A senior White House technology official publicly criticised Chinese startup Moonshot AI, saying it developed its new AI model Kimi K3 by illegally copying Anthropic’s Fable model. The claim came as Moonshot AI prepares to publicly release Kimi K3.

On July 22, Cryptopolitan, a blockchain outlet, reported that White House Office of Science and Technology Policy Director Michael Kratsios said on X, formerly Twitter, that Moonshot AI built a dedicated internal system to carry out large-scale extraction from U.S. models.

He claimed Moonshot AI repeatedly changed its access routes to avoid tracking and conducted large-scale, covert distillation to take U.S. technology. Kratsios said, "Large-scale covert industrial distillation to steal America’s proprietary technology and undermine American research is unacceptable."

The method at issue is a training technique known in the AI industry as "distillation". It teaches a weaker model to mimic performance by learning from the outputs of a stronger model. The technique is widely used when applied in limited ways in open environments, but Kratsios took issue with Moonshot AI allegedly conducting it privately at industrial scale. He views it as effectively theft rather than routine engineering.

The claim also links to findings Anthropic released in February. Anthropic said at the time that more than 3.4 million Claude conversations were traced to Moonshot AI, and it alleged hundreds of fake accounts were used in the process. It said some account records matched public profiles of senior Moonshot AI employees. Anthropic said the activity targeted Claude’s reasoning, coding and vision capabilities.

Kratsios also claimed Nvidia GB300 servers were used in the work. The servers were deployed in Thailand, and the GB300 is part of Nvidia’s high-performance chip line. He pointed out that such chips are barred from being brought into China under U.S. export controls.

Moonshot AI unveiled Kimi K3 on July 16. The company described it as an open-weights model with 2.8 trillion parameters, and it halted new Kimi sign-ups two days after the release. The full model weights are due to be released by July 27.

Anthropic sees the release itself as a risk. It said models produced through copying may not retain the original model’s safety guardrails, and once weights are released they can spread beyond control. Anthropic said it is concerned that open-weights models can slip out of anyone’s control.

Moonshot AI has not issued a public response so far. Kratsios also did not specify possible sanctions. In January 2025, AI czar David Sacks claimed DeepSeek copied an OpenAI model, but DeepSeek has fully denied it. The Chinese government also dismissed Anthropic’s February findings as baseless.

The dispute comes as the United States and China prepare AI talks in September. The United States still far outpaces China in the scale of AI investment, but Chinese research labs continue to narrow the performance gap with the United States. In that context, allegations of copying around Kimi K3 show the two countries’ AI rivalry is spreading beyond model performance to training data, chip controls and the spread of open models.