The Trump White House is choosing between two opposite China-AI strategies

Last week, the Chinese AI lab Moonshot released a new flagship model, Kimi K3, that the Trump administration says rivals the top US systems from Anthropic and OpenAI. The release landed inside a White House that has spent months arguing about how to respond to exactly this moment, and the fight it triggered is now the most concrete test yet of how the United States plans to keep its lead in frontier AI.

According to multiple people familiar with the deliberations, reported by Wired, the administration is split between parts of the White House that want stricter controls on Chinese AI and a Commerce Department that considers those restrictions unworkable. The two camps are not arguing about whether to act. They are arguing about which opposite strategy to pick.

The flashpoint is a training technique called distillation, in which one AI model is trained on the outputs of another. Distillation is a normal method for building smaller, cheaper models from larger ones, and US labs use it on their own systems. The White House's case is that when a Chinese lab runs distillation against an American model covertly and at scale, it stops being research and becomes theft of the kind that warrants a sanctions response.

Treasury Secretary Scott Bessent told Fox Business that covert distillation could trigger sanctions and called it "IP theft." The White House's Office of Science and Technology Policy amplified the charge on X, arguing that Moonshot built Kimi K3 by distilling Anthropic's Claude. Anthropic has separately alleged that the Chinese tech company Alibaba carried out the "largest known distillation attack" against it, according to the BBC.

That is one theory of how to keep the US ahead: treat the route to a frontier model as a controlled technology, and sanction anyone who takes it without permission. It is the AI-as-controlled-tech view: build a wall, police the route, hit the cheaters.

Commerce Secretary Howard Lutnick is pushing the opposite bet. Rather than wall off the technology, the Commerce route is to push Anthropic, OpenAI, and the rest of the US frontier labs to release more open-weight models — AI systems whose trained parameters are published so anyone can run, inspect, or fine-tune them — and race China on cost, access, and diffusion. Earlier this month, the administration lifted export controls on certain Anthropic models, and Anthropic added a new security measure to stay in the administration's good graces, the kind of accommodation that fits a Commerce-led, openness-first doctrine.

The two strategies are not actually on equal footing inside the government. Commerce, through its Bureau of Industry and Security, is the agency that writes and enforces US export controls on chips and AI. A White House that wanted to sanction distillation would still need BIS to write the rule, and the same agency is being told by its own political leadership to lean the other way. The result, as one source put it to Wired, is that Lutnick is "more freewheeling" than his deputy on the question. The White House has not yet sent a formal request for input to Commerce at all.

There is also a deeper problem with the sanctions route that neither side has resolved publicly. Distillation is not a settled legal category. Most uses of it, including inside US labs, are normal engineering. Drawing a line around "covert" distillation, and proving a particular Chinese model was built that way, is a much harder forensic question than the current rhetoric suggests. Alibaba and Moonshot have both rejected the characterization, and a sanctions designation built on disputed training-method claims would face immediate legal pushback.

That is why the next few weeks matter more than the past few months. People familiar with the deliberations expect the White House to take some form of presidential action, though not necessarily an executive order, after the Anthropic–Alibaba allegation. The form of that action, whether a sanctions designation, a new BIS rule, a directive to push open-weight releases, or some combination, will tilt US AI policy toward one pole for the rest of the decade.

The other tell to watch is whether Lutnick's open-weight push actually produces a release. The administration has not committed to specific incentives, and US frontier labs are not, as of now, lining up to publish their best models. If no major US lab publishes a flagship open-weight model by the end of the year, the policy debate quietly defaults to the sanctions track regardless of which theory the White House says it prefers.

For now, the most honest read is that the Trump administration is not split between two coherent camps. It is split between two theories of how to maintain US AI leadership, wall off the route or open the road, and the Kimi K3 release has made that choice unavoidable.