Treasury Secretary Scott Bessent told Fox Business on Tuesday that the United States would examine whether Chinese AI models have stolen capabilities from American rivals, and that Washington has the ability to sanction overseas AI companies that steal. The next day, Michael Kratsios, the president’s science and technology adviser, posted on X that his office has information Moonshot AI distilled capabilities from Anthropic’s Fable AI model to build its Kimi K3. The message is clear enough: senior government officials are deploying the authority of the White House to protect one private company’s model from a rival’s competitive product, and the Treasury Department is promising enforcement tools to match. This is not an intellectual-property dispute being handled by courts. It is state power being exercised as a competitive weapon, and not even bothering to disguise the target.
It is worth being precise about what distillation is, because the word “stealing” does a great deal of work in Kratsios’s formulation and very little of it is honest. Distillation — querying a model’s outputs to train a smaller, cheaper model that approximates some of its performance — is a standard technique in the field. It is closer, in engineering terms, to reverse engineering than to corporate espionage. A distillation attack does not extract weights or architecture; it extracts behavior. The weaker model learns to produce similar answers. Whether this constitutes misappropriation depends on what, exactly, was appropriated, and the answer — pattern-approximation from freely available API outputs — is less obviously theft than it is the sort of competitive practice that has driven every information-technology market since engineers started studying each other’s products. The prior accusation Anthropic itself leveled at Alibaba in June used similar language about the “largest known” distillation effort. The language escalates; the engineering it describes is routine.
The hardware allegation is a separate and harder question. Kratsios accused Moonshot of gaining access to restricted cutting-edge Nvidia servers the company was not authorized to use. If true, this involves actual export-control violations — tangible chips, tangible shipping records, tangible contracts — and deserves the scrutiny it will get. Export controls on advanced semiconductors exist for documented national-security reasons, and violating them is not a gray-area engineering dispute. But Kratsios bundled this claim with the distillation accusation in the same post, and the bundling is the tell. If the hardware claim were the genuine concern, it would stand on its own — it does not need the distillation charge to give it weight. Bundling the two in a single post lets the stronger charge lend credibility to the weaker one, and lets the whole package serve a purpose neither claim serves alone: the framing of Chinese AI capability as fundamentally illegitimate.
That framing is doing real work right now. Moonshot’s Kimi K3 drew global attention last week because it appeared to narrow the gap between Western and Chinese AI. The competitive response — from Anthropic, from Nvidia, from the administration — is not to outperform the rival but to delegitimize it. If K3’s capabilities came from theft rather than engineering, then no American company needs to match them. The administration has spent months building this very posture, and the Kratsios accusation is its sharpest expression yet: the White House itself announcing, on a social-media platform, that a Chinese startup’s competitive product is stolen goods.
The irony is structural, and it runs deep. Anthropic — like every major American AI company — trained its models on copyrighted material scraped from the internet without the consent of the writers, artists, photographers, and publishers who produced it. The copyright question is live in multiple federal courts right now. The companies that built their models on unlicensed copyrighted material are now asking the U.S. government to sanction a rival for learning from their own unlicensed output. The extraction flows one direction; the IP claim flows the other. This is not intellectual-property protection in any principled sense. It is monopoly-rent defense dressed in the vocabulary of property rights.
Xi Jinping endorsed open-source AI and launched a China-led global standards body last week. Whether that initiative is genuinely redistributive or merely a different flavor of platform capture is a question worth asking. But the strategic trajectory is legible: China is building an alternative AI ecosystem with its own rules, its own standards, and its own bodies of legitimacy. The American response — a White House adviser accusing a Chinese startup of stealing on X, a Treasury Secretary threatening sanctions on cable television — is not going to slow that trajectory. It will accelerate it. Every foreign government watching the United States deploy its sanctions apparatus to protect one company’s model weights will draw the obvious conclusion about whose technology stack it is safe to depend on.
There is a public hearing before the Senate Commerce Committee next month on AI export controls and competitive policy. The hearing matters because hearings are where the record gets built. Whether Moonshot’s hardware procurement violated export controls is a question the Commerce Department’s Bureau of Industry and Security is equipped to investigate and should. Whether distillation from a publicly accessible API constitutes theft is a question for courts, not for the White House press office. The administration has decided to skip both processes and deliver its verdict on X. That decision tells you what this is about. It is not about protecting intellectual property. It is about protecting market position, using the Treasury Department as enforcement and the science adviser as a megaphone.