Kimi K3 Too Powerful? White House Accuses Moonshot AI of Illegally Using NVIDIA GB300 Chips

Foreign media reported on Wednesday (22nd) that U.S. White House officials have accused Chinese artificial intelligence (AI) startup Moonshot AI of improperly using U.S. AI models and export-controlled NVIDIA (NVDA-US) GB300 chips to develop Kimi K3, an AI model that shocked the technology industry last week with its advanced capabilities.

Michael Kratsios, Director of the White House Office of Science and Technology Policy, posted on social media Wednesday stating that Moonshot AI acquired servers equipped with GB300 chips and likely used the equipment in Thailand, possibly to train AI models. He argued that the company may have violated U.S. export control regulations and technology companies’ terms of service in developing Kimi K3.

The GB300 is part of NVIDIA’s Blackwell generation products, which the U.S. prohibits from being sold to Chinese companies. Although the Trump administration has allowed limited sales of NVIDIA’s H200 chips to China, more advanced Blackwell products remain under export ban.

The U.S. Department of Commerce’s Bureau of Industry and Security (BIS) issued a memo in May reaffirming that even if Chinese companies are located outside China, the sale of advanced processors to them remains restricted. The Trump administration is currently investigating whether companies are exploiting loopholes to ship advanced AI chips to overseas subsidiaries of Chinese firms.

NVIDIA has previously denied the existence of such loopholes and stated it enforces regulations as if no loopholes exist, but has not yet responded to Kratsios’s latest allegations. Moonshot AI has also not responded to the claims.

Kimi K3’s performance has challenged the market’s prevailing view that Chinese AI capabilities significantly lag behind those of OpenAI and Anthropic. Moonshot AI claims this open-weights model surpasses most competitors in overall capability, trailing only Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6. The news triggered a sell-off in tech stocks, with market reactions similar to the impact caused by DeepSeek last year.

Kratsios also accused Moonshot AI of building a complex internal platform to conduct large-scale “distillation” of U.S. AI models. Model distillation involves using the outputs of a more powerful “teacher model” to train a smaller “student model,” thereby replicating some of its capabilities.

U.S. Treasury Secretary Bessent expressed concerns on Tuesday about Chinese open-source AI firms’ distillation practices, calling the unauthorized acquisition of U.S. model capabilities intellectual property theft. He warned that if foreign models steal U.S. corporate technology, the Trump administration has the ability to impose sanctions.

OpenAI and Anthropic have recently made similar accusations, claiming Chinese competitors are systematically and without authorization using outputs from top U.S. models to develop chatbots at lower cost. Anthropic previously named Moonshot AI as conducting model distillation. Both chambers of the U.S. Congress are now advancing legislation that could blacklist or sanction Chinese companies found to improperly distill U.S. models.