I am a physicist by training, educated in the 1980s and working through the 1990s, when computing resources were anything but abundant. We did not have what today feels like an almost infinite amount of processing power at our disposal. Quite the opposite. Computing was expensive, memory was limited, storage was constrained, and waiting for a badly designed piece of code to finish could be painfully educational. As a consequence, much of the code we developed for data processing, modelling, and analysis was written with efficiency in mind. Performance mattered because the underlying computing infrastructure forced it to matter. You learned to think carefully about algorithms, memory, execution time, and what was actually necessary to solve the problem. Inefficient code was not simply inelegant; it could make an otherwise sensible analysis practically impossible.

Fast-forward to today, and the world looks very different. Compared with what we had available in the 1980s and 1990s, computing resources can almost appear infinite. On top of that infrastructure, we now have extraordinarily sophisticated AI models capable of doing much of the work that previously required significant amounts of specialised programming, analysis, and manual effort.

With the right human guidance, these models can often do the work dramatically faster and, depending on the experience of the person using them, sometimes better as well. The combination of increasingly capable models and vast underlying computing infrastructure has fundamentally changed what an individual can accomplish.

I am certainly not arguing that we should return to the way we worked in the 1980s or 1990s. I have no nostalgia for waiting hours for calculations that today take seconds, nor do I believe that artificial scarcity is somehow virtuous in itself. Abundant computing power is an extraordinary enabler of innovation.

But abundance has a downside. When computing resources become cheap, plentiful, and easily accessible, there is less pressure to care deeply about efficiency. Why spend weeks making an algorithm twice as efficient if you can simply add more compute? Why optimise memory usage if more memory is readily available? Why rethink an architecture if another rack of GPUs solves the problem?

Abundance can, in other words, breed a certain degree of complacency.

And that brings us to a much more interesting question. What happens when abundance is taken away?

If access to advanced processors, manufacturing technology, software tools, or other critical enablers is deliberately restricted for a company or even an entire country, the intuitive assumption is that innovation will slow down. And initially, it probably will.

But constraints also change behaviour!

When the obvious solution is no longer available, there is suddenly a powerful incentive to find another one: better algorithms, more efficient architectures, smarter use of memory, improved training techniques, alternative hardware, or entirely different approaches to the problem.

History repeatedly shows that scarcity can be a remarkably effective catalyst for ingenuity as described in work of Keupp & Gassmann and Acar, Tarakci & van Knippenberg.

The paradox is therefore that restricting access to a superior technology may not simply delay the party being restricted. Under the right conditions, it may encourage breakthroughs that reduce the importance of the very technology being withheld.

The strategically uncomfortable question is not only whether you can deny someone access to your technology. It is whether, by doing so, they might eventually teach themselves how to live without what they cannot get.

THE OBSESSION WITH INTELLIGENCE.

The AI debate remains strangely obsessed with one question:

Who has the most intelligent model? Is it OpenAI? Anthropic? Google? DeepSeek? Qwen?

For consumers, that may make some sense.

Industry does not necessarily need the most intelligent model. It needs intelligence that is good enough, economically deployable and deeply integrated into the business.

Take a telecom operator. Its most valuable AI applications are unlikely to be poetry or general knowledge. They will be diagnosing network faults, optimizing radio networks, supporting field technicians, improving cybersecurity, automating customer care, analyzing contracts, migrating legacy software and helping engineers plan networks.

For these tasks, does the operator really need the smartest general-purpose model in the world?

Or would it rather have a model that is perhaps slightly less clever, but cheaper, controllable, auditable, deployable inside its own infrastructure and deeply integrated with its own data and operational systems?

That is a very different proposition. The competition may therefore not be about the best model. It may be about the best industrial AI system.

A proprietary frontier model is essentially intelligence-as-a-service. An open-weight model allows the enterprise to combine a capable foundation model with its own data, workflows, engineers, infrastructure and domain knowledge.

Suddenly, it becomes my AI system, rather than somebody else’s intelligence embedded in my process.

This is why the biggest threat from Chinese open-weight models such as DeepSeek and Qwen may not be that they beat OpenAI or Anthropic on another benchmark.

It may be much more fundamental: What if foundation-model intelligence becomes a commodity?

If a telecom operator, manufacturer or bank can obtain something approaching 80% (or whatever appropriate for a given business) of frontier capability at a fraction of the cost, customize it and run it at massive scale without paying a toll on every inference, the economic center of gravity begins to move. Away from whoever owns the foundation model, and towards whoever owns the data, workflows, customers and domain knowledge.

That also changes the US-China discussion.

The USA model is increasingly built around massive capital, massive compute and highly concentrated frontier laboratories (see the Stanford AI Index 2026 and CB Insights).

The Chinese model appears to be evolving under stronger constraints: more emphasis on efficiency, more open-weight models, intense competition and faster diffusion of ideas.

In my opinion, the real question is which innovation system is better at creating, spreading and industrializing useful intelligence. Jeffrey Ding has made a closely related argument for general-purpose technologies more broadly: historically, the great powers that gained most from a new technology were not those that invented it first, but those that diffused it fastest across their economies.

The US clearly dominates resources and scaling power today.

China may be building strength in something different: diversity of experimentation and speed of knowledge diffusion.

And there is an interesting irony. If Chinese developers, because they cannot access unlimited compute, discover how to achieve the same result with one-fifth of the resources, American companies can use the same techniques and combine them with much greater compute.

So this may not ultimately be a contest between American and Chinese models. It may be a contest between two innovation structures, concentrated scale versus distributed experimentation.

And the bigger economic question may not be who owns the smartest model.

It may be who captures the value once intelligence itself is no longer scarce.

CLOSED, OPEN OR OPEN-WEIGHT?

Before going further, it is worth clarifying three terms that are often used rather loosely.

A closed or proprietary model is controlled by its developer. You access the intelligence through an API or application, but you normally cannot download the model, run it independently or fundamentally modify it. OpenAI’s GPT models, Anthropic’s Claude and Google’s Gemini broadly follow this model.

An open-weight model makes the trained model weights available. Subject to the licence, you can download the model, run it on your own infrastructure, fine-tune it, quantise it and integrate it deeply into your own systems. Qwen, DeepSeek, Llama, Gemma and many Mistral models fall into this category.

Important is to note that open-weight does not necessarily mean open-source.

The weights tell you what the model has learned. They do not necessarily tell you how it was created. The training data, filtering, reinforcement-learning process, synthetic-data generation and parts of the training code may remain proprietary.

A genuinely open-source AI model goes further by making substantially more of that underlying recipe available and allowing users to study, modify, and redistribute the system.

So the distinction is in theory quite simple:

Closed: you use somebody else’s model.

Open-weight: you can take the model with you.

Open-source: you can also see much more of how it was built.

Platforms such as Hugging Face have become important because they make millions of open-weight and open-source models easy to discover, download, test and deploy. Nvidia’s September 2026 agreement to acquire Hugging Face for ca. USD 13 billion is therefore strategically interesting in itself: the world’s dominant AI-chip company is acquiring one of the most important distribution platforms for open AI models.

For this article, the most important distinction is between proprietary and open-weight AI. Because once the weights become portable, the ownership boundary changes. Such a model can increasingly become part of your infrastructure, your data, and your processes rather than simply somebody else’s intelligence accessed through an API.

THE FOUR WISE MEN.

There is no shortage of people willing to tell us who will win the AI race. Depending on whom you listen to, the decisive factor is access to the most advanced GPUs, the ability to spend hundreds of billions of dollars on compute, possession of the best frontier model, control of an open AI ecosystem, or simply having enough very smart people relentlessly trying to solve the problem.

Four recent conversations are particularly interesting because they come from people sitting at very different places in the AI value chain: Kai-Fu Lee, Dario Amodei, Jensen Huang and Sam Altman. They certainly do not agree on everything. In fact, some of their views are almost diametrically opposed.

And that is exactly what makes them interesting.

Kai-Fu Lee provides perhaps the most provocative perspective. Lee has the unusual advantage of having spent much of his career deeply embedded in both the American and Chinese technology ecosystems. He worked at Apple, Microsoft and Google, established Microsoft Research Asia, ran Google China, and today leads Sinovation Ventures and 01.AI. In AI Superpowers (2018) he argued that AI was moving from an age of discovery, where fundamental research dominates, to an age of implementation, where engineering, data and speed of deployment matter more, and that this shift favors China. His recent interviews are essentially an update of that thesis for the era of open-weight models.

In a February 2026 Financial Times interview with Eleanor Olcott, Kai-Fu Lee described the difference between the American and Chinese approaches to AI using an analogy I particularly like. America, in his description, has a handful of exceptionally gifted students, OpenAI, Anthropic, Google and xAI, each convinced that it can solve the ultimate AGI problem first. They are competing for the Nobel Prize, so to speak, and they are prepared to spend extraordinary amounts of money trying to get there.

China looks rather different.

Kai-Fu Lee describes the Chinese ecosystem more like a study group. One company develops something, publishes a model, others study it, improve it, modify it, and contribute ideas back into the ecosystem. The companies are competitive, certainly, but knowledge flows much more freely between them, particularly through models commonly described as open source, although open-weight is often the technically more accurate description.

There is another important difference. Chinese companies generally cannot afford to behave as if capital and compute are unlimited.

As Kai-Fu Lee points out, Alibaba cannot simply decide to lose another $10 billion next quarter in pursuit of some distant AGI prize. American frontier laboratories backed by enormous pools of capital have considerably more freedom to do precisely that. Chinese AI companies therefore operate with stronger pressure to produce useful products, improve efficiency and generate commercial returns.

Scarcity changes behaviour.

Kai-Fu Lee does not argue that China currently leads the United States in frontier AI research. Quite the opposite. He acknowledges that much of the fundamental frontier innovation still originates in the United States. But he argues that strong Chinese engineering capabilities allow Chinese teams to understand those breakthroughs, reproduce them, modify them and sometimes find cheaper or more efficient ways of reaching comparable outcomes. His analogy is the moon landing: doing it first is extraordinarily difficult; doing it second becomes easier simply because somebody has demonstrated that it is possible.

Kai-Fu Lee develops this argument further in a September 2026 Bloomberg Weekend interview with Mishal Husain, “Former Google China Chief Kai-Fu Lee: China Will Win the AI Race for Reach.” Here, he compares the emerging AI market with smartphones. The American proprietary models may become the iPhone: technologically excellent, tightly controlled and highly profitable. Chinese open-weight models could become Android: perhaps capturing less profit per user, but spreading much further.

The strategic question is therefore not only who produces the most intelligent model. It is whose intelligence becomes embedded most widely across the world’s applications and industries.

Then we have Dario Amodei, CEO of Anthropic, who looks at almost exactly the same situation and reaches a rather different strategic conclusion.

In his June 2026 extended interview with Emily Chang for Bloomberg’s The Circuit, Dario Amodei again defended restrictions on exporting advanced AI chips to China. His argument is straightforward: access to frontier compute remains strategically important, and the United States should not voluntarily enable China to achieve parity or leadership in the most advanced AI capabilities.

For Dario Amodei, the frontier matters enormously.

He argues that there is a substantial economic premium associated with the most intelligent models. Open models somewhat behind the frontier can certainly create considerable value, but the exponential improvement of frontier models means that the largest opportunities continue to migrate towards the leading edge. From that perspective, restricting access to the hardware required to train the largest models is not merely industrial policy. It is a way of protecting a strategic technological advantage.

And Dario Amodei is not talking about modest amounts of computing infrastructure. In the same interview, he describes Anthropic planning for roughly tenfold annual growth in compute and nevertheless finding demand increasing even faster than expected. Compute, in this worldview, is not a peripheral ingredient of AI progress. It is one of its fundamental production factors.

If Kai-Fu Lee is essentially saying, constraints may make China more inventive, Dario Amodei is saying, yes, perhaps, but starving the competition of the most important production resource still matters.

Maybe both could be right.

Which brings us to Jensen Huang, CEO of Nvidia.

Jensen Huang has an obvious commercial interest in selling as many AI accelerators as possible, including to China, so his position should not be regarded as entirely detached from Nvidia’s business model. Nevertheless, it would be equally foolish to dismiss the views of the person whose company sits at the centre of the global AI compute infrastructure.

Jensen Huang has repeatedly warned that restricting Chinese access to American technology does not mean that China simply stops developing AI.

In a July 2026 Axios interview, against the backdrop of another wave of concern about Chinese models, Jensen Huang went further. He argued that Chinese open models were already excellent and should be used rather than banned. More fundamentally, he rejected the idea that keeping Chinese models out of Western markets would somehow make the United States safer or more competitive. His argument is that open models expand the overall AI market, accelerate adoption, encourage innovation and ultimately create more demand for infrastructure. Not less.

There is an important economic insight hidden in Jensen Huang’s position.

Making AI radically cheaper does not necessarily destroy the value of compute. It may do precisely the opposite. Economists have a name for this: the “Jevons paradox”, first observed by William Stanley Jevons in 1865, when more efficient steam engines increased rather than reduced Britain’s coal consumption. Satya Nadella, Chairman and CEO at Microsoft, invoked it the week DeepSeek‘s R1 was released.

If the cost of performing an intelligent task falls by a factor of ten, we should not automatically assume that the world will consume one-tenth as much compute. We may instead perform one hundred times as many intelligent tasks. Anyone familiar with the history of telecommunications should recognise the mechanism. Making a minute of communication dramatically cheaper did not reduce the importance of telecom networks. It created vastly more communication.

Efficiency and consumption can grow simultaneously.

And then there is Sam Altman, CEO of OpenAI, whose perspective is particularly interesting because OpenAI represents perhaps the purest expression of the American frontier-model strategy.

In a July 2026 CNBC interview, Sam Altman was asked directly about Chinese models catching up with American frontier laboratories. His answer was remarkably relaxed:

“The Chinese open-source models are getting very good.”

He nevertheless maintained that OpenAI would continue to produce the world’s best models and that customers would continue to value access to the frontier.

But another part of the same interview is arguably more important for the argument here. Asked about rising compute and memory costs, Sam Altman acknowledged that infrastructure cost had become a headwind and said that OpenAI therefore had to produce greater algorithmic gains to compensate while continuing to reduce the price of intelligence delivered to customers.

Read that again in the context of our discussion about scarcity.

Even one of the world’s best-funded AI companies, with access to extraordinary amounts of computing infrastructure, increasingly cares about doing more with less.

Efficiency may come back into fashion.

So what do our Four Wise Men actually tell us?

Probably not who will win (Sorry!)

Kai-Fu Lee tells us that scarcity, engineering intensity, openness and competitive collaboration can accelerate catch-up. Dario Amodei tells us that frontier capability and access to enormous computational resources still matter. Jensen Huang reminds us that restricting technology does not eliminate the demand for it and may accelerate alternative ecosystems. And Sam Altman demonstrates that even at the frontier, economics eventually forces attention back towards efficiency.

The interesting conclusion is therefore not that compute does not matter. Clearly, it does.

Nor is it that human talent alone determines innovation. Capital matters. Technology matters. Access to knowledge matters. Competition matters. The number of people capable of contributing matters. And, perhaps counterintuitively, constraints may matter as well.

This is where the discussion becomes considerably more interesting than simply counting Nvidia GPUs. Because if we genuinely want to understand whether the United States or China has the stronger long-term innovation engine, we need to ask a different question:

What actually determines the capacity of a society, an ecosystem or a company to produce technological breakthroughs?

THE BREAKTHROUGH FORMULA.

If the Four Wise Men teach us anything, it is that technological breakthroughs are unlikely to be explained by a single variable. That would be a fairly naive expectation.

It would certainly be convenient if they were. But naive.

Count the GPUs. Count the dollars. Count the engineers. Whoever has the biggest number wins. Innovation has never been quite that cooperative.

There is a substantial body of research suggesting that new ideas emerge from a combination of people, resources, incentives, competition, accumulated knowledge and the ability of ideas to move between people and organisations.

Paul Romer made Human Capital central to technological progress already in his seminal work on endogenous growth. Importantly, it is not simply population that matters, but the number of people with relevant skills actually contributing to the creation of new knowledge.

So relevant human capital matters.

But so does Participation.

Having capable scientists, engineers and entrepreneurs is one thing. Having large numbers of them actively experimenting, publishing, developing models, starting companies and challenging established approaches is something else. Nicholas Bloom, Charles Jones, John Van Reenen and Michael Webb make a related point in their work on declining research productivity: technological progress depends not only on the number of researchers, but on how productively research effort is converted into new ideas.

Then there are, obviously, Resources.

E.g., Compute. Capital. Energy. Data. Semiconductor technology. Cloud infrastructure. Access to sophisticated tools, etc.

Richard Windsor, in his recent analysis of Huawei’s Ascend processors, argues that China’s hardware disadvantage remains enormous. Based on Huawei’s own roadmap, he estimates its 2026 AI hardware to be more than three times as expensive to build and around nine times more power-hungry per unit of compute than Nvidia’s Blackwell generation. His conclusion is that China’s hardware disadvantage is likely to remain for the foreseeable future, giving the US and its allies a substantial advantage in AI.

Anthropic makes an even stronger version of essentially the same argument: compute is the most important ingredient of frontier AI, and sufficiently tight export controls could turn America’s current advantage into a durable 12-to-24-month capability lead by 2028.

Maybe, maybe, maybe not?

But technological history gives us plenty of reasons to be suspicious of what seems to be extrapolations of the past into the future. A large advantage today does not automatically become a permanent advantage tomorrow.

Despite the huge difference in available compute, Stanford’s 2026 AI Index found that, as of March 2026, the leading US model was only 2.7% ahead of the leading Chinese model on its comparison of frontier performance.

That does not mean compute is unimportant. Clearly, it really is! (i.e., within a given comparable context).

It does suggest that the conversion of compute into useful intelligence is not fixed. Epoch AI has estimated that, since 2012, the compute needed to reach a given level of language-model performance has halved roughly every eight months, meaning that algorithmic progress alone has delivered compute-equivalent gains comparable to, or faster than, the hardware gains from Moore’s law over the same period. A fixed GPU inventory therefore buys a moving amount of capability.

Innovation also depends on Competition and Diversity of approaches.

Philippe Aghion and colleagues found an inverted-U relationship between competition and innovation: too little competition encourages complacency, while too much may reduce the incentive or ability to invest.

Diversity matters because breakthroughs often come not from inventing everything from scratch, but from combining existing knowledge in unexpected ways. Brian Uzzi and colleagues, analysing almost 18 million scientific papers, found that highly impactful science often combined conventional knowledge with something unusually novel.

Having more independent teams exploring different paths may therefore matter. Which brings us back to Kai-Fu Lee’s Chinese “study group”. Because generating knowledge is only half the story.

The other half is how quickly knowledge spreads (i.e., knowledge diffusion).

Innovation economists have studied knowledge spillovers for decades. Cohen and Levinthal’s concept of absorptive capacity is particularly relevant: organizations differ enormously in their ability to recognize valuable external knowledge, absorb it and turn it into something useful.

Open-weight AI potentially changes that dynamic. A model that can be inspected, adapted, modified and built upon does more than lower the price of intelligence. It allows many more people and organizations to experiment with it.

More participants create more experiments. More experiments create more opportunities for somebody to discover something unexpected.

But good ideas also need to be recognized, funded, industrialized and scaled. Furman, Porter and Stern’s work on national innovative capacity makes precisely this point: innovation depends not simply on R&D expenditure, but on the wider ecosystem connecting research institutions, companies, capital and markets.

A brilliant idea that never escapes the laboratory remains a brilliant idea.

And finally, the factor that, for me, got this article started:

the Constraints.

Scarcity is obviously not automatically good. Remove enough resources and innovation stops.

Economists have long recognized the mechanism. John Hicks argued in 1932 that a rise in the price of one factor of production induces inventions that economize on that factor, and Daron Acemoglu later formalized this as directed technical change: innovation flows toward saving whatever input has become relatively scarce. Applied to AI, restricting compute does not stop innovation; it redirects it toward compute-saving algorithms. Research does suggest that the right kinds of constraints can change how problems are solved. Marcus Keupp and Oliver Gassmann found that resource constraints can stimulate radical innovation by forcing companies towards novel combinations and different approaches. Oguz Acar, Murat Tarakci and Daan van Knippenberg similarly conclude from a broad review of the literature that constraints can either suppress or stimulate creativity and innovation, depending on their nature and severity.

Having fewer GPUs may encourage you to design a more efficient model.

Having no GPUs probably does not … Right 😉

So the outlines of an AI innovation machine begin to emerge.

It needs capable people. Enough of them must participate. They need resources. They need competing ideas and different approaches. Knowledge needs to spread. Successful ideas need to be recognized and scaled. And constraints may change how efficiently all those ingredients are used.

None of these factors operates independently.

Which is why comparing the United States and China becomes considerably more interesting than comparing their GPU inventories.

I want to turn these ingredients into a simple framework for thinking about the likelihood of a breakthrough. The reader should note that I am not suggesting that technological breakthroughs can be predicted by an equation. I think that innovation is far too messy for that. But a simple framework can help us think about what makes breakthroughs more or less likely.

I will call it the Breakthrough (likelihood) Formula:

where B represents the likelihood of meaningful technological breakthroughs.

The logic is fairly intuitive and if we look at the thoughts of our Four Wise Men we can map their ideas to the following:

Perspective	Central proposition	Our formula
Kai-Fu Lee	Constraints + open models drive Chinese efficiency and collective learning	D × K (Diversity x Knowledge).
Dario Amodei	Frontier compute remains strategically decisive	R (Resources)
Jensen Huang	Open models expand the AI ecosystem and ultimately compute demand	K → D → R (Knowledge, Diversity & Resources)
Sam Altman	Compute scarcity creates concentration; solve it through massive abundance	C ↑ → η(R)* ↑ (Constraints, More Efficient use of Resources)
This blogs proposition	Breakthroughs depend on the interaction of all these mechanisms	B = f(H·P,R,D,K,S∣C)

Human capital, H, is the pool of people with the relevant skills and knowledge. But population alone is not enough. What matters is how much of that talent actually participates in research, engineering, entrepreneurship and experimentation. That is Participation, (P).

Together, H·P gives us something like the effective innovation population.

China’s sheer scale matters here. According to the US National Science Foundation, China awarded approximately 53,000 science and engineering doctorates in 2022, compared with about 45,000 in the United States in 2023. More strikingly, China produced around 30,000 engineering and computer-science doctorates, roughly twice the US number. That talent is increasingly visible in AI output: China accounted for 17.8% of global AI publications in 2024 and 20.6% of AI citations.

But the American comparison may illustrate why population alone is misleading.

The United States does not draw its innovation talent only from its own population of roughly 340 million people. It has historically been extraordinarily successful at attracting scientists, engineers, entrepreneurs and students from the rest of the world. In 2025, China and India alone accounted for 71% of international science-and-engineering master’s students and 44% of international S&E doctoral students in the United States. China itself supplied around 38,000 doctoral students.

Stanford’s 2026 AI Index still finds the United States home to more AI researchers and developers than any other country.

So China’s advantage is the enormous scale of its domestic talent pipeline. However, America’s counterweight is its ability to turn a much smaller domestic population into a considerably larger effective innovation population by attracting global talent. Whether this continues in the future remains to be seen.

That distinction will become important when we compare the two innovation machines.

The Resources, R, captures compute, capital, energy, data, semiconductor technology and access to sophisticated infrastructure and tools.

Here the United States has an enormous advantage—but not across every dimension.

On capital and frontier compute, the gap is striking. Stanford’s 2026 AI Index estimates US private AI investment at about USD286 billion in 2025, about 24 times China ca. USD12 billion, although Stanford explicitly cautions that private investment substantially understates China’s state-backed funding.

The United States also hosts 5,427 data centers, more than ten times any other country, and the global AI compute ecosystem remains dominated by American-designed accelerators. It should be noted that the number of data centers is obviously not the same thing as installed AI compute capacity.

China’s biggest weakness is therefore not a lack of money, engineers or electricity. It is access to the most advanced AI compute: leading GPUs, HBM memory and the semiconductor manufacturing technology required to produce them at scale.

But China has formidable resources elsewhere. Its electricity system alone supplied more than 9,500 TWh in 2025, more than twice the roughly 4,430 TWh generated in the United States, with China accounting for 58% of the increase in global electricity demand that year. China has an enormous industrial infrastructure, hyperscale domestic technology companies and a huge data-generating economy. Data itself is much harder to score cleanly. Both countries possess extraordinary amounts, but regulation, accessibility and the ability to combine datasets differ substantially.

So R is not simply “how many GPUs do you have?”.

The US currently has a very large advantage in the resources most important for frontier model training. China has a much larger resource base than the GPU comparison alone might suggest. And that distinction may matter increasingly over the longer term.

But innovation is also a search process. The more genuinely different approaches being explored, the larger the technological possibility space being searched. That is Diversity, D.

This is not simply the number of AI companies or models. What matters is whether different organizations have enough talent, resources and freedom to make genuinely different technological bets. The United States remains remarkably strong. Stanford’s 2026 AI Index counted 59 notable US models in 2025 compared with 35 from China, and the US continues to have a much larger startup and venture ecosystem.

But frontier AI is becoming increasingly expensive. Therefore also increasingly concentrated. OpenAI, Anthropic and xAI alone raised 38% of all AI funding in 2025, according to CB Insights. At the very frontier, much of the competition increasingly revolves around a relatively small group: OpenAI, Anthropic, Google, xAI and Meta.

China has fewer notable models overall, but its model ecosystem has become intensely competitive. DeepSeek competes with Alibaba’s Qwen, ByteDance, Moonshot, MiniMax, Zhipu, Tencent and others. Reuters describes low-cost, open-weight development as increasingly becoming the norm, with DeepSeek’s success prompting competitors to cut prices, open models and pursue different specialisations.

The comparison is obviously not straightforward!

America may have more AI experimentation in absolute terms. China may have more independent experimentation relative to the frontier resources available to it. Open weights potentially amplify that diversity by lowering the cost for yet more developers to try something different, which is the distinction of what D is intended to capture.

And discovering something useful is only part of the story. Other people need to learn from it. That is Knowledge diffusion, K: how quickly useful ideas, techniques and discoveries move through the ecosystem and become inputs into somebody else’s work.

This is where open-weight AI becomes particularly interesting.

The United States has an extraordinary knowledge-diffusion infrastructure. Its universities, GitHub ecosystem, research conferences, startups and movement of people between companies have historically allowed ideas to spread very quickly. Stanford’s 2026 AI Index counts 5.6 million open-source AI projects on GitHub, with US-based projects still attracting the highest level of engagement.

But there is an important tension at the frontier. Stanford also notes that the most capable American models are becoming the least transparent. OpenAI, Anthropic and Google increasingly disclose little about training code, datasets, model architecture or even basic training details.

China appears to be taking a somewhat different path. DeepSeek, Alibaba’s Qwen, Moonshot, Zhipu and others have made open-weight releases central to their competitive strategies. Those models can be downloaded, adapted, fine-tuned, compared and incorporated into other systems. Kai-Fu Lee’s “study group” analogy captures the potential effect nicely: one company’s innovation can become another company’s starting point rather than remaining entirely behind a proprietary wall.

And the diffusion is increasingly global, not merely Chinese. Reuters has reported growing international adoption of Chinese open-weight models because of their performance, low cost and ability to be customised. Stanford similarly concludes that open-source development is redistributing participation in AI beyond the traditional US technology centres.

The United States probably still has the stronger overall knowledge ecosystem. China may increasingly have the stronger diffusion mechanism around frontier open-weight models.

That interaction between D and K may turn out to be one of the most interesting differences between the two AI innovation systems.

Finally, good ideas must be recognized, funded, industrialized and scaled. That is Selection and Scaling, S.

The United States is probably the world’s strongest system for selecting and financing technological winners. In 2025, US private AI investment reached ca. USD286 billion, versus USD12+ billion in China, and 1,953 newly funded AI companies emerged in the US compared with only 161 in China. Once an idea demonstrates promise, America’s venture-capital markets, hyperscalers and enormous technology companies can put extraordinary amounts of capital behind it very quickly.

China’s strength is somewhat different. It has demonstrated an extraordinary ability to move technologies from engineering into large-scale industrial deployment. Stanford notes that China now leads the world in industrial robot installations, while recent developments in AI hardware, robotics and advanced manufacturing have made Chinese factories an increasingly important destination for foreign technology executives wanting to understand how quickly new technologies can be prototyped and industrialised.

Chinese AI companies are also becoming better at commercialization. Zhipu AI, for example, increased first-half 2026 revenue by around 400%, although Chinese foundation-model companies still trail OpenAI and Anthropic substantially in monetization.

So S may also need some nuance.

America is exceptionally good at identifying a promising technological company, funding it enormously and scaling it globally.

China is exceptionally good at turning technology into products, infrastructure and industrial capacity at very large scale.

And then we have the Constraints, C. I deliberately place constraints somewhat outside the other ingredients because constraints are not simply another resource where more is always better.

Too few constraints may encourage complacency. Moderate constraints can force efficiency, substitution and unconventional thinking. Severe constraints eventually become destructive.

Having fewer GPUs may encourage you to design a better algorithm. Having no GPUs probably does not.

The formula is therefore not intended to calculate whether China has a breakthrough probability of 63% and America 71%. That would be pseudo-precision.

It is a framework for asking a much more useful question:

How strong is an innovation ecosystem across the different ingredients that make breakthroughs possible? And, maybe more importantly, how do those ingredients interact?

The United States may have an overwhelming advantage in resources and scaling. China may compensate partly through a much larger effective engineering population, more independent experimentation, faster knowledge diffusion and stronger incentives to improve efficiency.

Which is precisely why counting GPUs alone may tell us much less about the future of AI innovation than many people assume.

BREAKTHROUGH POTENTIAL: USA VS CHINA.

If we apply the Breakthrough Formula qualitatively, the answer is less obvious than the difference in GPU inventories might suggest.

The United States has the clear advantage in resources (R). It has vastly more frontier compute, much deeper private capital markets, the leading semiconductor ecosystem and an extraordinary ability to concentrate money, infrastructure and global talent behind promising ideas. Its ability to select technological winners and scale them globally is probably unmatched.

China’s strengths sit elsewhere. Its domestic engineering and scientific talent pool is enormous (H·P). Competition between serious AI developers is intense. Open-weight models encourage rapid knowledge diffusion, while restrictions on advanced compute create unusually strong incentives to improve efficiency and explore alternatives.

In very simplified terms:

Factor	USA	China
Effective innovation population (H·P)	Exceptional, amplified by global talent.	Exceptional, driven by enormous domestic scale.
Resources (R)	Very strong advantage.	Large, but frontier-compute constrained.
Diversity (D)	Very high, although frontier AI is increasingly concentrated.	High and relatively distributed across serious competitors.
Knowledge diffusion (K)	Exceptional overall knowledge ecosystem.	Potentially stronger around open-weight frontier models.
Selection & Scaling (S)	Exceptional in funding and global commercial scaling.	Exceptional in industrialization and large-scale deployment.
Constraints (C)	Relatively limited compute constraint.	Strong pressure towards efficiency and substitution.

My assessment today would therefore still put the United States ahead in overall breakthrough potential. But probably not by anything approaching the margin implied by its advantage in compute.

But that distinction matters!

The American AI innovation machine is extraordinarily powerful because it can concentrate resources and scaling power behind a relatively small number of frontier bets.

The Chinese machine increasingly appears capable of compensating through scale of engineering talent, diversity of experimentation, rapid knowledge diffusion and constraint-driven efficiency.

We might reduce the difference almost provocatively to:

The USA is all about Resources (R) × Scaling (S).

While

China’s relative strength may lie more in Effective Innovation Population (H·P) × Diversity (D) × Knowledge Diffusion (K), under stronger constraints that reward efficiency.

Obviously, reality is considerably more complicated. Neither country owns one side of that equation exclusively. But the comparison highlights the central argument of this article:

The country with the most compute does not necessarily have a proportionate advantage in breakthrough potential.

And perhaps the most important part is that neither innovation machine is static.

• China can acquire more compute.
• America can adopt Chinese efficiency improvements.
• Open-weight innovations can cross borders almost instantly.
• Talent moves.
• Algorithms diffuse.
• Bottlenecks change.

THE VARIABLES DO NOT LIVE IN ISOLATION.

There is one important caveat to the Breakthrough Formula that is worth keeping in mind.

The variables H, P, R, D, K, S and C are not independent of one another.

More resources can attract more talent. More talent can increase diversity of experimentation. Greater diversity becomes more valuable when knowledge diffuses quickly. Stronger selection and scaling mechanisms can attract still more capital and researchers. And constraints can change how efficiently almost every other resource is used.

The innovation machine therefore contains feedback loops. The same is true of the United States and China. They are competitors, but they are not isolated innovation systems.

Chinese researchers build on American papers and architectures. American companies learn from Chinese open-weight models and efficiency techniques. Researchers move between universities and companies. Open models cross national borders almost instantly. Semiconductor restrictions may slow the movement of hardware, but they are considerably less effective at stopping ideas.

A breakthrough in one ecosystem can therefore increase the productivity of the other.

If Chinese developers discover how to achieve comparable model performance with dramatically less compute, American frontier laboratories can adopt the same techniques and combine them with their much larger compute resources. Likewise, a major American architectural breakthrough can rapidly become the starting point for Chinese experimentation.

So the US-China AI competition is not simply two independent machines racing towards the same finish line. They compete, copy, learn from and increasingly accelerate one another. That mutual dependence may turn out to be just as important as their differences.

THE MOST GPUs MAY NOT WIN.

So where does all this leave us?

Certainly not with the conclusion that compute does not matter. It matters enormously, and today the United States still has, in my assessment, the stronger overall AI innovation machine. Its combination of frontier compute, capital, semiconductor technology, global talent and ability to scale technological winners is formidable. But the Breakthrough Formula suggests that resources alone do not determine the likelihood of innovation. Human capital, participation, diversity of experimentation, knowledge diffusion, selection and scaling all matter as well. The US advantage in GPUs is therefore real, but it should not automatically be interpreted as an equally large, or permanent, advantage in breakthrough potential.

The second point is that the competition is dynamic. China’s restrictions on advanced compute may slow its progress, which is of course their intention. But they also create powerful incentives to use compute more efficiently, develop alternative architectures and reduce dependence on the very technologies being restricted. At the same time, the American and Chinese innovation systems do not operate independently. Chinese researchers build on American breakthroughs, while American companies can adopt efficiency improvements developed in China and combine them with much larger compute resources. Hardware can be restricted considerably more effectively than ideas. The two systems compete, copy, learn from and ultimately also strengthen one another.

And finally, I keep coming back to the obsession with intelligence. We spend enormous effort debating whether OpenAI, Anthropic, Google, DeepSeek or Qwen has the most intelligent model. For industrial AI, that may eventually become one of the less interesting questions. If capable foundation-model intelligence becomes cheap, widely available and increasingly open, economic value will migrate towards what remains scarce: proprietary data, workflows, customers, domain knowledge and the ability to integrate AI deeply into real businesses. A telecom operator, manufacturer or bank may not need the world’s most intelligent model. It may need one that is good enough, economically attractive, controllable and deeply embedded in how the company actually works.

There is, however, another side to this that deserves a paper of its own. If Chinese open-weight models increasingly become embedded inside telecommunications networks, factories, banks and critical infrastructure, questions of provenance, supply-chain integrity and trust will inevitably follow. We have spent years debating high-risk vendors in telecommunications. It would be rather ironic if we became extraordinarily careful about who supplies the radio equipment, while paying much less attention to the provenance of the AI models increasingly allowed to understand, recommend and eventually act upon the operation of the network. In the days of Odysseus, the Trojan horse was made of wood. In the age of AI, perhaps it arrives as a highly capable model downloaded for free. This is not an allegation that Chinese models contain hidden backdoors. It is an argument that model provenance deserves the same scrutiny we already apply to other components of critical infrastructure. In other words, applying AI models in an industry setting, requires the models to be fully open, irrespective of the origin of the model.

The United States has built perhaps the world’s most powerful machine for concentrating talent, capital and compute behind technological winners. China appears to be building something different: an enormous engineering base, intense competition, strong incentives for efficiency and an increasingly important open-weight ecosystem through which knowledge can diffuse quickly.

Which machine ultimately proves more innovative remains an open question.

But I would be very careful assuming that the answer can be found simply by counting GPUs. I believe such an approach would be a mistake.

REFERENCES.

Open Source Initiative, “The Open Source AI Definition – 1.0,” Open Source Initiative, 2024. Open Source Initiative definition of open-source AI.

Hugging Face, “Hugging Face Hub Documentation,” Hugging Face, 2026. Hugging Face documentation describing its model, dataset and application platform.

Jensen Huang, “NVIDIA to Acquire Hugging Face,” NVIDIA, 3 September 2026. NVIDIA announcement of its proposed acquisition of Hugging Face.

NVIDIA Corporation, “Current Report on Form 8-K,” U.S. Securities and Exchange Commission, 2 September 2026. NVIDIA regulatory filing on the proposed Hugging Face acquisition.

Eleanor Olcott, “01.ai’s Kai-Fu Lee: Why China will beat the US in consumer AI,” Financial Times, 4 February 2026. Financial Times interview with Kai-Fu Lee.

Kai-Fu Lee, “AI Superpowers: China, Silicon Valley, and the New World Order,” Houghton Mifflin Harcourt, 25 September 2018. Kai-Fu Lee’s book on the shift from the age of discovery to the age of implementation and China’s implementation advantages.

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.

Emily Chang, “Inside the Mind of Anthropic CEO Dario Amodei,” The Circuit, Bloomberg Originals, 17 June 2026. Bloomberg Originals interview with Dario Amodei.

Zachary Basu and Mike Allen, “Nvidia’s Jensen Huang defends Chinese AI amid Kimi panic,” Axios, 22 July 2026. Axios interview with Jensen Huang.

Blake Alcott, “Jevons’ Paradox,” Ecological Economics, July 2005. Ecological Economics review of the rebound effect, whereby efficiency gains increase rather than reduce total consumption.

Jeff John Roberts, “What is Jevons paradox? The reason Satya Nadella says DeepSeek’s new AI is good news for tech,” Fortune, 27 January 2025. Fortune report on Nadella’s Jevons paradox comment and the DeepSeek market reaction.

Julia Boorstin, CNBC Exclusive Interview with OpenAI CEO Sam Altman, Squawk on the Street, 9 July 2026. CNBC interview transcript with Sam Altman.

Paul M. Romer, “Endogenous Technological Change,” Journal of Political Economy, 1 October 1990. Journal of Political Economy paper on endogenous technological change.

Nicholas Bloom, Charles I. Jones, John Van Reenen and Michael Webb, “Are Ideas Getting Harder to Find?” American Economic Review, April 2020. American Economic Review paper on research productivity.

Richard Windsor, “China vs. USA – Chips for Free?” Radio Free Mobile, 31 August 2026. Radio Free Mobile analysis of Huawei and US AI hardware.

Anthropic, “2028: Two scenarios for global AI leadership,” Anthropic, 14 May 2026. Anthropic analysis of US-China AI leadership and compute.

Anson Ho, Tamay Besiroglu, Ege Erdil, David Owen, Robi Rahman, Zifan Carl Guo, David Atkinson, Neil Thompson and Jaime Sevilla, “Algorithmic Progress in Language Models,” Epoch AI / arXiv, 9 March 2024. Epoch AI estimate that the compute required to reach a given language-model capability has halved roughly every eight months since 2012. Companion summary: Epoch AI blog post on algorithmic progress in language models.

Georgia Adamson, Saif Khan, Tao Burga and Tim Fist, “Should the US Sell Blackwell Chips to China? Assessing the impacts of exporting the B30A AI chip,” Institute for Progress, 25 October 2025. Institute for Progress analysis of the US-China AI compute advantage.

Gregory C. Allen, “Mismatch of Strategy and Budgets in AI Chip Export Controls,” Center for Strategic and International Studies, 29 October 2024. CSIS analysis of export controls and a durable US technology lead.

Stanford Institute for Human-Centered Artificial Intelligence, “Technical Performance,” The 2026 AI Index Report, 2026. Stanford AI Index analysis of US and Chinese model performance.

Philippe Aghion, Nicholas Bloom, Richard Blundell, Rachel Griffith and Peter Howitt, “Competition and Innovation: An Inverted-U Relationship,” The Quarterly Journal of Economics, May 2005. Quarterly Journal of Economics paper on competition and innovation.

Brian Uzzi, Satyam Mukherjee, Michael Stringer and Benjamin Jones, “Atypical Combinations and Scientific Impact,” Science, 25 October 2013. Science paper on knowledge recombination and scientific impact.

Wesley M. Cohen and Daniel A. Levinthal, “Absorptive Capacity: A New Perspective on Learning and Innovation,” Administrative Science Quarterly, March 1990. Administrative Science Quarterly paper on absorptive capacity.

Jeffrey L. Furman, Michael E. Porter and Scott Stern, “The Determinants of National Innovative Capacity,” Research Policy, August 2002. Research Policy paper on national innovative capacity.

Daron Acemoglu, “Directed Technical Change,” The Review of Economic Studies, October 2002. Review of Economic Studies paper on how relative factor scarcity redirects innovation toward technologies that save on the scarce input.

John R. Hicks, “The Theory of Wages,” Macmillan, 1932. Original statement of the induced-innovation hypothesis: a rise in the price of a factor of production spurs inventions economising on that factor.

Marcus Matthias Keupp and Oliver Gassmann, “Resource Constraints as Triggers of Radical Innovation: Longitudinal Evidence from the Manufacturing Sector,” Research Policy, September 2013. Research Policy paper on resource constraints and radical innovation.

Oguz A. Acar, Murat Tarakci and Daan van Knippenberg, “Creativity and Innovation Under Constraints: A Cross-Disciplinary Integrative Review,” Journal of Management, January 2019. Journal of Management review of constraints, creativity and innovation.

National Science Board, “The State of U.S. Science and Engineering 2026: STEM Talent, Education, Training, and Workforce,” National Science Foundation, 2026. National Science Foundation analysis of US and Chinese science and engineering talent.

Stanford Institute for Human-Centered Artificial Intelligence, “Research and Development,” The 2026 AI Index Report, 2026. Stanford AI Index analysis of notable AI model development in the United States and China.

CB Insights, “State of AI 2025 Report,” CB Insights, 2026. CB Insights analysis of concentration in AI funding and frontier model development.

Jeffrey Ding, “Technology and the Rise of Great Powers: How Diffusion Shapes Economic Competition,” Princeton University Press, 2024. Princeton University Press book arguing that great-power advantage in general-purpose technologies depends on diffusion capacity rather than frontier innovation.

Eduardo Baptista and Che Pan, “A year on from DeepSeek shock, get set for flurry of low-cost Chinese AI models,” Reuters, 12 February 2026. Reuters analysis of competition and open-weight model development in China.

Reuters, “China’s Zhipu AI revenue quintuples in first half, loss narrows,” Reuters, 31 August 2026. Reuters analysis of competition among Chinese AI model developers.

James Landay, “Open-Weight Models Aren’t Enough. We Need Truly Open Source AI Models for Science and Society,” Stanford Institute for Human-Centered Artificial Intelligence, 4 August 2026. Stanford HAI discussion of open-weight models, transparency and reproducibility.

Reuters, “China weighs silicon curtain around sought-after AI models,” Reuters, 8 July 2026. Reuters analysis of the international importance of Chinese open-weight AI models.

Reuters, “A new, inexpensive Chinese AI model is catching up with Anthropic, OpenAI on their home turf,” Reuters, 2 July 2026. Reuters analysis of global adoption of Chinese open-weight AI models.

Stanford Institute for Human-Centered Artificial Intelligence, “Economy,” The 2026 AI Index Report, 2026. Stanford AI Index analysis of AI investment and entrepreneurial activity.

International Energy Agency, “Electricity 2026,” IEA, 2026. IEA analysis of global and Chinese electricity demand.

Stanford Institute for Human-Centered Artificial Intelligence, “The 2026 AI Index Report,” Stanford University, 2026. Stanford AI Index comparison of US and Chinese AI development and industrial deployment.

Reuters, “Tech pilgrims flock to China as global innovation race heats up,” Reuters, 3 September 2026. Reuters analysis of China’s rapid prototyping and industrial technology ecosystem.