“The victorious army first realizes the conditions for victory, and then seeks to engage in battle.” (Sun Tzu, “The Art of War, Chapter IV, ‘Disposition of the Army”).
Sometimes an article starts with months of research. This one started with an email and my slight, compulsory need to clear my thoughts on ideas that have been rummaging for a couple of years by now.
My boss had forwarded a Bloomberg interview with Kai-Fu Lee discussing the global AI race and the rapidly narrowing gap between American and Chinese AI models. Another friend (and occasional boss) dismissed it succinctly as “Chinese propaganda” and instead pointed to a detailed and well-argued analysis by Richard Windsor, arguing that China’s structural disadvantage in advanced semiconductors and AI compute made the story considerably less impressive. Could it really be that simple?
That triggered something.
I wrote back:
“Arghh … you trigger my compulsion “disorder” …. Been thinking about this topic for years. As a physicist studying in the 80s, we didn’t have much computing infrastructure available and were forced to write very efficient, highly capable code. Today, there is no need to waste time on that in the Western world, as computing resources are abundant. This breeds complacency imo.”
The language was perhaps more appropriate for an email between friends than for an article, but the thought behind it has bothered me for years.
When I studied physics in the 1980s, computing power and memory were scarce. You couldn’t simply throw another thousand processors at a badly formulated problem. If your code was inefficient, you improved the code. If the calculation was too large, you looked for another algorithm. We didn’t particularly enjoy constraints, but they forced us to think carefully about what we were actually trying to compute.
Forty years later, AI has turned that logic almost upside down.
The United States has built an extraordinary AI ecosystem around an abundance of capital, advanced GPUs, hyperscale data centers and some of the world’s best technology companies and researchers (albeit many with European and Asian nationalities). Scaling has worked spectacularly well. Exponentially bigger models, larger model training runs and more compute have repeatedly delivered capabilities that many of us would have considered science fiction only a few years ago.
China has been forced to conduct a different experiment.
Restrictions on access to the most advanced GPUs and semiconductor technologies have created a genuine disadvantage for China. That should not be trivialized. Frontier compute matters. Advanced semiconductors matter. High-bandwidth memory matters. The Western semiconductor and cloud ecosystems remain formidable.
But there is another possibility that is much more interesting than simply extrapolating today’s disadvantage into the future:
What if scarcity changes the direction of innovation itself?
If you cannot compete by assembling ever-larger clusters of the world’s best GPUs and HBMs, you have a very strong incentive to extract more intelligence from the compute you actually have. Model architecture matters more. Quantization matters more. Mixture-of-experts matters maybe even more. Hardware/software co-design matters as well. Open-weight models accelerate experimentation and diffusion. Suddenly, intelligence per GPU, per watt, and per dollar becomes at least as interesting as the absolute intelligence that can be squeezed out of the compute architecture.
That’s where a relatively innocent email exchange got somewhat out of hand. My first attempt to answer the question became far more comprehensive than I intended. In The AI Innovation Machine, I argued that:
Technological leadership cannot sensibly be reduced to counting GPUs or comparing AI investment budgets.
Resources matter enormously, but they are only part of an innovation system. Breakthrough potential also depends on the available human capital, how much of that talent actually participates, the diversity and intensity of competition, the diffusion of knowledge, the ability to select and scale successful ideas, and the constraints under which researchers and companies operate.
The comparison between the United States and China becomes much less obvious when viewed through that broader lens. Maybe not in the past, maybe not today, but ignoring the other breakthrough drivers assumes the naive notion that the future will simply reflect the past.
The United States has an extraordinary advantage in frontier compute, private AI investment, and much of the semiconductor and hyperscale infrastructure underlying today’s AI boom. China, however, has an enormous engineering and scientific talent pool, intense domestic competition, a rapidly developing open-weight ecosystem, massive industrial deployment opportunities, and stronger incentives to squeeze more intelligence from fewer computational resources.
The frontier-model gap is already much smaller than the resource gap would suggest. Stanford’s 2026 AI Index estimated that US private AI investment reached ca. USD 286 billion in 2025, more than 23 times China’s ~ USD 12 billion. Yet by March 2026, its comparison of leading US and Chinese models put the performance gap at only 2.7%. The two numbers aren’t directly comparable, of course, but the contrast is remarkable. It begs the question: Why is the observed model-performance gap nowhere near as large as the measured resource gap? (State subsidies may play a role, of course … still, it should prompt us to think).
That doesn’t prove China will win anything. It simply tells us inputs and outcomes are not the same thing.
My earlier article “The AI Innovation Machine” tried to understand the drivers behind that apparent paradox. This one is deliberately simpler.
LET’S MAKE THE QUESTION SIMPLER.
Frameworks, equations, and increasingly elaborate models of innovation can be dangerous: eventually, you can make a relatively simple question unnecessarily complicated.
There are strong reasons to believe China could become the world’s leading AI power. Compute scarcity may stimulate innovation. Chinese companies are becoming remarkably competitive in frontier models despite inferior access to advanced hardware. Open-weight models could accelerate knowledge diffusion. China has enormous human and industrial resources. And history should make us cautious about assuming that today’s technological hierarchy will still exist ten or twenty years from now. Huawei’s rise from a lower-cost follower of Western telecom-equipment vendors to a global technology leader is a useful reminder of how quickly that hierarchy can change.
Equally powerful reasons support the opposite.
Frontier compute really does matter. Semiconductor manufacturing is extraordinarily difficult. Knowing the physics behind EUV lithography is very different from industrializing a machine containing around 100,000 components that must operate with extraordinary precision and reliability. High-bandwidth memory, advanced packaging, and semiconductor manufacturing equipment remain important bottlenecks. American hyperscalers have enormous resources. Western frontier laboratories concentrate exceptional talent. Any efficiency breakthrough discovered in China can eventually be adopted in America and combined with vastly more available compute.
A fundamental asymmetry in the debate also fascinates me.
The West must continuously renew its technological barriers. China only has to eliminate each critical dependency once before moving to the next.
China may also be playing on a different clock. Western industrial policy is inevitably influenced by elections, changing governments, budgets, and corporate reporting cycles. Chinese semiconductor and technology policy can pursue objectives measured in decades. That persistence can waste enormous amounts of money when the strategy is wrong, but it can be extraordinarily powerful when the objective is right.
That difference produces something we now have enough evidence to examine properly, and it is the one part of this argument I have not seen made elsewhere. When China misses an industrial target, the standard Western reading is that the strategy failed. A decade of evidence from Made in China 2025 suggests a different reading. Targets were missed. The starting position for the next attempt moves on anyway with even more ambitious targets. I call this the ratchet, and I will return to it below at length, because I think it matters as much as, or perhaps more than, any of the individual benchmark scores discussed below.
So rather than decide the answer before asking the question, let’s make the strongest case for both sides.
Ten reasons China could win the AI race.
And then:
Ten reasons it won’t.
I suspect the interesting answer lies somewhere in the collision between the two. But first let us get the following out of the way:
WHAT DOES “WINNING” THE AI RACE ACTUALLY MEAN?
Before going any further, we should probably define what winning means. There is no finish line in AI and no referee who will eventually declare the United States or China the winner.
Nor do I mean simply producing the world’s most intelligent model.
A country could lead the frontier benchmarks while another captures more economic value, deploys AI more broadly across its industries, develops more cost-effective models, builds a more self-sufficient technology ecosystem, or integrates AI more successfully into society. As discussed later, the most intelligent model may not be the most commercially successful if another model is smart enough for most applications at a fraction of the cost. And technological leadership should also include the ability to capture those benefits without creating unacceptable societal, security, or potentially catastrophic risks. A country that develops the most capable AI but fails to translate it into broad socio-economic benefit, or cannot control the risks it creates, has hardly “won” the race.
That leadership has several dimensions: (1) frontier capability, how intelligent and capable the best models become, (2) economics, how cheaply useful intelligence can be produced and deployed, (3) innovation capacity, how rapidly the ecosystem generates, absorbs and improves new ideas, (4) industrial diffusion, how broadly AI improves companies, factories, infrastructure, public services and the wider economy, (5) strategic autonomy, how dependent the ecosystem remains on technologies controlled elsewhere, and (6) societal integration and safety, how successfully AI can be adopted at scale while preserving human control, institutional stability, security and appropriate safeguards against catastrophic outcomes.
Winning, in other words, is not simply about building the most intelligent model. It is about converting AI capability into sustained economic and societal advantage without losing control of the risks along the way.
No country needs to lead every dimension simultaneously. Indeed, that is part of this article’s argument. The United States could remain ahead at the absolute intelligence frontier while China becomes stronger in low-cost deployment, open-weight models, or industrial AI. Conversely, China could close much of the model-performance gap while remaining critically dependent on Western semiconductor technology and therefore still be strategically disadvantaged.
So when we ask whether China could win, we are really asking something broader:
Which innovation system will become better at turning talent, ideas, compute, capital and constraints into useful intelligence, and then turning that intelligence into sustained economic, technological and societal advantage without losing control of the risks?
That is maybe a much more interesting race than simply counting GPUs or comparing benchmark scores.
10 REASONS CHINA COULD WIN THE AI RACE.
This connects to the Breakthrough Model I developed in The AI Innovation Machine. In that framework, breakthrough potential depends not simply on resources (R), but on the interaction between human capital (H), participation (P), competitive diversity (D), knowledge diffusion (K), selection and scaling (S), and the constraints (C) acting on the innovation system.
The ten reasons that follow can be read through exactly that lens. China’s case is essentially that advantages in H, P, and D, increasingly rapid K through open-weight models, industrial S, and, perhaps counterintuitively, the innovation pressure created by C, could partly compensate for its current disadvantage in frontier-AI R. The interesting question is whether those factors multiply strongly enough to change the outcome.
The reader should take note that the emphasis on diffusion rather than invention is not new. Jeffrey Ding’s work on technological revolutions argues that great-power leadership has historically been determined less by who invented first than by which state managed to spread a general-purpose technology across its entire economy. My K and S parameters are an attempt to capture that same mechanism for AI.
1. SCARCITY FORCES EFFICIENCY.
China’s biggest weakness may also be its most powerful incentive to innovate.
The United States can increasingly solve AI problems by adding compute: more GPUs, larger clusters, longer training runs, and enormous amounts of capital. China has considerably less access to the world’s most advanced AI hardware. That is unquestionably a disadvantage, but it also changes what Chinese researchers are incentivized to optimize.
When compute is scarce, intelligence per GPU, per watt, and per dollar matters more.
DeepSeek provides an early indication of what this can look like. Its developers have explicitly focused on techniques such as mixture-of-experts architectures, reduced-precision training, efficient attention mechanisms, and hardware-aware optimization to extract more capability from constrained computational resources.
This does not mean that scarcity is inherently good. Having fewer GPUs may force you to develop better algorithms. Having fewer GPUs may force you to develop better algorithms. Having none prevents you from running the experiment.
But somewhere between abundance and deprivation lies a potentially powerful innovation pressure:
When you cannot afford to solve the problem by making the computer bigger, you are forced to make the idea better. In that sense, China’s compute constraint may become an innovation advantage rather than only a handicap.
2. EXPORT CONTROLS MAY REDIRECT INNOVATION RATHER THAN STOP IT.
Western semiconductor restrictions are designed to slow China’s progress in advanced AI by restricting access to frontier GPUs and the technologies required to manufacture them.
And they work. At least today.
China remains behind in leading-edge semiconductor manufacturing, EUV lithography, high-bandwidth memory, and several critical parts of the semiconductor equipment ecosystem.
But restrictions create a second-order effect that is much harder to predict: they dramatically increase the economic value of eliminating the dependency.
China is consequently investing heavily in domestic semiconductor fabrication, memory, lithography, semiconductor equipment, packaging, and AI accelerators. Its semiconductor equipment localization has risen rapidly, while companies such as SMIC (Semiconductor Manufacturing International Corporation), Huawei, CXMT (ChangXin Memory Technologies), AMEC (Advanced Micro-Fabrication Equipment Inc. China), and Naura are attacking different parts of the technology stack.
There is an important distinction here. China does not need to copy Western technology. Patents, trade secrets, and inaccessible supply chains make that difficult both legally and practically.
It needs to reproduce the capability.
And those are not necessarily the same thing.
The uncomfortable question for Western policymakers, then, is not whether export restrictions hurt China. They clearly do.
It is whether they hurt China faster than they increase China’s incentive to eliminate the dependency altogether.
The West must maintain the technological barrier continuously. China only has to break through each critical barrier once.
3. CHINA HAS AN ENORMOUS HUMAN-CAPITAL FACTORY.
AI innovation ultimately comes from people.
China’s engineering, mathematics, computer science, and scientific education system produces an extraordinary number of technically trained graduates. According to the US National Science Foundation, China awarded around 53,000 science and engineering doctorates in 2022, compared with approximately 45,000 in the United States. More strikingly for AI and technology, China awarded around 30,000 engineering and computer-science doctorates, roughly twice the US total of 15,000.
MacroPolo’s Global AI Talent Tracker provides another indication of the scale: researchers of Chinese origin, measured by undergraduate education, accounted for 47% of the top-tier AI researchers in its 2022 NeurIPS sample. That does not mean China possesses 47% of the world’s best AI researchers. Many of those Chinese-born researchers work in the United States, and America’s extraordinary ability to attract global talent remains one of its greatest advantages.
But innovation is partly a probability problem.
Not every engineer produces a breakthrough. Not every PhD develops a new model architecture. Not every research team finds something important.
But the larger the population capable of participating, the more opportunities there are for exceptional people and exceptional ideas to emerge.
This was one of the central arguments in The AI Innovation Machine: human capital matters, but so does the percentage of that human capital actually participating in the innovation process.
If breakthrough talent sits in the tail of the distribution, China’s unusually large pool of engineers and researchers gives it more chances to produce the outliers who change the game.
4. COMPETITION CREATES MORE EXPERIMENTS.
China has no national AI laboratory pursuing a single, centrally determined approach.
DeepSeek, Alibaba/Qwen, ByteDance, Tencent, Baidu, Moonshot AI, MiniMax, Zhipu, and numerous others compete intensely for models, applications, talent, customers, and capital.
That matters because innovation is partly a search problem.
Nobody knows beforehand which architecture, training technique, reasoning method, or application model will produce the next breakthrough. The more genuinely independent approaches being tried, the larger the solution space being explored.
Competition also creates pressure. A successful innovation from one company quickly becomes something its competitors must understand, reproduce or improve upon.
The important metric may therefore not simply be how much money the largest laboratory has. It may also be how many sufficiently well-resourced laboratories are trying sufficiently different things.
That is competitive diversity.
And in innovation:
China’s intense domestic competition means more experiments, more failures, and more chances that one of those experiments produces an unexpected breakthrough.
5. OPEN WEIGHTS TURN INNOVATION INTO A PRODUCT INNOVATION MULTIPLIER.
This may be one of China’s most underestimated advantages.
Many of the leading American frontier models remain proprietary. You can use them, usually through an API, but you cannot necessarily inspect their weights, modify them freely, deploy them wherever you want, or build your own technology directly on top of them.
China has developed a remarkably strong open-weight ecosystem around model families such as Qwen and DeepSeek.
That changes how innovation spreads. When an improvement appears in an open-weight model, thousands of researchers, companies, and developers can experiment with it, fine-tune it, distill it, optimize it for particular hardware, and build applications around it.
The original innovator therefore does not conduct all subsequent experimentation. The ecosystem does. We can already see this effect in the numbers. By August 2026, Hugging Face counted more than 151,000 derivative models built on Qwen, around 2.6 times Meta’s overall derivative-model footprint and 4.7 times the number built specifically on Llama. During the first seven months of 2026, developers were adding roughly 180–210 new Qwen derivatives every day. The overwhelming majority were created not by Alibaba, but by the wider developer community.
For companies, this matters enormously. A telecom operator, bank, or industrial company may ultimately care less whether OpenAI’s frontier model is marginally smarter than Qwen than whether it can deploy a sufficiently capable model inside its own infrastructure, control its data, customize it, and avoid paying continually for proprietary API access.
Open weights are therefore not simply a philosophical position about openness.
Closed models keep most experimentation inside the company that built them. China’s open-weight ecosystem turns each model release into a platform for experimentation across the wider economy.
6. CHINA CAN INNOVATE ACROSS THE ENTIRE STACK.
An AI model does not exist in isolation.
Modern AI depends on a chain: Semiconductors → Memory → Packaging → Networking → Compilers → Software frameworks → Model architectures → Inference systems → Applications.
China is increasingly being forced to work across the entire chain.
That is difficult because weakness in one layer can constrain the entire system. But it also creates incentives for hardware/software co-design: if you cannot simply purchase the world’s best GPU, you have stronger reasons to redesign software and models around the hardware you can manufacture or obtain.
DeepSeek’s engineering is interesting here, too. Its efficiency comes not from one miraculous algorithm but from multiple architectural and systems-level choices working together.
Huawei provides another example. Its Ascend accelerators do not need to become identical copies of NVIDIA GPUs. Huawei needs to create a sufficiently competitive combination of silicon, interconnect, software, and models to perform the required workload.
That changes the problem from:
How do we reproduce NVIDIA?
to:
How do we reproduce the capability NVIDIA currently enables?
China’s opportunity is not necessarily to replicate every Western technology, but to redesign the system so critical Western components matter less.
7. CHINA HAS A MASSIVE INDUSTRIAL LABORATORY FOR AI.
Much of today’s discussion treats AI as if its ultimate purpose were to generate words, images and software.
That may be only the beginning.
AI is increasingly moving into factories, robots, autonomous systems, logistics, telecommunications, energy systems and physical infrastructure. Here China possesses something the United States cannot simply reproduce by building another data center: an enormous manufacturing economy in which AI can interact with the physical world.
China already accounts for more than half of annual global industrial-robot installations and operates more industrial robots than any other country.
That creates millions of opportunities to experiment with machine vision, predictive maintenance, industrial optimization, robotics, autonomous logistics, and embodied intelligence.
Data generated by real machines operating in real factories can feed back into models, which improve the machines, which generate more data.
If the next AI revolution moves from generating tokens to controlling atoms, China’s industrial base becomes part of its AI infrastructure.
China’s manufacturing scale may give it a structural AI advantage: millions of machines, factories, robots, and industrial processes in which AI can be tested, improved, and deployed at scale.
8. THE WINNER MAY NOT BE THE MOST INTELLIGENT MODEL, BUT THE CHEAPEST MODEL SMART ENOUGH.
AI benchmark discussions naturally focus on which model ranks first.
Businesses tend to ask a different question:
What does it cost to solve my problem?
Suppose one model delivers an intelligence level of 100 at a cost of 100, while another delivers 95 at a cost of 20.
For some frontier scientific applications, that missing five percent may be decisive.
But for millions of enterprise applications, customer service, network operations, document analysis, coding assistance, translation, industrial automation or internal knowledge systems, the second model may be economically superior. This is already becoming an operational, not theoretical, choice. AT&T describes using an AI gateway that routes workloads toward the most cost-effective model for each task while developing open models specifically for telecom use cases. In other words, the optimization target increasingly becomes cost per successfully completed task, not simply access to the world’s most powerful model. Recent (i.e., 2026) FT analysis goes further, citing evidence that smaller models can reduce inference costs by 60–80% for suitable routine workloads.
That is why the rapid narrowing of the US-China frontier-model gap is so interesting. China’s AI-specific resource base remains considerably smaller in several critical areas, yet model performance has converged surprisingly quickly.
If Chinese developers keep improving intelligence per unit of resource, they don’t necessarily need to build the world’s single smartest model.
They need to make sufficiently capable intelligence dramatically cheaper to deploy.
China may not need to build the world’s most intelligent model. If it can build models that are smart enough for most applications at dramatically lower cost, that could become a major commercial advantage.
9. CHINA PLAYS A LONGER GAME.
Technology development takes time. Semiconductor technology takes a lot of time.
Western companies certainly make investments over decades, and Western governments can pursue long-term industrial policies. But democratic governments also operate through elections, changing administrations, annual budgets, and shifting political priorities. Public companies live with quarterly reporting and continual capital-market scrutiny.
China’s strategic industrial objectives can operate on a different clock.
That longer horizon is not rhetorical. China’s current Fifteenth Five-Year Plan for 2026–2030 explicitly calls for decisive breakthroughs across the integrated-circuit supply chain and other critical technologies, while major national science and technology programs are being designed with 2035 in view.
But perhaps the more interesting point is not simply that China plans further ahead. It is what happens when those plans miss their targets.
Western analysis often evaluates Chinese industrial policy by asking whether it achieved the stated target. That is an obvious measure, but perhaps not always the most revealing one.
The more interesting question may be where the capability floor sits when the target is missed.
We now have something useful we did not have ten years ago. Enough evidence to test that idea. In November 2025, the U.S.-China Economic and Security Review Commission (USCC) published a detailed retrospective on Made in China 2025, the industrial strategy launched a decade earlier. Its conclusion was remarkably nuanced. China broadly achieved its ambitious objectives in only about half of the targeted sectors. It fell substantially short in some of the hardest areas, particularly advanced semiconductors. But the Commission nevertheless concluded that China had become more innovative, moved significantly up the global value chain, and strengthened its position as a global manufacturing powerhouse.
The semiconductor numbers illustrate why simply asking whether a target was “met” or “missed” can be misleading.
China had targeted around 50% semiconductor self-sufficiency by 2020. According to the USCC assessment, it achieved only 16.6%. Domestic chipmaking equipment likewise reached only around 16% self-sufficiency by the third quarter of 2024, far below the original ambition.
On a conventional scorecard, that is failure. But underneath the missed targets, the starting position moved.
China’s share of global foundational-logic wafer capacity increased from approximately 19% in 2015 to 33% in 2023, while its mature-node semiconductor capacity expanded more than four times faster than global demand. The Commission estimates that Chinese state-led semiconductor investment since 2014 had exceeded USD 150 billion by 2024. SMIC meanwhile became the world’s third-largest semiconductor foundry by revenue.
That does not look like technological success against the original target. It looks more like failure to meet the timetable while substantially raising the starting point for the next timetable.
That is what I think of as a capability ratchet: ambitious targets may repeatedly be missed, but each attempt leaves behind accumulated skills, infrastructure, suppliers, technology and experience, so the next cycle starts from a higher floor.
The target is set above the existing capability. Investment follows. The target is missed. But the fabs, engineers, suppliers, equipment, infrastructure, operating experience, and accumulated know-how created during the attempt do not disappear.
The next cycle therefore begins from a higher floor.
Target → investment → partial achievement → higher capability floor → new target.
The target can be missed repeatedly while each cycle leaves China less dependent than the one before.
For Western policy, that creates an uncomfortable possibility: export controls may succeed in keeping China below the frontier while simultaneously raising the floor from which China attacks the next one. And this pattern is not confined to semiconductors.
The USCC found that China met or exceeded many of its objectives in electric vehicles, electrical equipment, biopharma, high-performance medical devices, shipbuilding, and space-related technologies. Even in sectors where it missed formal targets, it often made substantial gains in market share, localization, and technological capability.
Other independent assessments reach broadly similar conclusions. A 2025 Rhodium Group study commissioned by the U.S. Chamber of Commerce found that China had substantially reduced import dependencies and built globally competitive positions across several strategic industries, while remaining dependent on foreign technology in particularly difficult areas such as advanced semiconductors, aerospace components, and high-end industrial equipment. Importantly, 62% of foreign companies surveyed expected their Chinese competitors to catch up technologically within five to ten years.
The European Union Chamber of Commerce in China reached a similarly uncomfortable conclusion: Made in China 2025 coincided with substantial improvements in China’s advanced-manufacturing capabilities and helped turn China into a manufacturing superpower, even as it generated overcapacity, market distortions, and high costs for foreign competitors.
This matters because China did not respond to the targets it missed by abandoning the strategy.
It doubled down.
China established its first major national semiconductor investment fund in 2014. A second followed in 2019. A third, and substantially larger, fund followed in 2024 with registered capital of approximately RMB 344 billion (ca. USD 47.5 billion). The third fund alone is larger than either predecessor.
But the capability ratchet is not automatic. The same USCC assessment provides an important counterexample: commercial aviation. China has provided sustained state funding, political priority, and a large domestic market, yet COMAC still depends substantially on foreign technology and has not achieved anything resembling the progress seen in areas such as EVs, telecommunications equipment, or mature semiconductors.
In other words: the ratchet can jam. That suggests that the decisive issue is not simply whether China is willing to spend enough or wait long enough. It is what kind of technological problem China is trying to solve.
Where progress depends heavily on manufacturing scale, process learning, supplier development, capital investment, and repeated engineering cycles, capability can accumulate even when headline targets are missed.
Where the binding constraint is deeply accumulated tacit knowledge, extreme precision, tightly controlled intellectual property, or supplier capabilities concentrated inside a handful of incumbents, persistence may be far less effective.
That distinction matters enormously for AI. The useful question is therefore probably not whether China meets every semiconductor and AI objective in its 2030 plans. It almost certainly will not.
The more interesting question is:
Where will China’s capability floor be in 2030 when it misses?
China could still remain well below the US frontier while being radically stronger than it is today: domestic accelerators sufficiently capable for much of inference, a memory supply less vulnerable to export restrictions, substantially more domestic semiconductor equipment, and AI software increasingly optimized around Chinese hardware.
That would not constitute victory.
But it would not resemble the technological dependency structure that Western policy is trying to preserve today.
And that leaves one particularly important unresolved question.
Foundational semiconductors ratcheted. Jet engines largely did not. Which one does advanced lithography resemble?
Nobody knows yet.
If advanced lithography is primarily constrained by tacit engineering knowledge, extreme precision, and an extraordinarily difficult supplier ecosystem, China’s capability floor may remain below the Western frontier for a very long time.
But if enough of the problem can be broken down into engineering challenges that improve through investment, experimentation, supplier development, and repeated manufacturing cycles, the floor may keep moving upward even as China repeatedly misses its stated targets.
That classification may matter more to the long-term AI race than whether one Chinese frontier model scores a few percentage points higher or lower on today’s benchmark.
There is an interesting echo of Sun Tzu’s strategic thinking here:
“The victorious army first realizes the conditions for victory, and then seeks to engage in battle.”
China does not necessarily have to win the semiconductor race in 2026 or 2028.
It needs to keep shrinking each successive dependency, build the industrial ecosystem around the remaining gaps, and remain willing to try again when the first timetable proves too optimistic.
In a technology race measured in decades, China’s ability to keep raising the capability floor, even while missing the target, may itself become a competitive advantage.
10. HISTORY SAYS CONSTRAINTS SOMETIMES PRODUCE BREAKTHROUGHS.
There is a broader reason not to assume that China’s technological constraints will simply translate into a permanently lower rate of innovation.
History suggests that constraints can sometimes produce breakthroughs.
Josef Taalbi’s study “What Drives Innovation? Evidence from Economic History” of Swedish product innovation from 1970 to 2007 found that many innovations emerged as responses to specific problems, discrete events or new technological opportunities, with the economic and energy crises of the 1970s stimulating problem-driven innovation.
This does not mean that scarcity is inherently good for innovation. Clearly it isn’t. Too little capital, too little compute, too few skilled people, or insufficient access to technology eventually makes innovation harder, not easier. Research on innovation under constraints points to precisely this balance: some constraints can stimulate creativity and radical innovation, while excessive constraints suppress it.
But the right kind of constraint changes behavior.
When an existing solution becomes too expensive, unavailable, or simply impossible, engineers must question assumptions that would otherwise remain unchallenged. They redesign algorithms. They substitute materials. They change architectures. They integrate hardware and software differently. Sometimes they discover that the supposedly indispensable resource was indispensable only because the previous system had been designed around its abundance. A substantial literature on so-called frugal innovation documents precisely this phenomenon: resource constraints can change not merely the cost of a solution, but its design, engineering, and architecture.
That is essentially what I experienced as a physics student in the 1980s. Computing resources were scarce, so wasting compute was not an attractive option. We had to think harder about the mathematics, the algorithms, and the code.
China may now be experiencing something similar on an incomparably larger scale.
Advanced GPUs are constrained? Improve model efficiency.
Frontier semiconductor manufacturing equipment is unavailable? Develop domestic alternatives.
EUV lithography cannot be purchased? Push DUV further while developing alternative lithography capabilities. This is already happening. TechInsights confirmed that SMIC manufactured Huawei’s Kirin 9000S using its second-generation 7nm process without EUV, relying instead on considerably more complex DUV multiple-patterning. The approach carries yield, cost, and cycle-time penalties, but it demonstrates something important: denying access to the preferred technological route does not necessarily eliminate the destination. It can increase the incentive to find another route.
High-bandwidth memory is constrained? Improve memory utilization, packaging, and system architecture.
NVIDIA’s ecosystem is difficult to access? Build an alternative hardware and software stack. If you have sufficient resources, you can make up for less efficient hardware with quantity.
None of these responses guarantees success. But each restriction increases the economic value of finding another route.
And we have seen something resembling this Chinese technology trajectory before.
Twenty-five years ago, Huawei and ZTE were widely regarded as lower-cost followers of Ericsson, Nokia, Siemens, Alcatel, Nortel and Motorola. Today Huawei is the world’s largest RAN supplier and, in several areas of mobile-network technology, operates at or near the technological frontier. The lesson is not that semiconductors will necessarily follow the same trajectory. They are arguably a much harder industrial challenge. Today’s technological gap should not automatically be mistaken for a permanent technological hierarchy.
Another important difference lies between invention and catch-up. China does not have to rediscover the physics behind EUV lithography, high-bandwidth memory, or advanced semiconductor manufacturing. Much of the underlying science is known. The harder problem is engineering: finding implementations that avoid protected or inaccessible technologies and then turning those implementations into reliable, high-throughput, high-yield industrial systems. That remains enormously difficult. But it is different from starting from zero.
Western technology restrictions are intended to preserve an existing advantage by making critical resources scarcer for China. They may succeed. But every resource made scarce also becomes more valuable to substitute, redesign around, or ultimately make less important.
The constraints meant to slow China may also push it toward the efficiency breakthroughs that narrow the gap.
10 REASONS CHINA MAY NOT WIN THE AI RACE.
The same Breakthrough Model also provides the counterargument. The parameters do not inherently favor China. America’s extraordinary frontier-AI resources (R), its ability to attract and concentrate global human capital (H), deep participation through universities, startups and frontier laboratories (P), competitive diversity (D), rapid international knowledge diffusion (K), and exceptional ability to finance and scale successful ideas (S) remain formidable advantages.
And constraints (C) cut both ways. They may stimulate Chinese efficiency and substitution, but beyond a point, they simply restrict the experiments China can run. The ten reasons below therefore ask whether the US and wider Western ecosystem’s advantages across the same breakthrough drivers ultimately outweigh the innovation pressure that scarcity creates for China.
The first ten reasons make a credible case that scarcity, talent scale, competition, openness, industrial depth, and strategic patience could allow China to overcome some of its current disadvantages.
But Scarcity can stimulate innovation, but it can also simply leave you with fewer resources.
It is easy to romanticize constraints when looking backward at successful innovations. We remember the engineers who found ingenious ways around limitations. We hear much less about the thousands of experiments that were never attempted because the necessary resources simply weren’t available.
So let’s reverse the argument. If we make the strongest possible case that China could win the AI race, intellectual consistency requires making an equally strong case for why it may not. And that case is maybe unsurprising, and also formidable (imo).
1. AT THE FRONTIER, COMPUTE STILL MATTERS.
American technology companies are spending hundreds of billions of dollars building AI infrastructure for a reason.
Compute works.
The extraordinary improvement in AI capabilities over recent years has been driven not only by better algorithms but by scaling: more compute, more data, larger training runs, increasingly sophisticated post-training, and enormous amounts of experimentation. Kaplan and coworkers established the foundational language model scaling laws underlying this progress in their paper, “Scaling Laws for Neural Language Models,” demonstrating that model performance improves predictably with increases in model size, training data, and compute. These scaling relationships have helped shape the expansion of computational resources devoted to frontier AI.
Efficiency improvements do not eliminate that relationship.
If a Chinese laboratory discovers how to achieve the same intelligence using half the compute, that is an important innovation. But if an American laboratory can adopt the same technique and apply it to ten times as much compute, scarcity may have produced the innovation without producing a lasting competitive advantage.
Compute also buys something less visible: experiments.
Frontier AI research involves trying ideas that fail. Organizations with enormous computational resources can run more experiments, test larger architectures, and discard unsuccessful approaches without betting the company on each attempt.
Scarcity may make individual experiments more efficient.
America’s compute abundance lets its frontier labs run more experiments, test more ideas, and scale the winners faster than China can.
2. KNOWING THE PHYSICS IS NOT THE SAME AS MASTERING THE FACTORY.
China has excellent physicists, semiconductor engineers, and materials scientists. Many Chinese engineers have studied or worked inside the Western semiconductor ecosystem. The fundamental physics behind advanced semiconductor manufacturing is hardly a secret.
That does not make the technology easy to reproduce.
EUV lithography illustrates the problem perfectly. An advanced lithography system combines around 100,000 components involving extreme ultraviolet light generation, extraordinary optical precision, vacuum systems, wafer positioning, metrology, materials science, contamination control, and sophisticated software.
And all of it must work together.
Producing EUV light in a laboratory is one achievement. Building equipment that can expose wafers repeatedly, with nanometre-scale accuracy, high throughput, acceptable yield and industrial reliability is something very different.
The same applies to high-bandwidth memory, advanced packaging and semiconductor manufacturing equipment.
China does not necessarily need to reproduce Western implementations—and patents and trade secrets provide additional incentives to find different solutions. But it must reproduce the capability, economics, reliability, and manufacturing scale.
Meanwhile, ASML, TSMC, SK hynix, Samsung, NVIDIA, and the rest of the frontier ecosystem are not standing still.
China must catch up while the West keeps moving ahead. Turning semiconductor catch-up into a race against a continuously advancing frontier.
3. AMERICA STILL ATTRACTS AN EXTRAORDINARY SHARE OF THE WORLD’S BEST AI TALENT.
China’s human-capital numbers are formidable.
But where talented people are born and where they ultimately conduct their most important research are two different things.
MacroPolo’s Global AI Talent Tracker illustrates the paradox beautifully. China has become the largest country of origin for top AI researchers, while the United States remains the dominant destination for elite international AI talent. In the report “Have Top Chinese AI Researchers Stayed in the United States?” from Carnegie Endowment, we can read:
Researchers from China have long been one of the largest contributors to cutting-edge artificial intelligence research at American companies and universities. Studies of the top AI research papers have shown the authors originally from China contributing as much—if not more—to American AI output than authors from the United States.
That attraction is an enormous American advantage.
The US innovation system has historically been remarkably effective at attracting exceptional scientists, engineers, and entrepreneurs from China, India, Europe, and elsewhere, and placing them in universities, startups, and technology companies with extraordinary resources.
OpenAI, Anthropic, Google DeepMind, Meta, NVIDIA, and the leading American universities consequently concentrate talent drawn from far beyond America’s domestic population.
China may therefore produce more of the raw human capital.
America’s historical superpower status has persuaded much of the world’s human capital to work there.
However, an important question remains: can America take this advantage for granted? The US research system depends unusually heavily on foreign talent. In 2024, temporary visa holders earned 61% of US doctorates in computer and information sciences, 54% in engineering and 52% in mathematics and statistics. Yet there are emerging warning signs. International science-and-engineering enrollment at US universities fell 9% between 2024 and 2025, while new international student enrollment overall fell 17% in fall 2025. Visa delays and denials remain among the most frequently cited obstacles to international enrollment. Despite the broader decline, China remained the largest source of international Science & Engineering (S&E) doctoral students in the US in 2025, with about 38,000 students. Chinese S&E master’s enrollment fell about 8% between 2024 and 2025. The United States is competing technologically with China while also educating tens of thousands of the Chinese scientists and engineers who make up that competition.
At the same time, federal research funding and international scientific collaboration have become more uncertain. In 2026, NSF was on course to issue around 30% fewer new research grants than the previous year. The lowest number in more than four decades. New restrictions have also been introduced on collaboration with institutions considered national-security risks.
It is too early to conclude that these developments are causing elite AI researchers to leave, or avoid, the United States. America remains the world’s most powerful magnet for international scientific talent. But that advantage is not assured.
China may produce enormous amounts of technical talent, but America’s ability to attract, retain, and concentrate the world’s best researchers gives the US and the wider Western ecosystem a powerful advantage.
4. CAPITAL BUYS MORE AND BIGGER EXPERIMENTS.
Competitive diversity matters. But experimentation costs money. And at the AI frontier, it costs extraordinary amounts.
The gap in private AI investment between the United States and China is enormous. Stanford’s 2026 AI Index estimated US private AI investment at roughly USD 286 billion in 2025 compared with approximately USD 12 billion in China.
Chinese state investment makes the total resource comparison less extreme than those private numbers suggest. Nevertheless, America’s capital markets provide something centrally directed industrial policy struggles to reproduce: enormous amounts of risk capital distributed across many competing companies and investment decisions.
Some of that money will undoubtedly be wasted. But waste is partly what experimentation looks like before we know which experiment succeeds.
A system capable of financing ten enormous failures may also finance the eleventh company that discovers something transformative.
America’s deeper capital markets let its frontier labs fund more experiments, absorb more failures, and keep backing the few ideas that ultimately work. An advantage China may not easily replicate at the same scale.
5. OPEN WEIGHTS WORK BOTH WAYS.
China’s open-weight ecosystem may accelerate innovation enormously.
It also creates a strategic problem.
If DeepSeek, Qwen, or another Chinese laboratory discovers a fundamentally more efficient model architecture and publishes enough information for others to reproduce it, American researchers can learn from the breakthrough too.
Ideas travel much more easily than GPUs.
That means scarcity-driven Chinese innovation may diffuse into an American ecosystem possessing substantially greater computational resources.
Imagine China discovers an architecture capable of delivering today’s frontier performance with one-fifth of the compute. That would be a remarkable achievement. But what happens when an American frontier laboratory adopts the same architecture and gives it five or ten times more compute?
Open weights therefore create a paradox. They increase China’s ability to diffuse innovations throughout its own ecosystem and internationally, but they can also shorten the period during which those innovations give China a unique advantage.
The enormous derivative ecosystems already forming around Qwen and other open-weight families demonstrate how rapidly model architectures and techniques can escape the organizational boundaries of their original developers.
China’s open-weight ecosystem may accelerate Chinese innovation, but it also makes China’s breakthroughs easier for the resource-rich West to adopt and scale.
6. THE WESTERN AI STACK REMAINS FORMIDABLE.
China is increasingly developing alternatives across the entire AI technology stack.
But today, much of the strongest stack remains concentrated in the United States and allied economies.
NVIDIA dominates frontier AI accelerators and has spent nearly two decades building CUDA and the surrounding developer ecosystem. American hyperscalers operate enormous cloud and AI infrastructure. TSMC leads advanced semiconductor fabrication. ASML remains uniquely capable in production EUV lithography. SK hynix, Samsung, and Micron dominate advanced memory. Western and Japanese companies remain important across semiconductor manufacturing equipment, EDA software, networking, and specialized components.
No single country owns this entire stack.
That is precisely the point.
China is not competing only against the United States. In many critical technologies, it competes against an interconnected US-European-Japanese-Korean-Taiwanese technology ecosystem built over decades.
China may eventually reproduce much of it domestically.
But doing so simultaneously across multiple interdependent layers is an extraordinary industrial challenge.
China is not trying to replace a single Western dependency. It is competing against an integrated US-led technology ecosystem built over decades.
7. INDUSTRIAL SCALE MAY NOT BE WHERE THE DECISIVE AI BREAKTHROUGH HAPPENS.
But what if the decisive breakthrough doesn’t happen there?
The next major step in artificial intelligence could instead come from autonomous software engineering, mathematics, scientific discovery, AI agents, synthetic-data generation, or even AI systems increasingly capable of improving the technologies used to build subsequent AI systems. This is no longer purely speculative. METR’s work on AI task-completion horizons shows that frontier models can complete increasingly long and complex software-engineering tasks autonomously, with the effective task horizon expanding rapidly across successive model generations. METR’s analysis of AI task-completion horizons provides one of the clearest empirical indicators of this progression.
In that world, America’s strengths in software, cloud computing, frontier models, universities and research laboratories may matter considerably more than China’s manufacturing scale.
This produces an interesting division. China may possess an extraordinary environment for developing AI that operates machines.
America may possess an extraordinary environment for developing AI that improves AI.
If AI progress increasingly becomes recursive, better AI helping researchers design better models, chips and software, the advantage of having the strongest frontier models could compound.
Ding’s diffusion-centric work complicates this picture. His research argues that China’s ability to generate technological innovation has historically exceeded its ability to diffuse general-purpose technologies throughout the wider economy. That evidence largely predates the 2023–2026 surge in Chinese open-weight models, however. The rapid growth of Qwen derivatives may therefore be early evidence of a different diffusion mechanism. Although global model adoption is not the same thing as economy-wide domestic diffusion.
Although not completely one-to-one with Ding’s domestic diffusion arguments, the Qwen derivative numbers (i.e., global diffusion) may be the first serious evidence that China is closing precisely that deficit. If diffusion capacity determines technological leadership, the open-weight strategy is a way of acquiring it without first fixing the domestic bottleneck. China may be routing around its diffusion deficit rather than closing it.
China’s industrial laboratory is an enormous asset, but only if the next decisive AI frontier is substantially industrial.
8. THE LAST 5% OF INTELLIGENCE MAY BE WORTH MORE THAN THE FIRST 95%
The economic argument for efficient Chinese models appears compelling.
Why pay 100 for intelligence of 100 if you can obtain intelligence of 95 for a cost of 20?
For many applications, you probably shouldn’t. But AI capability may not translate linearly into economic value. A model that is 95% as capable as the frontier may be virtually indistinguishable for summarising documents, customer service or translation.
But suppose the missing 5% determines whether an AI system can reliably write production software, conduct scientific research, manage complex autonomous tasks, or operate without continuous human supervision.
Then the economic difference between 95 and 100 could be enormous.
Capability thresholds can also be nonlinear. A system that must successfully execute twenty dependent steps has only about a 36% chance of completing the sequence if each step is 95% reliable. At 99% reliability, that rises to roughly 82%. What looks like a small improvement at the model level can therefore become the difference between an unreliable assistant and a practically useful autonomous system. The same phenomenon exists throughout technology. Small differences near a critical performance threshold can suddenly enable entirely new applications.
That is why frontier leadership still matters, even if cheaper models dominate large parts of the market.
China may reach the first 95% more efficiently. But if the final 5% unlocks capabilities we do not yet have, America’s frontier advantage could matter disproportionately.
9. AMERICA’S RESOURCE ADVANTAGE IS ENORMOUS
China is not resource-poor.
It has enormous electricity production, manufacturing capacity, infrastructure, universities, technology companies, and a huge domestic market. As discussed earlier, focusing only on GPUs dramatically understates the resources available to China’s innovation system.
But resources specifically optimized for frontier AI tell a very different story.
The United States currently possesses an extraordinary concentration of AI capital, compute infrastructure, and access to frontier hardware.
Stanford’s 2026 AI Index estimates that US private AI investment reached ~ USD 286 billion in 2025, compared with ca. USD 12 billion in China, a difference of more than 23 times. Stanford correctly cautions that private investment understates China’s total resources because the comparison does not fully capture Chinese government guidance funds and state-directed investment. Nevertheless, the difference in private capital available to fund competing AI companies, infrastructure, and experiments is enormous.
And the capital is increasingly being converted into physical compute.
Epoch AI estimates that five hyperscalers control around 71% of the world’s AI compute. The largest known AI data center already has computing capacity equivalent to approximately 1.1 million NVIDIA H100 GPUs, while gigawatt-scale AI data centers now require investments on the order of USD 38 billion per gigawatt of IT capacity.
This scale continues to increase remarkably quickly. Epoch estimates that the compute used to train frontier language models has grown by approximately 5× per year since 2020, while the cost of frontier training has increased by around 3.5× per year. A separate Epoch analysis finds that much of this historical increase in training compute has come from deploying ever-larger numbers of accelerators in parallel.
That matters because algorithmic efficiency and compute abundance are not substitutes in the simple sense that one eliminates the need for the other.
Both can improve simultaneously.
Epoch estimates that pre-training compute efficiency has itself improved by roughly 3× per year, meaning that the same level of performance can be achieved with dramatically less compute than before. Yet frontier laboratories have not responded by using less compute. They have largely reinvested those efficiency gains into training still more capable systems.
This is potentially the most difficult part of the scarcity argument for China.
Suppose a Chinese laboratory discovers an architecture that achieves the same capability using one-fifth of today’s compute. That is an important breakthrough. But unless the innovation remains proprietary, an American laboratory can potentially adopt the same architectural improvement and apply it to a much larger computational base.
Efficiency therefore reduces the amount of compute needed to reach a given capability. It does not necessarily reduce the advantage of having more compute available.
This matters because the argument in The AI Innovation Machine was never that resources don’t matter. It was that resources alone do not determine innovation. That distinction cuts both ways.
China may compensate for fewer frontier-AI resources through efficiency, competition, human capital, open-weight innovation, and different approaches to system design. Indeed, much of this article explains why that compensation could be surprisingly powerful.
But compensation is not elimination.
If the two innovation systems eventually become similarly efficient, similarly creative, and similarly capable of exploiting their human capital, the advantage shifts back toward the United States and the wider Western technology ecosystem. With substantially more frontier compute and capital available, Western laboratories can finance more experiments, operate larger clusters, and apply the same algorithmic breakthroughs at greater scale.
China may learn how to do more with less. But if the West can adopt the same innovations while still having much more, abundance becomes an advantage again. (if we assume that the West/the USA has an abundance lead compared to China).
10. TOO MUCH SCARCITY SIMPLY BECOMES SCARCITY.
And finally we return to where we started.
Constraints can stimulate innovation. Researchers facing limitations sometimes discover approaches that would never have been pursued in an environment of abundance.
But this relationship has limits.
The research literature points to a nuanced relationship rather than a simple “scarcity is good” conclusion. Keupp and Gassmann show that resource constraints can, under some conditions, trigger radical innovation, while Acar, Tarakci and van Knippenberg demonstrate more broadly that constraints can both stimulate and inhibit creativity depending on their nature and intensity.
The innovation literature does not suggest that ever-increasing scarcity produces ever-increasing creativity. Rather, there appears to be a point beyond which constraints stop focusing ingenuity and begin preventing experimentation.
A physicist with a slower computer may write a better algorithm.
A physicist without access to a computer cannot run the calculation.
The same applies to AI. Restricted access to frontier GPUs may encourage Chinese researchers to build more efficient models. But insufficient compute can also prevent them from testing next-generation ideas.
A lack of EUV may spur remarkable advances in DUV multi-patterning. But eventually the additional complexity, cost, and yield penalties may overwhelm the advantage.
Limited HBM (High Bandwidth Memory) may encourage better memory management. But memory bandwidth remains a physical constraint.
Scarcity can redirect innovation. It cannot repeal physics.
That leaves us with perhaps the most important uncertainty in this entire discussion:
China may innovate around constraints, but if those constraints eventually cap what it can build and test, the frontier advantage stays with the West.
We simply don’t know yet!
But we are not equally ignorant about everything. The frontier-capability race is genuinely hard to call. The question of whether China’s constraints eventually cap what it can build is genuinely open. What the last decade does tell us is how to read the answer when it arrives: not by whether China hits its 2030 targets, which it most probably will not do that, but by where the floor sits when it misses them.
REFERENCES & FURTHER READING.
Kim K. Larsen, “The AI Innovation Machine – Why the Most GPUs May Not Win,” TechNEconomY Blog, 7 September 2026. Analysis of the broader drivers of AI innovation beyond compute resources alone.
Mishal Husain, “Former Google China Chief Kai-Fu Lee: China Will Win the AI Race for Reach,” Bloomberg Weekend, 4 September 2026. Bloomberg interview with Kai-Fu Lee on the global AI race.
Richard Windsor, “China vs. USA – Chips for Free?” Radio Free Mobile, 31 August 2026. Radio Free Mobile analysis of Chinese and US AI hardware economics.
Stanford Institute for Human-Centered Artificial Intelligence, “Technical Performance,” The 2026 AI Index Report, 2026. Stanford AI Index analysis of US and Chinese frontier-model performance.
Stanford Institute for Human-Centered Artificial Intelligence, “Research and Development,” The 2026 AI Index Report, 2026. Stanford AI Index analysis of model development and training compute.
Stanford Institute for Human-Centered Artificial Intelligence, “Economy,” The 2026 AI Index Report, 2026. Stanford AI Index analysis of US and Chinese AI investment.
Epoch AI, “Trends in Artificial Intelligence,” Epoch AI, updated 2026. Analysis of frontier training compute, algorithmic efficiency, training costs, hyperscaler compute concentration and AI data-center scale.
DeepSeek-AI, “DeepSeek-V3 Technical Report,” arXiv, 27 December 2024. DeepSeek-V3 technical report on model architecture and compute efficiency.
Chenggang Zhao et al., “Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures,” arXiv, 14 May 2025. DeepSeek analysis of hardware-aware AI model co-design.
MacroPolo, “The Global AI Talent Tracker 2.0,” Paulson Institute. MacroPolo analysis of the global distribution and movement of top AI talent.
International Federation of Robotics, “World Robotics 2025 – Industrial Robots,” IFR, 25 September 2025. IFR data on China’s industrial robot installations and operational stock.
Dell’Oro Group, “RAN Market Stabilized in 2025,” Dell’Oro Group, 17 February 2026. Dell’Oro assessment of global RAN vendor rankings and Huawei’s market position.
ASML, “EUV Lithography Systems,” ASML. ASML history of EUV development and industrialization.
Reuters, “China Sets Up Third Fund With USD 47.5 Billion to Boost Semiconductor Sector,” Reuters, 27 May 2024. Reuters report on China’s third and largest semiconductor Big Fund
Congressional Research Service, “Made in China 2025 and China’s Industrial Policies,” CRS, updated 1 June 2026. CRS analysis of China’s industrial-policy objectives through 2035 and 2049.
Daron Acemoglu, “Directed Technical Change,” The Review of Economic Studies, October 2002. Economic analysis of how relative factor scarcity can redirect technological innovation.
Marcus Matthias Keupp and Oliver Gassmann, “Resource Constraints as Triggers of Radical Innovation: Longitudinal Evidence from the Manufacturing Sector,” Research Policy, September 2013. This is already reference #5 in The AI Innovation Machine.
Oguz A. Acar, Murat Tarakci and Daan van Knippenberg, “Creativity and Innovation Under Constraints: A Cross-Disciplinary Integrative Review,” Journal of Management, January 2019. This is already reference #6 in The AI Innovation Machine.
Sujai Shivakumar, Charles Wessner and Thomas Howell, “China’s Localization Drive in Semiconductors Gains Impetus from Allied Chip Export Controls,” Center for Strategic and International Studies, 24 March 2026. CSIS analysis of how semiconductor export controls are accelerating Chinese localization.
Adina Yakefu, Apolinário and Irene Solaiman, “State of Open Models: Summer 2026 Observations,” Hugging Face, 14 August 2026. Hugging Face analysis of Chinese open-model scale, Qwen adoption, and downstream model development.
National Science Board / National Center for Science and Engineering Statistics, “The State of U.S. Science and Engineering 2026 – STEM Talent: Education, Training, and Workforce,” National Science Foundation, 2026. NSF international comparison of science, engineering, and computer science doctorate production.
Matt Sheehan and Sophie Zhuang, “Have Top Chinese AI Researchers Stayed in the United States?” Carnegie Endowment for International Peace, 3 December 2025. Analysis of Chinese-origin elite AI researchers and US-China talent flows.
Andy Markus, “The Tokenomics Equation: Balancing Cost and Performance,” AT&T, 23 July 2026. AT&T’s production approach to model routing, cost optimization, and open AI models.
Financial Times, “How Big Is the Open-Model Threat to AI Hyperscalers?” Financial Times, 11 September 2026. Analysis of enterprise adoption and the cost economics of smaller and open-weight AI models.
Xinhua / State Council of the People’s Republic of China, “Outline of the 15th Five-Year Plan for National Economic and Social Development of the People’s Republic of China,” 13 March 2026. China’s 2026–2030 plan for integrated circuits, artificial intelligence and technology self-reliance.
Xinhua / Government of the People’s Republic of China, “China to Make Breakthroughs in Core Technologies, Achieve Sci-Tech Self-Reliance,” 5 March 2026. Official summary of China’s plans for full-chain breakthroughs in integrated circuits and other strategic technologies.
Reuters, “China Starts Production of Home-Grown Immersion DUV Chipmaking Tools, Source Says,” Reuters, 28 July 2026. Report on China’s emerging domestic immersion-lithography capability.
Financial Times, “Huawei Drives China’s Push to Make Advanced Chips,” Financial Times, 8 September 2026. Investigation of Huawei’s effort to build a domestic semiconductor and lithography supply ecosystem.
U.S.-China Economic and Security Review Commission, “Made in China 2025: Evaluating China’s Performance,” 14 November 2025. It provides the overall scorecard, semiconductor performance, localization figures, and analysis of why some long-term industrial strategies succeeded while others did not. USCC — Made in China 2025: Evaluating China’s Performance.
Camille Boullenois, Malcolm Black and Daniel H. Rosen, “Was Made in China 2025 Successful?” Rhodium Group, 5 May 2025, prepared for the U.S. Chamber of Commerce. Particularly valuable for import dependence, technological catch-up, and the finding that 62% of surveyed foreign firms expect Chinese competitors to catch up within five to ten years. Rhodium Group — Was Made in China 2025 Successful?.
European Union Chamber of Commerce in China, “Made in China 2025: The Cost of Technological Leadership,” 16 April 2025. Useful independent European assessment of China’s progress in advanced manufacturing, together with the costs of overcapacity and market distortion. European Chamber — Made in China 2025: The Cost of Technological Leadership.
Janne M. Korhonen, “Overcoming Scarcities Through Innovation: What Do Technologists Do When Faced With Constraints?” Ecological Economics, 2018. Analysis of scarcity-induced technological innovation and substitution under resource constraints.
Soumodip Sarkar and Sara Mateus, “Value Creation Using Minimal Resources – A Meta-Synthesis of Frugal Innovation,” Technological Forecasting and Social Change, June 2022. Meta-analysis of how innovation creates value under resource constraints.
Xu Yan and Minyi Huang, “Leveraging University Research Within the Context of Open Innovation: The Case of Huawei,” Telecommunications Policy, March 2022. Study of Huawei’s transition from technology follower to global technology leader.
TechInsights, “China’s SMIC Plays 7 nm Card,” TechInsights, 2023, updated analysis 2026. Technical analysis of SMIC’s use of DUV multiple-patterning to manufacture 7nm-class semiconductors without EUV.
TechInsights, “HiSilicon Kirin 9000s (SMIC 7nm, N+2) Process Flow Analysis,” TechInsights, 3 September 2026. Process analysis of China’s most advanced commercial logic manufacturing without EUV.
TechInsights, “Lithography: Paving the Way for Technological Advancement and Autonomy,” TechInsights, 30 June 2026. Analysis of DUV, EUV, and alternative lithography approaches, including China’s response to EUV restrictions.
Josef Taalbi, “What Drives Innovation? Evidence from Economic History,” Research Policy, Vol. 46, No. 8, October 2017, pp. 1437–1453. Historical evidence on innovation as a creative response to problems, crises, and new technological opportunities.
Jared Kaplan et al., “Scaling Laws for Neural Language Models,” arXiv, 23 January 2020. Foundational analysis of relationships between AI performance, model scale, data, and compute.
Samuel Carrara, Silvia Bobba, Darina Blagoeva et al., “Supply Chain Analysis and Material Demand Forecast in Strategic Technologies and Sectors in the EU – A Foresight Study,” European Commission, Joint Research Center, Publications Office of the European Union, 16 March 2023, JRC132889. European Commission JRC analysis of strategic technology supply chains.
METR, “Measuring AI Ability to Complete Long Tasks,” METR, 2025–2026. Empirical analysis of the increasing task-completion horizon of frontier AI models. See also “Measuring AI Ability to Complete Long Software Tasks” by Thomas Kwa et al.
Jeffrey Ding, “Technology and the Rise of Great Powers: How Diffusion Shapes Economic Competition,” Princeton University Press, 20 August 2024. Theory and historical evidence that technological leadership depends less on which state invents first than on which state adopts general-purpose technologies at scale across its whole economy.
Jeffrey Ding, “The Diffusion Deficit in Scientific and Technological Power: Re-assessing China’s Rise,” Review of International Political Economy, Vol. 31, No. 1, January 2024, pp. 173–198. Argues that assessments of China’s technological power overweight innovation capacity and underweight diffusion capacity, and finds China well short of science and technology superpower status on a diffusion-centric reading.
AI MODELS DISCUSSED IN THIS BLOG.
CLOSED & PROPRIETARY MODELS.
OpenAI (USA), “GPT-5,” OpenAI, 7 August 2025. OpenAI overview of the GPT-5 model family used in ChatGPT.
Google DeepMind (USA), “Gemini,” Google DeepMind. Google DeepMind overview of the Gemini family of multimodal AI models.
Anthropic (USA), “Claude,” Anthropic. Anthropic overview of the Claude family of proprietary AI models.
OPEN-WEIGHT MODELS.
Meta AI (USA), “Llama,” Meta. Meta’s official Llama model and resources page, and Hugging Face Meta Llama model resources.
Alibaba Cloud / Qwen Team (China), “Qwen,” Alibaba Cloud. Official Qwen model documentation, and Hugging Face Qwen model resources.
DeepSeek-AI (China), “DeepSeek,” DeepSeek. Official DeepSeek model and research site, and Hugging Face DeepSeek model resources.
Mistral AI (France), “Open Models,” Mistral AI. Mistral AI’s open-model documentation and Hugging Face Mistral AI resources.
MORE FULLY OPEN / OPEN RESEARCH MODELS.
Allen Institute for AI (USA), “OLMo: Open Language Model,” Ai2. AI2’s OLMo project provides open model weights, training code, evaluation tools, and training data for OLMo models, along with Hugging Face OLMO resources.
BigScience (France), “BLOOM: A 176B-Parameter Open-Access Multilingual Language Model,” BigScience, 2022. Hugging Face BLOOM model resources.
Technology Innovation Institute (UAE), “Falcon LLM,” TII. Falcon model family and research resources and Hugging Face Technology Innovation Institute models.
ABBREVIATIONS.
| Abbreviation | Meaning | Short explanation |
|---|---|---|
| AI | Artificial Intelligence | Computer systems capable of tasks associated with human intelligence, including reasoning, language, prediction and decision-making. |
| API | Application Programming Interface | A software interface that allows applications to access a model or service without hosting it themselves. |
| CUDA | Compute Unified Device Architecture | NVIDIA’s software platform and programming ecosystem for using NVIDIA GPUs for general-purpose and AI computing. |
| DUV | Deep Ultraviolet | Lithography technology using deep-ultraviolet light to manufacture semiconductor circuits. It predates EUV but remains extensively used. |
| EDA | Electronic Design Automation | Specialized software used to design, simulate, and verify semiconductor chips. |
| EU | European Union | Political and economic union of European member states. |
| EUV | Extreme Ultraviolet | Advanced semiconductor lithography using 13.5-nanometre wavelength light to manufacture leading-edge chips. |
| FT | Financial Times | International business and financial newspaper referenced in the article. |
| GPU | Graphics Processing Unit | Highly parallel processor originally developed for graphics and now central to training and running modern AI models. |
| HBM | High Bandwidth Memory | High-speed stacked memory placed close to AI accelerators to provide the very high data throughput required by advanced AI workloads. |
| IFR | International Federation of Robotics | An industry organization that produces global statistics on industrial robots and automation. |
| JRC | Joint Research Center | The European Commission’s science and knowledge service. |
| METR | Model Evaluation & Threat Research | Research organization studying frontier AI capabilities, including how long and complex a task AI systems can complete autonomously. |
| NSF | National Science Foundation | US federal agency supporting scientific and engineering research and education. |
| PhD | Doctor of Philosophy | Advanced research doctorate; used in the article when comparing scientific and engineering human capital. |
| R&D | Research and Development | Activities aimed at creating new knowledge, technologies, products or processes. |
| RAN | Radio Access Network | The part of a mobile network connecting user devices to the operator network through radio base stations. |
| S&E | Science and Engineering | Collective category used in US education and workforce statistics. |
| SMIC | Semiconductor Manufacturing International Corporation | China’s largest semiconductor foundry and a central company in China’s advanced-chip ambitions. |
| TSMC | Taiwan Semiconductor Manufacturing Company | The world’s leading advanced semiconductor foundry and manufacturer of many frontier AI processors. |
| US / USA | United States / United States of America | Used throughout the article when comparing the American and Chinese AI ecosystems. |
| USCC | U.S.-China Economic and Security Review Commission | US congressional commission assessing the economic and national-security implications of US-China relations. |
| AMEC | Advanced Micro-Fabrication Equipment Inc. China | Chinese semiconductor-equipment manufacturer particularly important in etching and deposition technologies. |
| CXMT | ChangXin Memory Technologies | Major Chinese memory-chip manufacturer working to reduce China’s dependence on foreign DRAM and advanced-memory suppliers. |
