US threatens sanctions over China AI
Treasury says China AI sanctions are on the table after the White House named Moonshot, alleging it distilled Anthropic's Fable and used banned Nvidia chips.
Treasury Secretary Scott Bessent said on July 21 that the US will examine open-weight AI models from Chinese companies for signs of intellectual-property theft and could sanction the firms behind them if it finds evidence. The specific accusation is distillation: training a new model on the outputs of an American frontier model to approximate its capabilities without the underlying research cost. Bessent framed the administration’s position as pro-open-source in general, but not tolerant of foreign firms copying US-developed technology by that route.
The threat lands days after Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model that multiple benchmarks put close to OpenAI and Anthropic’s frontier tier, the latest in a run of Chinese releases (Kimi K3, Qwen, GLM, DeepSeek) that have closed the open-weight gap faster than most US labs expected. No specific company or model has been named as a sanctions target yet, and no enforcement action has been announced.
Update, July 22-23: The administration went from general warning to a specific accusation. White House OSTP director Michael Kratsios said on X that Moonshot “developed a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of access to avoid detection,” naming Anthropic’s Fable as the model Moonshot allegedly distilled to build Kimi K3. Kratsios separately alleged Moonshot acquired Nvidia GB300-equipped servers and accessed GB300 chips in Thailand, in apparent violation of US export controls. Bessent followed up, saying sanctions “will be on the table” for companies that cross the line into “industrial-scale distillation attacks.” Moonshot has not confirmed or denied the specific claims.
AI researchers who spoke to reporters are skeptical of the distillation timeline: Braden Hancock of the Laude Institute noted Fable only became publicly available July 1, leaving roughly two weeks for Moonshot to distill data, train, and ship a model at Kimi K3’s capability level, which he called implausible. Nathan Lambert of the Allen Institute for AI added that pulling frontier-grade training data through API queries at scale would be “insanely expensive” and slowed by response latency, and that distillation matters less as competing labs close in on the frontier, where reinforcement learning does more of the work than imitation.
For operators who’ve been routing between Chinese and Western models on cost, this is a policy risk to price in alongside the technical and data-residency risks already inherent to that choice: a sanctions regime aimed at specific labs or models could cut off API access, hosting partnerships, or even weight downloads with little notice, the same way export licensing has already swung back and forth on H200 chip access between the US and China. Nothing here changes today’s access, but a routing strategy built entirely around the cheapest Chinese open-weight model now has a new single point of failure to watch. Nearly 200 US startups are already pushing back against a broad restriction, so the policy outcome here is genuinely contested, not a foregone conclusion.