Trump administration flags Kimi K3 model as evidence of Anthropic 'heist', triggering export-control review on Chinese AI-chip IP.

Beijing-based startup Moonshot AI recently unveiled Kimi K3, a 2.8 trillion-parameter open-weight AI model that delivers performance approaching Anthropic's Fable model at a fraction of the cost. Demand exploded so rapidly that the company was forced to temporarily pause paid subscriptions due to compute constraints.

Kimi K3's release immediately drew scrutiny from the Trump administration. White House Office of Science and Technology Policy Director Michael Kratsios accused Moonshot of executing a "large-scale, covert industrial distillation" campaign against Anthropic's proprietary Fable model, alongside accessing restricted Nvidia GB300 AI chips. Treasury Secretary Scott Bessent threatened sanctions against Chinese AI companies for stealing American intellectual property through this technical process. Undersecretary of State Jacob Helberg called it a "heist of invaluable American Intellectual Property."

In AI development, distillation typically involves using outputs from a larger "teacher" model to train a smaller "student" model through supervised fine-tuning—a standard technique employed by virtually all AI labs. When companies systematically replicate a rival's model outputs by launching thousands of automated API queries, often while switching IP addresses to evade detection, AI researchers refer to this as "adversarial distillation."

The central question remains: did Moonshot use adversarial distillation against Anthropic's frontier models to create Kimi K3?

Experts are divided. Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI, expressed skepticism: "I don't think you get a model this strong and this quickly on the heels of Fable doing strictly distillation. There's just not even frankly time, right? Fable's only been publicly available since July 1st. You can't distill that much data, train a model, and release it in two weeks."

Ryan Greenblatt, chief scientist at Redwood Research, challenged this assessment: "Many people seem to be arguing this is implausible because Fable hasn't been out for long. But Moonshot could have obtained unauthorized access before Fable was released via actors with prerelease access (Glasswing) or possibly by hacking Anthropic directly." Greenblatt also noted that test results suggest Kimi K3's pretraining data may have involved distillation of older Anthropic models, though not necessarily Fable specifically.

Anthropic had previously accused Alibaba, a different Chinese company, of conducting adversarial distillation "on an industrial scale," according to a letter obtained by Ars and sent to Senators Elizabeth Warren (D-Mass.) and Tim Scott (R-SC).

The Kimi K3 controversy extends beyond distillation to hardware access—arguably the more decisive factor. If Moonshot accessed restricted Nvidia AI chips, it would have violated both U.S. export controls and China's own domestic ban on these chips, imposed to stimulate domestic chip production.

Compute—the physical resources required to run AI models, including chips, electricity, high-bandwidth networking, and cooling facilities—lies at the heart of the U.S.-China AI competition. With China advancing rapidly in AI model innovation, the global AI race may ultimately depend on which side commands greater computing resources. As Liang Wenfeng, founder of Chinese AI company DeepSeek, explained on an investor call: "Our gap with the U.S. is mainly resources. The talent gap is essentially also a compute gap."