In the late 1950s, the United States was gravely concerned that a missile gap was emerging between itself and the Soviet Union. The conventional wisdom held that the Soviet Union was poised to overtake the United States in missile production, nullifying the United States’ significant nuclear edge. U.S. officials eventually learned that the gap did not exist; compared with the United States, the Soviet Union in fact had a paltry missile arsenal. Yet in the interim, the missile gap was a convenient belief to hold: the putative gap justified an arms race with the Soviet Union—one desired by military brass and an engorged military-industrial complex, but that soon proved terrifyingly risky.
The United States may repeat that mistake in its assessment of China’s artificial intelligence (AI) capabilities. As the country awakens to the risks of AI, a debate rages over whether to pause or pace the technology’s development—that is, to halt or slow the rate of progress in AI to develop better mechanisms to control it. “Pause AI development,” Senator Bernie Sanders (I-VT) demanded of the chief executives of eight AI companies this summer. “It is not too late to avoid disaster.” On the other end of the political spectrum, tech billionaire Elon Musk endorsed Anthropic chief executive Dario Amodei’s call to pace frontier AI development in three words: “Dario is right,” he wrote on X in September.
But others disagree with this view, usually arguing that American AI development should not be slowed because of China. The thinking is that if the United States paces its AI development, China’s own AI sector will surge ahead and unacceptable consequences will follow. Therefore, the United States, should continue to move quickly. People making this argument include President Donald Trump—who insists that “whoever wins AI, wins”—and House Speaker Mike Johnson, who cautions that with hasty regulation “we will lose the race to China,” and Nvidia chief executive Jensen Huang, who has flatly declared that “China is going to win the AI race.”
The China argument has taken on decisive significance. But the argument is flawed in ways reminiscent of the flaws animating Washington’s feverish perception of a missile gap during the Cold War. It both overstates China’s capabilities and misreads the actual risks that would be incurred in the unlikely eventuality that China does pull ahead.
China Can Pace, Not Surpass
Yes, Chinese companies have trained AI models that observers describe as being mere months behind those of the United States. But the evidence points toward China at best being able to pace, not surpass, the United States.
A nation’s AI power can be divided into three pillars: computing power, algorithms, and data. Across them, the resources available to the United States and its allies outpace those that are available to China.
Take computing power, which can be used to train or service AI models—where the United States enjoys a clear lead. The United States has at its disposal computer chips that are far superior to those of China and possesses in aggregate roughly ten times as much computing power as China does. That advantage alone could prove decisive. It caps the resources that China has available to train or serve its own AI models. And it even stymies China’s ability to leverage its undeniably potent energy-generation capacity to bolster its AI sector, as its own chips are up to two and a half times less energy efficient.
That advantage should prove enduring. Barring a Chinese takeover of Taiwan, the United States appears likely to outpace China in its ability to build computing power: my coauthored book shows that the United States and its allies—who are deeply enmeshed in U.S. technology supply chains—together dominate the high-technology ecosystem required to build this power; their companies generate 84 percent of global high-technology profits and generate 64 percent of global value-added in high-technology sectors. The gargantuan financial system of the United States has meanwhile deployed investments toward the AI buildout that for now greatly outmatch those of Beijing. And even if grassroots opposition impedes data center construction in the United States, that will not stop U.S. companies from building data centers outside the country.
China does possess both the talented AI researchers to create training algorithms and the data on which to train models. But the United States competes with or outcompetes China on both. The United States employs 42 percent of the world’s top AI researchers to China’s 28 percent—even though China now produces nearly half of them—and has a continued knack for attracting and keeping China’s most talented researchers in Silicon Valley. The one pillar on which China may have an advantage is data, where Beijing could compel the release of vast quantities of data to its own AI labs. Yet so far, that data has not offered China a lead over the United States. And as American AI labs embed their models across more industries and perfect their training algorithms, the amount of data they are exposed to—and can train off—will naturally and quickly grow.
Surfing in America’s Wake
Why then have Chinese models performed so well? It helps that China is second only to the United States in all three pillars of AI power. But the more important reason is that China’s AI industry has skillfully surfed in the wake of the American AI ecosystem. Evidence from the U.S. government, the AI labs, and external observers all point toward Chinese labs having trained their own models off the intelligence of the models of the United States through a process known as distillation, where a less sophisticated model is improved by training off the outputs of a more sophisticated model. However, China’s surfing has only been sufficient for it to, at best, closely trail the United States: China has never held the AI frontier—since 2023, Chinese models have trailed the best American models by roughly seven months on average, and China’s surfing strategy means that it is likely that U.S. pacing will slow China as well.
The Race Is Not the Risk
Even in the unlikely event that China does temporarily surpass the United States, the consequences of that alone may not be as catastrophic as many suggest. AI on its own is powerful, but it has not yet been shown that having a slightly better AI model grants decisive economic or military advantages—for instance, a well-deployed weak AI model could be more effective than a poorly-deployed strong AI model. The truly grave concern is that AI models may soon be able to improve themselves—a process that Amodei says “is starting to happen across the industry”—which could in the next decade ignite an intelligence explosion and grant substantial powers to whoever is possessing them. But a recursive self-improvement loop of this nature could just as quickly result in either the United States or China losing control of the models they develop, transcending their geopolitical competition and threatening both countries. It is this risk that has sharpened discussions of AI pacing in the United States. In short, the version of China’s AI takeoff that is most likely to harm the United States is also the version that is most likely to harm China. If China approaches that threshold, it is therefore plausible that it too may choose to pace; if it doesn’t, it may enjoy substantial advantages, but it may also lose control of the technology to its own detriment.
During the Cold War, the fear engendered by the missile gap prompted the U.S. to escalate the nuclear race with the Soviet Union, and the two sides were soon brought dangerously close to the precipice; today, a similar fear about China could accelerate the U.S.-China AI race. To be sure, the stakes of that race may not be as high as those of the nuclear race during the Cold War, and indeed a full AI pause may not be wise. But arguments in the United States against a pause should not fall into the same analytical trap of the Cold War: competing to the point of blindness, underestimating its own strengths and a rival’s weaknesses, charging toward the realization of a promethean technology whose risks it does not yet appreciate. That is a cycle that history has taught us to be wary of. We should avoid it.
