US threatens Moonshot with sanctions over distilling Anthropic's Fable

On July 22 the U.S. Treasury threatened sanctions against China's Moonshot AI after the White House accused the company of improperly distilling Anthropic's Fable model. It is the first time the administration has tied a specific Chinese model to an Entity List threat over intellectual-property theft.

Key takeaways

Treasury Secretary Scott Bessent warned of sanctions and Entity List designations for "industrial-scale distillation" that crosses into IP theft.
The claim: Moonshot distilled Anthropic's Fable model, publicly available since July 1, 2026.
The backdrop is the release of Moonshot's open-weight Kimi K3 model in the week of July 22.
The White House's Michael Kratsios says Moonshot used Nvidia GB300 servers in Thailand, breaching export controls.
Neither Moonshot nor Anthropic commented.

The claim: distilling the Fable model

At the core of the dispute is an accusation of knowledge distillation. It is a technique in which a smaller model (the student) learns by mimicking the outputs of a larger model (the teacher). In itself it is a standard machine-learning method, but doing it on the outputs of someone else's commercial model can violate its terms of service and intellectual-property rights.

The White House says Moonshot used Anthropic's Fable model this way, a model made public only on July 1, 2026. In the same window, in the week of July 22, Moonshot released its own open-weight Kimi K3. That timeline is central here: the short gap between Fable's launch and Kimi K3's debut is what the administration says makes the accusation plausible.

When firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table.

Scott Bessent, U.S. Treasury Secretary.

The export angle: GB300 and Thailand

The second axis of the accusations is hardware. Michael Kratsios, the White House science and technology policy chief, says Moonshot acquired servers equipped with Nvidia GB300 chips and accessed them in Thailand. GB300s are banned for sale to Chinese companies, so accessing them outside China could amount to circumventing U.S. export controls.

That is an important distinction. One charge concerns intellectual property (distillation), the other concerns hardware export controls (access to GB300s). The administration merges them into a single picture of a company that allegedly both copied a model and dodged hardware limits. For now, though, these are government claims without published technical evidence.

A wider fight inside the administration

The sanctions threat does not happen in a vacuum. Back in April 2026 the administration signaled steps to curb the distillation of Chinese models from U.S. systems, something OpenAI and Anthropic had pushed for. At the same time, the government camp itself is split over open models.

David Sacks, a former administration AI adviser, argued that large U.S. firms want the government to eliminate their open-source competition. Chinese models such as Kimi are free and less restrictive to use, creating real price pressure on paid U.S. rivals. The Moonshot dispute is therefore also a fight over whether Washington is defending intellectual property or the interests of specific companies.

The hardware context has been shifting too. In recent months the administration loosened some export controls, letting Nvidia sell more to China, only to sharpen its rhetoric against specific firms soon after. That inconsistency makes a clear read hard: at times the priority is selling chips, at times protecting the U.S. technological edge. The Moonshot threat fits the latter narrative, but without published evidence it remains mostly a political signal.

Why it matters

This is the first case in which the U.S. administration publicly ties a specific Chinese model to a sanctions threat over distillation. It moves the argument about AI training methods from the technical plane to the geopolitical and commercial one. If the Entity List is actually used against an AI company over how it trained a model, it will set a precedent that is hard to reverse. The problem is that distillation is very hard to prove: similar outputs from two models are not yet evidence of copying, and the administration has not presented technical proof. The case also exposes the tension in U.S. policy on open weights: the same open models that some want to restrict as an IP threat, others treat as healthy competition. For companies training models, it means growing regulatory risk around training data and the provenance of the learning signal.

What's next?

Treasury named sanctions and the Entity List as options — the next step is a formal decision or the lack of one.
No response yet from Moonshot or Anthropic — statements from either could change the picture.
Open questions remain: technical proof of distillation and verification of the GB300-in-Thailand claim.