Two A.I.s: How the U.S. and China Are Building Different Machines

Ten years after the U.S. and China each issued their first national plans for artificial intelligence, the two countries are not building the same future. They are running two separate experiments on the same underlying technology, and the gap between them is widening. The forks show up in the policy logic, the funding model, the data the systems train on, and the question of whose lives the technology is meant to serve.

Kai-Fu Lee has spent a career working in both versions of the field. Today he is the chief executive of 01.AI, a Beijing-based artificial-intelligence company he founded in 2023, and he has watched the divergence up close. In a recent New Yorker profile, he framed the moment in plain language: "It's two different universes now." The line cuts against the wire-style narrative that the U.S. and China are racing toward the same finish line. They are not. They are walking out of the same lab in different directions.

The U.S. side of the fork is the one most American readers will recognize. In 2016, the Obama Administration released the federal government's first major A.I. plan, a document that, in retrospect, "barely registered." It dropped days after the Access Hollywood tape upended the 2016 campaign, and the next administration scrapped it. President Trump handed A.I. strategy largely to Silicon Valley, and Washington has, by and large, stayed out of the way since. The result is a market-led model: capital flows to whoever can ship the best model, regulators arrive after the fact, and the question of whose values get encoded into the system is left to the companies that build it.

China's 2016 plan, issued a few months later, was something else. Dense with Communist Party jargon, it laid out an industrial policy spanning agriculture, policing, transportation, and defense, with explicit targets: "major breakthroughs" by 2025 and "world leading" status by 2030. Lee calls the model "all-hands-on-deck." The state sets the goals, the state directs the capital, the state decides which applications ship first, and the state owns the data. The system is not racing the U.S. toward the same finish line. It is encoding a different answer to the question every A.I. system has to answer first: who is this technology for, and who decides?

Lee's biography is the most useful way to see what the fork looks like from inside. He was born in Taiwan in 1961, took a computer-science degree from Columbia in the 1980s, and studied A.I. at Carnegie Mellon. At Microsoft he urged Bill Gates to acquire a small search startup called Google; Gates passed. Lee later ran Google in China. In 2009 he moved to investing and tried to operate in both Beijing and Silicon Valley, but eventually closed the U.S. office. He has called A.I. "men's final step to understand themselves," but the more telling quote, given the moment, is the one about universes.

What is actually being encoded is the harder question. On the U.S. side, the values that get baked in are the values of the companies shipping the models, plus the slow grind of litigation, regulation, and public pressure. On the Chinese side, the values are the priorities of the state, with the efficiency and the blind spots that implies. Both systems are now training on data the other will never see, deploying applications the other has decided not to build, and writing policy the other country has no leverage over. The "race" framing, useful as it is for funding rounds and campaign speeches, papers over a more uncomfortable observation: the two A.I.s are not interchangeable.

The Hacker News thread on Lee's profile, predictably, treats the question as a horse race. The top-voted comments split between "China wins on data and deployment" and "the U.S. wins on chips and frontier research," as if the only thing at stake is which team ships the bigger model first. That is a useful debate to have, but it is not the one Lee is making. The fork is about more than performance benchmarks. It is about which institutions get to decide what A.I. is for, and which lives the technology is optimized to serve first.

Ten years on, the two 2016 plans are visibly diverging. The U.S. plan was effectively abandoned; China's plan is on a deadline. The architectures that result will not just compete with each other. They will carry their respective policy logics into every hospital, courtroom, classroom, and police department that adopts them, and those institutions will inherit a set of assumptions their governments never voted on. The next question is not who wins the race. It is who gets to keep asking what the race is for.