IA · 23 July 2026 · 4 min read

The Distillation War: U.S. Threatens Sanctions Over Chinese AI Models

In brief: U.S. Treasury Secretary Scott Bessent and White House officials have threatened economic sanctions against Chinese AI developers, alleging that Moonshot AI engaged in covert, large-scale distillation of Anthropic's Claude Fable 5. The clash highlights growing geopolitical friction as high-performing Chinese open-weight models challenge the business models of Silicon Valley's leading AI labs.

by Team Mocchi's

Tensions between Washington and Beijing over artificial intelligence have reached a new flashpoint. The U.S. Department of the Treasury and the White House have explicitly raised the prospect of economic sanctions and Entity List designations for Chinese AI laboratories accused of conducting systematic "distillation" attacks against frontier American models.

At the center of the dispute is Chinese startup Moonshot AI, creator of the open-weight model Kimi K3. As reported by TechCrunch, Treasury Secretary Scott Bessent warned that "open source is not open season on American IP," accusing the firm of conducting covert distillation attacks on an industrial scale.

Accusations of "IP Theft" Through Model Distillation

Model distillation is a widely used AI training technique in which a smaller, more efficient model learns from the outputs generated by a larger frontier model. While distillation is routinely employed across the industry for model optimization, U.S. officials argue that large-scale training on the synthetic outputs of proprietary commercial systems — specifically Anthropic's Claude Fable 5 — constitutes intellectual property theft.

The release of Moonshot's Kimi K3, distributed for free with open weights and capabilities rivaling top-tier Silicon Valley offerings, has alarmed policymakers. The widespread availability of high-performing open-weight models threatens to erode the commercial margins of leading U.S. AI labs, whose massive capital expenditures rely on subscription fees and pay-per-token API access.

Shadow Hardware and Export Control Evasion

Beyond software training practices, the White House has also scrutinized the hardware infrastructure powering Chinese AI development. Michael Kratsios, head of the White House Office of Science and Technology Policy, alleged that Moonshot accessed Nvidia GB300 servers hosted in Thailand to bypass U.S. export controls.

Nvidia's Blackwell-generation chips are strictly prohibited from export to Chinese entities under federal trade regulations designed to limit foreign military supercomputing capabilities. The alleged leverage of third-party data centers in Southeast Asia adds severe trade enforcement concerns to the ongoing copyright dispute.

A Divided Washington Over Open-Weight AI

Despite aggressive rhetoric from the White House, formulating a policy response has triggered debate within Washington. As revealed by WIRED, officials are split between national security hawks urging outright bans and the Commerce Department, which views broad bans on open-weight software as practically unenforceable.

Furthermore, as noted by MIT Technology Review, the rapid adoption of free, high-capability models from foreign labs is undermining the economic moats of American tech giants, leading some policymakers to advocate for U.S. open-source incentives as a counter-strategy.

Mocchi's take

The distillation dispute highlights an inevitable shift in the AI ecosystem: as technical knowledge democratizes, relying solely on trade secrets to protect proprietary models becomes increasingly fragile. For European companies, this geopolitical friction underscores the necessity of building flexible, model-agnostic software architectures rather than locking business logic into a single cloud API vendor. Leveraging performant open-weight models — or implementing hybrid deployment models — enables enterprises to preserve data sovereignty and operational resilience regardless of international regulatory shifts. Ultimate business value no longer resides in owning a foundational model, but in the capability to seamlessly integrate intelligence into core enterprise processes.

All articles on the Mocchi's blog