Full Breakdown
Trump Administration Accuses Moonshot AI of IP Theft via Model Distillation
7/23/2026, 6:29:00 AM
Core Event
On July 22 2026, Michael Kratsios, director of the White House Office of Science and Technology Policy, alleged that Beijing-based Moonshot AI “distilled” Anthropic’s Fable model to develop its Kimi K3 large-language model. Treasury Secretary Scott Bessent warned that sanctions could be imposed against companies engaged in such IP theft. Moonshot AI and the Chinese embassy did not respond to requests for comment.
Background & Context
Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model, on July 16 2026. Its coding and reasoning performance approached Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6, narrowing the United States’ lead in frontier AI. Anthropic previously accused Moonshot and two other Chinese labs of using 25,000 fraudulent accounts to generate 28.8 million interactions with its Claude models, labeling the activity a “distillation attack.”
U.S. export controls forbid the sale of Nvidia’s GB300 (Blackwell) chips to Chinese entities. Kratsios claimed Moonshot accessed such chips via servers in Thailand, suggesting a possible violation of those controls.
Data & Statistics
- Kimi K3 parameters: 2.8 trillion – the largest open-weight model released by a Chinese developer.
- Anthropic’s reported fraud: 25 k fake accounts, 28.8 M Claude interactions over six weeks.
- OpenRouter token routing (July 2026): Chinese models 46.4 % of tokens, U.S. models 35.7 %.
Official Statements & Responses
- Scott Bessent (Treasury Secretary): The administration “has the ability to sanction” firms that steal U.S. IP.
- Liu Chang (Chinese embassy spokesperson): Called the allegations “entirely unfounded” and affirmed China’s commitment to IP protection.
Criticism & Opposition
- Alon Yamin, CEO of Copyleaks, warned that policing distillation is “hard to police” and noted that many AI models are built on copyrighted content without permission.
- David Sacks, former White House AI czar, argued that restricting Chinese open-weight models could primarily benefit the “duopoly” of OpenAI and Anthropic.
Conflicting Reports & Gaps
- Hamish Low (Institute for AI Policy and Strategy) finds it “pretty overwhelmingly likely” that Kimi 3 was distilled from U.S. models.
- Nathan Lambert notes that Anthropic’s Fable was publicly available only on July 1 2026, a short window before Kimi 3’s launch on July 16 2026.
- No public forensic evidence directly links Fable outputs to Kimi 3’s weights; Moonshot has not disclosed technical details, and U.S. investigators have not released findings.
Verbatim Quotes
- “We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model,” — Michael Kratsios
- “This is more than a heist of invaluable American Intellectual Property,” — State Jacob Helberg
- “Illicit, adversarial distillation is IP theft and industrial espionage that supports adversary military and intelligence capabilities,” — Sarah Heck
What’s Next
- The administration is reviewing possible sanctions and a bipartisan amendment to the National Defence Authorization Act that would codify sanctions and export-control measures against Chinese AI firms.
- Treasury Secretary Bessent is slated to lead the first official U.S.–China AI dialogue in September 2026, ahead of President Xi’s planned visit.
- Internal deliberations consider adding Moonshot and similar companies to the Commerce Department’s Entity List and potential executive actions requiring U.S. firms to guarantee security when using foreign models.
