The US just threatened to sanction Chinese AI models directly, not just the chips that train them

US Threatens Sanctions on Chinese AI Models Over IP Theft

Until now, Washington's strategy for slowing Chinese AI was built on silicon: restrict the chips, and you restrict the models that depend on them. Treasury Secretary Scott Bessent just opened a second front. He said the US will examine Chinese open-source AI models for evidence of intellectual property theft and warned of sanctions if theft is established 1. The comments, delivered on Fox Business and first reported by Bloomberg, signal that Washington may start targeting the models themselves, not just the hardware behind them 1.

For anyone building products on open-weight Chinese models, or backing companies that do, this creates a category of regulatory risk that did not previously exist. A sanctionable model is a fundamentally different proposition from a restricted chip. Chips sit in data centers where customs agents can find them. Models are software that moves across borders in seconds.

Chinese open-source models have been narrowing the gap with US frontier labs. Moonshot AI's Kimi K3 is the latest example, part of a wave of releases that threaten the business models and fundraising prospects of companies like OpenAI and Anthropic 1. Axios separately reported Monday that the Trump administration was weighing a wholesale ban on Chinese open-source models, though that claim has been disputed 1.

Bessent drew a line between supporting open source and tolerating what he characterized as theft. "This administration supports open source models, but what we do not support is IP theft," he said. He added that the administration has the ability to sanction overseas models found to be stealing from American companies 1.

The dispute centers on a technique called distillation, which transfers capabilities from a larger model into a smaller, more efficient one 1. US AI companies have spent months warning that foreign actors copy their technology and redeploy it as open source. In April, the White House said it would work with AI firms to combat the practice 1.

But whether distilling from another company's model constitutes theft is an open question, and the sharpest critique came from Microsoft CEO Satya Nadella. "I find it ironic that the status quo is to then turn around and impose restrictive terms on distillation," he said, pointing out that labs depend on fair use rights to train on publicly available data 1. American labs built their competitive position by training on public content under fair use. Bessent's sanctions threat would draw a line through that same practice when the lab doing the distilling happens to be Chinese.

The legal backdrop blurs any clean narrative. Anthropic, one of the firms that could benefit from sanctions against Chinese competitors, received court approval this week to begin paying out its $1.5 billion settlement after a judge ruled the company illegally downloaded millions of copyrighted books for training 1.

Hugging Face CEO Clem Delangue dismisses distillation as the driver of China's ascent. He credits strong research teams and a more collaborative culture, noting that US companies also practice distillation and that if it were the decisive factor, other countries would have caught up long ago 1.

The science debate will not be settled by Treasury lawyers. What matters for the market is the direction of travel. Washington spent years treating advanced chips as the chokepoint, wagering that cutting off the hardware would slow Chinese AI development 1. Chinese models kept improving. Bessent's statement expands the battlefield from silicon to software. Whether that expansion becomes enforceable policy depends on what the examination actually finds. But for any company whose product depends on a Chinese open-weight model, the planning assumption just changed.