I increasingly think the United States winning the AI race is far from guaranteed, because the competition is no longer simply about who produces the most powerful model. The U.S. currently has major advantages in frontier models, advanced semiconductors, private capital, and its ability to attract global talent. But China has a different set of advantages that may become even more important as AI matures: enormous electricity generation and grid expansion, cheaper infrastructure, batteries, manufacturing capacity, rapidly improving domestic chips, a massive STEM pipeline, and an increasingly competitive AI ecosystem. The important question therefore isn’t simply who builds the smartest model; it’s who can produce and deploy useful intelligence at the lowest cost and at the greatest scale.
Talent is particularly important here. One of America’s greatest historical advantages has been its ability to attract extraordinarily talented scientists, engineers, researchers, and entrepreneurs from around the world—including large numbers from China, India, and the rest of Asia—and integrate them into American universities and companies. If immigration restrictions, visa uncertainty, or an increasingly hostile political environment weaken that advantage, the U.S. risks damaging one of the very mechanisms that made Silicon Valley dominant. China doesn’t even need every Chinese researcher currently in America to return home. If future generations of top Chinese graduates increasingly decide that they can build equally prestigious and lucrative careers in Beijing, Shanghai, Shenzhen, or Hangzhou instead of moving to the United States, America’s talent advantage gradually erodes while China’s compounds.
Then there is infrastructure. AI ultimately requires chips, data centers, electricity, cooling, networks, and enormous amounts of capital. The U.S. and its allies still possess a major advantage in advanced semiconductors, and that should not be underestimated. But China has extraordinary advantages in electricity generation, grid infrastructure, batteries, manufacturing, construction capacity, and the speed at which physical infrastructure can be deployed. If Chinese companies can eventually produce models that are only marginally behind the American frontier while operating them much more cheaply, China doesn’t actually need to “win” the benchmark race. A Chinese model that delivers 95% of the useful capability of the best American model at 20% of the cost could be far more attractive to much of the world than a technically superior but substantially more expensive American alternative. At that point, intelligence-per-dollar becomes more important than simply having the highest benchmark score.
This becomes even more significant when geopolitics enters the picture. The United States has historically benefited enormously from its network of allies and partners, but if Washington increasingly treats allies and developing countries through tariffs, threats, restrictions, and transactional pressure, it creates an opening for China. Beijing could approach countries in Southeast Asia, South Asia, Africa, Latin America, and the Middle East with a very different proposition: Chinese models, inexpensive inference, financing for data centers, energy infrastructure, hardware, and potentially greater flexibility over local deployment and data sovereignty. China may even be willing to compromise on certain intellectual-property or technology-control issues when doing so helps establish Chinese AI infrastructure abroad. Many governments ultimately aren’t going to care whether the world’s best benchmark belongs to an American or Chinese company. They are going to care about cost, reliability, access, sovereignty, and whether their data remains under their control.
Of course, China has its own geopolitical problem: countries may distrust Beijing’s intentions, worry about surveillance, censorship, data access, or becoming dependent on Chinese infrastructure. And the U.S. still has enormous advantages in semiconductors, capital markets, research universities, software ecosystems, and global alliances. That’s why I don’t think this necessarily ends with one country completely defeating the other. We could instead end up in a world where American companies remain at the absolute technological frontier while Chinese companies dominate large portions of global AI deployment because they make intelligence dramatically cheaper and easier to access.
That is ultimately why I think America’s current lead can be misleading. The real danger for the United States isn’t necessarily that China suddenly develops a dramatically smarter AI model. It’s that China gets close enough. If American AI is technically superior but Chinese AI is cheaper, open, energy-abundant, locally deployable, and backed by China’s enormous industrial capacity, then China could lose the frontier-model race while still winning much of the global economic race. The U.S. can absolutely prevent that outcome—but doing so requires understanding that AI supremacy will not be determined by algorithms alone. It will be determined by talent, chips, energy, infrastructure, capital, cost, alliances, and ultimately which country can distribute useful intelligence to the rest of the world most effectively.
Power and Pelf 