Analysis
Treasury Secretary Scott Bessent said the US could impose sanctions on China if Chinese AI models are found to have been trained on stolen American intellectual property, escalating a fight over Moonshot AI's Kimi K3 model from a market story into a potential trade-policy one. "We are finding watermarks of our U.S. large language models on many of the Chinese models, and that's unacceptable," Bessent said. "We're going to be looking at that in the coming days or weeks."
Bessent used the technical term distillation -- the practice of training a smaller, cheaper model on the outputs of a larger, more expensive one -- to describe the specific concern. Independent researchers have separately noted that Kimi K3's benchmarked capabilities appear to exceed what Moonshot AI's known compute resources should be able to produce through standard from-scratch training, fueling speculation that some of its training signal came from outputs of competitor models rather than being disclosed as original training data.
“Kimi K3's release helped trigger a 24% drawdown in chip stocks by undercutting the thesis that near-frontier AI requires frontier-level compute spend.”
The administration's stated position tries to thread a needle: "This administration supports open source models, but what we do not support is IP theft," Bessent said. "If we see, especially, that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft." That framing lets the administration keep backing US open-weight efforts while treating a specific allegation against a specific foreign lab as a trade-enforcement matter rather than a broader crackdown on open models generally.
The timing compounds an already unusually tense week for the open-vs-closed AI debate. Kimi K3's release helped trigger a 24% drawdown in chip stocks by undercutting the thesis that near-frontier AI requires frontier-level compute spend. In the same window, OpenAI's head of strategic futures argued -- then walked back -- that regulators should specifically fear open-weight models to protect frontier labs' capex advantage. Bessent's sanctions threat adds a third, more consequential thread: an actual government enforcement mechanism aimed at a specific competitor, rather than industry commentary about competitive dynamics.
What to watch: whether Treasury identifies specific evidence tying Kimi K3 to unauthorized distillation from a named US model, whether any sanctions actually get proposed versus remaining a rhetorical threat, and how Moonshot AI and the Chinese government respond.
