exactly right: that semiconductor manufacturing was itself a domain of tacit knowledge separable from chip design, and that a firm focused entirely on manufacturing — building depth, transmitting knowledge generation to generation — could outcompete any integrated rival whose attention was divided. TSMC's current dominance reflects 35 years of that focused accumulation. In each case the transition required a minimum of two generations of engineers: one to accumulate the failures and another to inherit their lessons. The first works at the frontier of the possible, impressive by the standards of a latecomer but still short of the global frontier. The second begins with the tacit knowledge the first spent careers acquiring, and can therefore push further.

China is somewhere in the middle of its first generation in precision aerospace and advanced semiconductors. The AECC's engine testing infrastructure is accumulating failures at scale — compressing into 15 years the experiential learning that GE and Rolls-Royce acquired over half a century. Whether test-cell failures are epistemically equivalent to decades of service experience is genuinely uncertain. Huawei's photonics laboratories are cultivating, within a Chinese institutional environment, the specialist engineering culture whose absence is the actual barrier. These efforts reflect a sophisticated understanding of the problem: that the bottleneck is not financial but institutional, and that institutions require time in ways that factories do not. The question is not whether China will cross the boundary. It is how long that crossing takes, and what the world looks like in the interval. China’s rise does not overturn the technological order. Really, it clarifies it.

For most of the modern era, the distinction between scale technologies and precision technologies was obscured because the same countries dominated both, above all the United States, Germany and Japan. China’s ascent has disaggregated that coincidence. A civilization optimized for scale has mastered scale technologies to an extraordinary degree, and it has met in precision technologies a resistance that capital alone cannot dissolve.

The world that results is structurally eccentric. Different systems excel in different technological regimes, and the tempo of progress in each regime is set less by ambition or investment than the pace at which institutional memory accumulates.