Why America Wins

In a word: compute

In our scenario, China lags the US by 3-6 months through 2028. They try to catch up by stealing American algorithmic secrets and finished model weights, but are never able to fully close the gap.

Many people thought this underestimated Chinese ability to compete:

China has amazing AI talent, a big infrastructure advantage, and top-notch companies. We were impressed by DeepSeek like everyone else. Nothing in our scenario should be interpreted as denigrating their accomplishments.

But in the end, it all comes down to compute.

The world’s best chips come from Taiwan Semiconductor Manufacturing Corporation. TSMC depends on equipment produced by America and its allies, putting it within the US sphere of influence.

Starting in 2022, America banned TSMC from selling advanced chips to China. The restrictions were weak, and China was able to use a combination of legal loopholes and smuggling to get a substantial number of chips regardless. But even the current level of export controls limit China’s compute, and once the White House starts to take the possibility of an AI arms race more seriously, they can tighten restrictions further.

As of 2024, we estimate that the United States had about 75% of advanced chips suitable for AI development. China had about 15%, and the rest of the world combined had 10%.¹ We expect these numbers to stay about the same through 2027, even accounting for China having about the same amount of success smuggling in chips as they do now, and ramping up domestic manufacturing.

Current export controls, despite their flaws, do impose a cost on Chinese AI companies. They don’t have easy access to the most cost effective, frontier chips. We tentatively estimate that compute for Chinese companies is around 60% more expensive in China than a counterfactual without sanctions.

In the future, if the US tightens the loopholes, and China succeeds in ramping up domestic manufacturing, their cost efficiency should decrease further.² When you combine this with the fact that public reports have Chinese companies spending around 4 times less than US companies on AI chips in 2025, we find that the gap between the US and China is on track to remain the same size.

If America has five times the compute, why do we think China will “only” be 3-6 months behind?

Both American and Chinese chips are currently divided among many different AI companies and non-AI applications. Google, probably the most compute-rich US company, only has about 15% of the world’s advanced compute (and doesn’t even use it all for AI); OpenAI only has 5-10%. In China, DeepSeek barely has 1%. AI prowess depends on the amount of compute concentrated in a single project. If China does a better job concentrating their compute than the US, they can establish a lead, or at least catch up.

In our scenario, we have China begin to concentrate their compute in 2026 and aggregate 10% of their 15% compute share in a single national AI effort; US compute remains scattered. We think companies are more likely to concentrate than disperse, so the 2026-2027 US leader may have more like 15-20% rather than the current 10-15%.

Since the Chinese project will use 10% of world compute, and the leading American project will use 15-20% of world compute, we forecast America having a modest compute lead over China.

This is despite some conservative assumptions:

• That the US continues to do a poor job enforcing chip sanctions.

• That China centralizes almost all of its compute, and the US barely centralizes at all.

• That Chinese compute centralization happens quickly - starting in 2026 and being near-complete by 2027.

If we relax any of these assumptions, America’s compute lead widens further.

Can China Make Its Own Compute?

Not quickly enough to matter by 2027.

China’s biggest chip company, Huawei, is able to make chips almost as good as TSMC/NVIDIA’s H100s. But these chips require advanced wafers that they currently source from TSMC. As chip sanctions tighten, Huawei’s supply of these wafers will be limited to what they can smuggle in. We include these TSMC-Huawei chips in our (low) estimate of Chinese compute.

Another Chinese company, SMIC, is trying to make homegrown wafers to reduce dependence on sanctioned TSMC components. So far, they’re still far behind the cutting edge: their process has lower yields and gives worse performance. Catching up fully is a daunting task. TSMC’s advantage relies on extreme UV lithography, an incredibly complex machine that only a single company in the Netherlands (ASML) has ever managed to manufacture. This is set to take many years.

We forecast a crucial period for the intelligence explosion in 2027 - 2028. During this period, SMIC will still be playing catchup, so chips will still be expensive to make and behind the frontier. China will likely still rely on some smuggling of chips and wafers to keep pace.

If America continues to do a mediocre job enforcing sanctions, we think China may be able to make significant amounts of homegrown chips in the early 2030s. If America starts robustly enforcing sanctions, and extending them to key equipment and components, that could delay Chinese chip independence until the late 2030s.

If we’re overestimating the speed of AI progress and the intelligence explosion doesn’t happen until the 2030s, then we agree China will be in a strong position.

What About Energy?

In our scenario, energy is not a big obstacle by 2027 - 2028.

China has a big energy advantage, but luckily for the US, the need for energy is downstream of the need for AI chips, and the US is likely to be able to pull together enough capacity to avoid bottlenecks.

Like in chip manufacturing, this story is only as strong as our timelines; a 2030s intelligence explosion is more favorable to China.

What About Talent?

Other commenters argue that we ignore human talent.

It’s a fair concern - does compute alone really determine progress?

Models of AI progress usually divide it into two bins: compute scaling and algorithmic improvement. With enough algorithmic improvement - technological advances in how to build AIs efficiently - you can make up for a compute shortfall.

China has 1.4 billion people. Might this give them enough research talent to speed up algorithmic progress and compensate for US compute advantages?

We think not, for three reasons:

• Algorithmic secrets are leaky, so it’s hard for there to be a big US-China gap.

• Insofar as there is a gap, the US is probably ahead.

• Algorithmic advances are partly bottlenecked by compute, so talent is less valuable than it seems.

Going through each in turn:

Algorithmic secrets are leaky: Everyone is constantly stealing everyone else’s algorithmic advances. America steals China’s. China steals America’s. OpenAI steals Anthropic’s. Anthropic steals OpenAI’s. This is why all the big companies in both countries are within a year or so of each other.

This isn’t necessarily cloak-and-dagger-style espionage. You can learn a lot about an AI just by talking to it, or reading the model card, or reading the papers that get published about it. And companies are constantly luring their competitors’ top researchers away with offers of higher salaries, then debriefing them for technical secrets.

Sufficiently motivated countries could crack down on this; indeed, our scenario has both America and China rapidly scaling up security over the next few years. But make the crackdown too intense, and it will slow progress - for example, isolate researchers on a secure military base and ban them from using phones, and many of them will quit. We don’t think this factor will entirely go away by the crucial 2027-28 period.

Insofar as there is a gap, the US is probably ahead. DeepSeek’s R1 was very impressive. While not exactly ahead of the top American models on absolute performance, its performance given its price was remarkable. This raised concerns that China was pushing ahead of the US on algorithmic technology.

But later developments somewhat alleviated these concerns. Although early reports claimed DeepSeek was trained on an incredible $6 million budget, later analysis suggested the real cost was probably closer to $1.6 billion. Rather than succeeding with only a tiny number of chips, DeepSeek succeeded because they bought a large number of chips just before the chip sanctions kicked in. Now that the chip sanctions are in place (however lossily), such successes will be harder to come by.

Head-to-head comparisons of R1 vs. similarly-timed Western AIs revealed that the latter were probably further along the price-performance frontier.

(source)

So while DeepSeek is an extraordinary achievement in the context of China’s earlier-stage AI ecosystem and more limited resources, we don’t think it reflects an absolute advantage (or even an absolute algorithmic advantage) over the United States.

Talent is less valuable than it seems. There’s an argument that, whatever the current situation, we should eventually expect Chinese talent to dominate. China’s population is 4x larger than America’s, and it graduates twice as many STEM PhDs. That’s a lot of smart people who could potentially go into AI.

On the other hand, America draws on a pool of talented immigrants from all over the world (for example, leading US AI researcher Ilya Sutksever was born in Russia, grew up in Israel, and studied in Canada). Adjusting for these factors, we’re not sure who wins here, or by how much.

Still, let’s say for the sake of argument that China eventually gets a 2-4x talent advantage. Is this enough to dominate America’s compute advantage and win the race?

We’ve thought about this question a lot, because it determines the speed of the intelligence explosion. Once AIs begin to contribute to AI R&D, “talent” increases by orders of magnitude, but compute stays constant. What happens? Compute starts bottlenecking research - not just the size of large training runs, but the ability of algorithmic progress researchers to test their new ideas.

How tight is the bottleneck? You can read our Takeoff Forecast for the details, but we model a situation where near-superintelligent AIs increase the size of the “talent” pool 1000x, and where each of these AI “employees” is as productive as the best human AI researchers (think Alec Radford or Ilya Sutskever). We find that even this extreme scenario only speeds progress by 25x. On this scale, China’s 2-4x talent advantage barely even registers. Compute constraints are a harsh master. And AI company budgeting implicitly agrees with this estimate: OpenAI spends 6x as much on compute as on labor, suggesting they find the former more important.

For all these reasons, we don’t think China has a talent advantage that can compensate for America’s compute advantage.

Sounds Like Chip Sanctions Are Pretty Important, Huh?

We are most concerned about misalignment risk, and would be happiest if America and China stopped racing and agreed to some sort of international framework for developing AI responsibly.

But for people more interested in the US-China arms race, we can’t think of anything more important than enforcing chip sanctions. We think this is the difference between America having a slight lead over China in 2027 that diminishes to near zero in 2030, vs. America having a strong lead over China in 2027 and maintaining it to 2035 and beyond.

In a world like our scenario, where near-term AI progress determines the future, small differences in chip sanction enforcement could mean the difference between total American domination of the future vs. total Chinese domination of the future. Even if you disagree with us about the importance of AI, advanced chips could be used to develop narrow AI systems for military applications. We think if the US government understood the stakes, they would make chip sanction enforcement their number one priority - higher than Ukraine, higher than Israel, higher than immigration, higher even than national defense.

Instead, the Bureau of Industry and Security, the agency charged with enforcing chip sanctions, is underfunded, with a budget of only $200 million, and China can smuggle in enough advanced chips to stay in the game. It would take only a few hundred million dollars to decisively win the AI race - far less than the billions being invested in Stargate, power plants, et cetera.³

If the administration wants to be friendly to China, they should use that goodwill to negotiate a bilateral treaty regulating AI development. If they want to play hardball, they should actually play hardball.

Summary

Almost all advanced chips come from Taiwan, within the US sphere of influence. The US uses export controls to try to keep these chips from China. The export controls are poorly enforced - but even so, America has 5x the AI-relevant compute of China. We expect this ratio to continue at least until the crucial 2027 - 2028 period. This gives top American companies an advantage over their Chinese competitors, even if China does a better job consolidating its compute.

China is trying to make its own advanced chips, but this is a long process, and they’re not on track to fully internalize the supply chain until 2030 at the earliest. Robust enforcement of sanctions on chip-making supply chain components could delay this until 2035 - 2040.

Although Chinese researchers are talented, we’re not sure they’re more talented than the global talent pool America has access to. Even if they were, compute limitations will probably remain decisive for plausible US-China talent gaps.

If the US wants to ensure it has a comfortable lead in AI, it should tighten chip sanctions, widening its lead during the 2027 - 2028 period and prolonging it until the late 2030s.

1. Recent work tracking over 500 AI supercomputers globally agreed with this geographical breakdown.

2. The Huawei AI CloudMatrix 384, reported on recently by Semianalysis, shows that the latest efforts by Huawei to match NVIDIA's frontier server needs to use around 2x the wafer area and 2x the memory area to get the same compute and memory bandwidth performance. Despite using TSMC (Taiwanese) wafers and Samsung (South Korean) memory. This might translate to around a 2x manufacturing cost today, but if they become reliant on domestic wafers (SMIC) and domestic memory (CXMT), this cost inefficiency might increase further.

3. CNAS recommend $57 million for improved export control enforcement, including $12 million for an AI chip registry and random sampling program, and $45 million for modernizing BIS’s enforcement capabilities in line with CSIS recommendation. We find it mind-boggling that the long-term global balance of power might hinge on America’s unwillingness to pay $57 million, and have generously rounded this up to “a few hundred million” to preserve our own sanity.