US Treasury Threatens Sanctions on China, Claims Its AI Firms Distilled Anthropic’s Fable Model

The US Treasury Secretary Scott Bessent has warned China that economic sanctions are on the table for what he describes as “IP theft.” He claims that Chinese AI developer Moonshot used improper techniques to train its Kimi K3 AI on Anthropic’s Fable model, though others also accuse it of breaching US export controls.

Late last week, Moonshot had what many analysts described as a “DeepSeek moment,” in which its latest model proved far more capable than anyone expected, delivering cheaper, more efficient results than some frontier models from Western firms, albeit at a slower pace. Microsoft quickly considered using Kimi K3 for its Copilot, but now the White House is considering banning foreign open-source models entirely, and may even sanction China over its development.

Although model distillation—using smaller models to learn from the outputs of a larger one—is a common AI training technique, Bessent claims Moonshot may have gone too far and “crossed the line into IP theft.” Describing its actions as akin to “industrial scale distillation attacks,” he said the US government would now look into Chinese open source models for more direct evidence of IP theft. If discovered, economic sanctions would be considered in response.

Tweet
(opens in a new tab)

Other components of the US administration have been calling foul on Moonshot’s AI efforts, too. The White House’s science and technology policy lead, Michael Krasios, has alleged Moonshot breached US export controls by training its AI using banned Nvidia GB300 GPUs hosted in Thailand.

He also said that while model distillation to create more efficient models was a legitimate method of optimization, using it covertly on a large scale to steal proprietary technology was unacceptable.

Although the White House investigation into the alleged IP theft may need to conclude before any concrete action is taken, there is mounting evidence that it may look to ban American firms from using Chinese frontier models entirely. Meanwhile, over 200 startups have urged the White House not to restrict AI access, claiming it would stifle innovation and disproportionately harm smaller AI firms while benefitting larger ones.

It would be a boon for OpenAI and Anthropic, they said, as those companies would effectively monopolize the frontier AI model market in the US, with no international alternatives.

Regardless of what follows, the advent of Kimi K3 has raised concerns and questions about the US AI industry. With a million tokens costing less than a third of the price of Fable, and fractions of the cost of the best OpenAI and Google models, what are these US companies doing differently or incorrectly? And do we really need all that extra compute if the models can be so much more efficient?