White House Points Finger at China's Moonshot AI Over Alleged Anthropic Model Theft
Michael Kratsios, Donald Trump's Assistant for Science and Technology, has publicly accused Chinese AI startup Moonshot AI of building its Kimi K3 model by distilling Anthropic's Fable — essentially bulk-querying a proprietary model to train a cheaper knockoff.
Kimi K3 is a 2.8-trillion-parameter open-weights model that raised eyebrows when it dropped on July 16. Its quality was good enough to spook investors, who promptly knocked US AI stocks down on fears that Chinese labs might be closer to the frontier than anyone comfortable admitted.
Kratsios took to X to make the accusation explicit. According to him, Moonshot AI didn't just run a few queries — it built a "sophisticated internal platform" specifically designed to conduct large-scale distillation against US models, with built-in methods for rotating access routes to dodge detection. That's not someone accidentally tripping over a usage policy. That's an operation.
He also alleged the company got its hands on Nvidia GB300 accelerators, hardware the US prohibits from being sold to China. The workaround, apparently, was accessing GB300s running in Thailand. Renting compute outside China to sidestep export controls is a well-worn trick at this point, but it's rarely been called out so directly by a sitting White House official.
Kratsios was careful to note that distillation itself isn't illegal. Compressing a large model into a smaller, more efficient one is standard practice across the industry. The objection is the "covert industrial" variety aimed at extracting proprietary technology from US competitors without their knowledge or consent.
Treasury Secretary Scott Bessent piled on with his own post, framing it as a defence of open-source principles while drawing a line in the sand. His argument: open source is not a standing invitation to strip-mine American IP. If Chinese firms are running what he called "distillation attacks" at industrial scale, sanctions and Entity List additions are on the table.
Whether that threat lands is another question. The US has already sanctioned most of China's significant tech players. The result has been a cottage industry of evasion — grey market hardware, offshore GPU farms, creative supply chain routing. Sanctions haven't stopped China from building capable AI labs, and there's little evidence that another round of designations would change the trajectory.
This isn't the first time Anthropic has pointed the finger at Moonshot AI. The company made the same distillation allegation back in February 2026, naming DeepSeek and MiniMax alongside Moonshot as repeat offenders.
Beijing, predictably, denies everything. It always does.