A new Chinese AI model just rattled Silicon Valley — and Washington says it knows exactly how it was built.
Michael Kratsios, director of the White House Office of Science and Technology Policy, said on July 22, 2026, that the administration has information that Beijing-based Moonshot AI covertly distilled Anthropic’s Claude Fable 5 model to build its newly released Kimi K3 system. Treasury Secretary Scott Bessent followed with a warning that sanctions and Entity List blacklisting could be on the table.
The Accusation
Kratsios alleged Moonshot built a “sophisticated internal platform” capable of running large-scale distillation against U.S. models, constantly switching access methods to dodge detection. Anthropic has separately said it identified more than 3.4 million Claude exchanges carried out through fraudulent accounts, with metadata reportedly tying the activity back to senior Moonshot employees. Bessent went further on Fox Business, claiming his office found “watermarks” of U.S. language models embedded in Chinese systems — a specific forensic claim that, notably, hasn’t been accompanied by any public methodology or data release.
The Chip Angle
The dispute isn’t just about software. Kratsios also alleged Moonshot obtained servers equipped with Nvidia’s restricted GB300 chips — hardware built on Nvidia’s Blackwell architecture that’s barred from sale to China — and accessed similar systems through Thailand, likely to help train its models.
Why Researchers Are Skeptical
Here’s the wrinkle: independent AI researchers have pointed out a timing problem. Claude Fable 5 was only re-released on July 1 after briefly going offline due to U.S. export controls, while Kimi K3 launched just 15 days later, on July 16 — a tight window for the kind of large-scale distillation being alleged. Neither the White House nor Anthropic has published logs, query records, or training-data evidence to substantiate the specific Fable-to-K3 claim, leaving the public case thinner than the rhetoric surrounding it.
What Makes Kimi K3 a Threat
Controversy aside, the model’s specs explain the alarm: Kimi K3 is the first open-weight model with 2.8 trillion parameters and a 1-million-token context window, undercutting closed-source U.S. competitors on price while performing close to frontier level on coding and reasoning tasks. It’s popular enough that Moonshot had to freeze new signups within two days of launch — its compute infrastructure couldn’t keep up with demand.
The Fallout on Both Sides
Washington is now weighing an outright ban on Chinese AI models and threatening Entity List placement for firms involved in covert distillation. Beijing’s Commerce Ministry fired back, accusing the U.S. of “AI hegemonism” and vowing countermeasures. Notably, the response inside the U.S. hasn’t been unified either: 179 American tech founders and startups signed a joint letter opposing a broad ban on foreign open-weight models, arguing it would hurt American innovation more than it protects it.
