Full Breakdown
U.S. Officials Accuse Moonshot AI of Illicit Use of Banned Nvidia Chips and U.S. AI Models
7/22/2026, 8:13:39 PM
Core Allegations
White House Office of Science and Technology Policy Director Michael Kratsios posted that Chinese startup Moonshot AI accessed “GB300-equipped servers” in Thailand to train its Kimi K3 system using Nvidia’s Blackwell chips, which are barred from sale to Chinese companies. He said the move violates export-control rules and the terms of service of U.S. AI models that Moonshot allegedly distilled. Treasury Secretary Scott Bessent warned the administration could sanction entities that engage in “IP theft” of American large-language models.
Background & Context
The Bureau of Industry and Security’s May memo reaffirmed that advanced AI processors such as Nvidia’s Blackwell line are restricted for export to Chinese firms, while earlier-generation H200 chips remain permissible. Nvidia has publicly denied any loophole that would allow advanced chips to reach Chinese subsidiaries.
Moonshot’s Kimi K3, unveiled in late July 2026, claims open-weight performance that surpasses most U.S. rivals except Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6.
Timeline
- April 2026 – Administration announced plans to address “adversarial distillation,” the practice of training a smaller model from the outputs of a larger one.
- May 29, 2026 – Treasury Secretary Scott Bessent warned that the U.S. would sanction overseas models that steal from American companies.
- Late July 2026 – Moonshot released Kimi K3, prompting immediate industry and policy reactions.
Data & Statistics
Official Statements & Responses
Nvidia’s spokesperson rejected the notion of a loophole that would allow advanced chips to reach Chinese subsidiaries, stating the company operates under the assumption that no such pathway exists. Moonshot AI has not issued a public response.
Criticism & Opposition
Anthropic’s letter to the Senate Banking Committee alleged that Alibaba-affiliated entities carried out the “largest known distillation attack” against its models.
Conflicting Reports & Gaps
Verbatim Quotes
- “This administration supports open source models, but what we do not support is IP theft. If we see, especially, that overseas models are stealing from our great companies, we have the ability to sanction them,” — Scott Bessent, treasury secretary
- “We are finding watermarks of our U.S. large language models on many of the Chinese models, and that's unacceptable,” — Scott Bessent, treasury secretary
What’s Next
U.S. and Chinese officials are slated to hold bilateral talks on AI governance in September, with Treasury Secretary Bessent expected to represent the United States. The discussions could shape future enforcement actions and export-control policies concerning advanced AI hardware and model-distillation practices.
