China, the United States, and the AI Race

Sign up to receive CFR President Mike Froman’s analysis on the most important foreign policy story of the week, delivered to your inbox every Friday afternoon. Subscribe to The World This Week.

China might have become the manufacturing floor for the global economy, but the West has taken some comfort from the assessment that the United States retains the lead when it comes to the quest for artificial intelligence (AI). That might depend, however, on how one defines the competition.

The United States tends to define it in terms of the race toward Artificial General Intelligence (AGI), that is, self-improving artificial intelligence which surpasses the cognitive power of human beings and is capable of executing real-world knowledge work tasks. By Trump’s AI czar David Sacks’ estimate, “China is not years and years behind us in AI. Maybe they’re three to six months,” but no one can really be certain—what that means, whether that’s true, and whether it really matters.

For one, how will we know when AGI has been achieved? The point at which AI crosses over into AGI, as models of ever-increasing sophistication are developed, might not be obvious.

Second, does it matter? If both the United States and China are going to achieve AGI, maybe as little as six months apart, does it matter who gets there first—other than to feed the vanity of the entrepreneur who achieves that milestone? What is going to happen in that period to position the winning country in a categorically different situation than otherwise might be the case?

And that leads to the third question: Are we, the United States, racing toward the wrong finish line? Indeed, we assume we are in a neck-to-neck race with China, but they might be racing on an entirely different course.

If the measure of success is building the biggest, most beautiful model, the United States is doing quite well. As U.S. firms invest hundreds of billions of dollars into the latest models, chips, and AI infrastructure, I was comforted to read the National Institute for Standards and Technology’s new AI benchmarking report, which found that the best U.S. model outperformed the best Chinese model, DeepSeek V3.1, across almost every benchmark, including by a 20 percent margin in software engineering tasks, a 35 percent margin in general operating costs, and an order of magnitude in various cybersecurity screenings.

But chatbots might not be the be-all and end-all when it comes to thinking machines and the strategic competition between the United States and China. There is a growing—if self-serving—argument among China’s leading technologists, officials, and researchers that the large language models captivating Silicon Valley do not represent the most strategic path to an AI-enabled future. Or, to paraphrase some Chinese experts on Weibo, “ChatGPT outputs are capitalist drivel.”

While China is certainly working to improve its large language models too, it is pursuing a somewhat different strategy—by choice or by necessity. China is less focused on large frontier models, such as ChatGPT-5, and more focused instead on wiring intelligence into the physical economy at scale. As Fareed Zakaria and Eric Schmidt have noted, Chinese President Xi Jinping often frames AI as “application-oriented,” and Beijing’s policies and procurement decisions reflect this vision—such as “city brain” pilots in Wuhan that fuse traffic cameras, internet-of-things (IoT) sensors, and other devices with autonomous vehicles.

The real action is in the manufacturing domain, where China is surging ahead in “embodied AI.” China operated roughly 2 million industrial robots in 2025 and installed about 295,000 more in 2024—more than the rest of the world combined—with a majority now made domestically in China. By contrast, U.S. factories installed about 34,000. These robots will all be powered or augmented by smaller-scale Chinese AI applications that don’t require the immense training compute or inference infrastructure of increasingly powerful Western chatbots.

China’s Ministry of Industry and Information Technology estimates that by the end of 2025, over 60 percent of large Chinese manufacturers will have adopted some form of “AI + Manufacturing” integration, and thousands of “AI-empowered” factories have already been certified nationwide. The country’s 14th Five-Year Plan calls for “comprehensive intelligent transformation” of industrial production, with AI embedded across 70 percent of key sectors by 2027, 90 percent by 2030, and 100 percent by 2035. This diffusion is already measurable on the ground: Nearly half of all new Chinese manufacturing equipment sold last year incorporated machine vision, predictive maintenance, or autonomous-control functions—evidence that AI is no longer confined to pilot projects but is becoming a default layer of the industrial economy.

The United States, obviously, has no such plan or benchmarks, but it is not hard to imagine armies of entrepreneurs across the United States developing new applications to deploy across the economy as AI advances. The United States is wagering on hundreds of billions of dollars of compute, hyperscale superclusters, and ever-larger language models in pursuit of AGI—systems so capable and creative that they might unleash an epoch of explosive economic growth and scientific discovery. But that is quite different from China’s approach of having a plan, backed up by a set of incentives and sanctions, to ensure the rapid diffusion and integration of AI across the whole of the industrial sector.

As Charlie Munger of Berkshire Hathaway once said, “Show me the incentive, and I’ll show you the outcome.” In the United States, AI has probably taken the form of consumer apps and enterprise software because that’s where the incentives—that is, the near-term profits—lie. China’s approach, by contrast, revolves around smaller-scale AI applications as an input to production rather than a product itself.

The strategic question is whether, over time, intelligence diffused through the physical economy proves more transformative than the wisdom of a future “ChatGPT-15.” There are certainly merits to both strategies, and we shouldn’t dismiss the transformative potential of large language models, particularly to advance basic research and scientific innovation, in addition to knowledge work. But we should probably be less confident and less complacent about the significance of our lead in model development.

The differences in AI strategy between the United States and China are also reflected in each country’s policy response, including their approach to export controls. Yesterday, China’s Ministry of Commerce turned the United States’ “small yard, high fence” doctrine on its head and unveiled what can only be described as a “big square, great wall” policy. The new measures go far beyond semiconductors, imposing export-license requirements on nearly the entire suite of rare earths and critical minerals which sit at the foundation of modern computing—from gallium (used in high-frequency chipsets and 5G amplifiers) and germanium (vital for infrared sensors and high-efficiency semiconductors) to graphite (the anode material in nearly all lithium-ion batteries), tungsten (used in chip etching and heat sinks), and various rare-earth magnets (critical for precision motors in chip fabrication equipment).

The asymmetry is striking. Washington has spent years weaponizing the leading edge of the AI technology stack—restricting China’s access to the most advanced GPUs which could facilitate China’s competition in our race toward AGI—while Beijing has consolidated control over the inputs—including critical minerals—which could help the United States compete in their race toward industrial dominance.

According to the International Energy Agency and the U.S. Geological Survey, China has a dominant global market position in nickel, cobalt, graphite, gallium, and germanium refining—materials essential to advanced chipmaking, sensors, and batteries. So too in manufacturing specialized, synthetic industrial diamonds, and processing heavy rare earths–areas in which they also own and control the intellectual property for processing techniques and equipment  In other words, the United States can cut China off from the chips of today, but China can make it vastly harder to build the chips and other advanced technologies of tomorrow.

The implications extend beyond semiconductors. As my colleague Rush Doshi noted, “This is basically like the United States’ Foreign Direct Product Rule, but with Chinese characteristics.” If China enforces these controls, even selectively, it could send shockwaves through the global supply chain for advanced computing, electric vehicles, and renewable energy systems—industries that depend on the same materials portfolio.

It also reframes the AI competition itself.  For all of America’s prowess in software and design, the uncomfortable truth is that in a world where China can twist the spigot on raw materials essential to chipmaking, China’s applied, industrial AI strategy—anchored in manufacturing and resource security—cannot be dismissed.

It is noteworthy that China has taken these escalatory actions on the eve of Trump’s meeting with Xi on the margins of the upcoming APEC summit in South Korea. Is it to seriously damage or retard the United States’ prospects for winning the AI race? Is it to have something to trade away in exchange for meaningful concessions from the United States? Or is it to show that China has learned from the Trump administration that, like tariffs, export controls can confer real leverage in negotiations over unrelated issues, such as Taiwan.

Just hours ago, in reaction to China’s export control announcement, Trump threatened to cancel his upcoming meeting with Xi and impose new tariffs, posting on Truth Social, “For every Element that they have been able to monopolize, we have two. I never thought it would come to this but perhaps, as with all things, the time has come.”

When Trump first imposed tariffs on China, many of us noted that China might retaliate by exploiting the chokepoints over certain products it controlled. Now we’re competing in the use of economic leverage as well. Time will tell who, if either, will prove “the winner” of that competition.

Let me know what you think about the future of the AI race and what this column should cover next by emailing [email protected]