U.S. threatens sanctions on Chinese AI models over alleged IP theft

First it was advanced microchips. Now, Washington is threatening to pull the plug on open-source code itself.

On Tuesday, Treasury Secretary Scott Bessent went on Fox Business to issue a direct warning to Chinese AI labs. He said the U.S. is launching an investigation into Chinese open-source models for potential intellectual property theft, threatening formal sanctions if violations are found.

His remarks follow an Axios report on Monday indicating that the Trump administration is weighing a wholesale ban on Chinese open-source models, although some observers dispute whether such an extreme move will materialize.

Model distillation under the microscope

The dispute centers on model distillation, a widespread technique where smaller models are trained on outputs from frontier systems. Bessent specifically noted that the government will investigate whether Chinese developers used outputs from OpenAI and Anthropic, citing embedded AI watermarks as potential evidence. Official bilateral AI talks between Washington and Beijing are reportedly scheduled for September.

The timing is hardly a coincidence. Chinese labs have been releasing increasingly powerful open-weight architectures, most notably Moonshot AI's Kimi K3. The 2.8-trillion-parameter system rivals top-tier models like GPT-5.6 and Claude, undermining the premium subscription models that American frontier labs rely on to fund their massive capital expenditures.

Yet the claim that distillation constitutes theft is sparking immediate pushback, even within the United States. Microsoft CEO Satya Nadella recently criticized this exact stance, calling it ironic that model providers invoke fair use to train on everyone else's public data while imposing restrictive anti-distillation terms on their own outputs.

The legal standing of these claims also looks shaky given current events. Just this week, Anthropic received final court clearance to start issuing payouts under a massive $1.5 billion settlement for illegally downloading copyrighted books during training.

Hugging Face CEO Clem Delangue also pushed back on the narrative during an appearance on TechCrunch's Equity podcast. He argued that distillation plays a minor role in building high-end AI, noting that if it were a magic shortcut, open-source models would be thriving in dozens of countries. China's real advantage, he said, lies in world-class research talent and an aggressive, collaborative open-source strategy.

Why it matters

If the White House follows through with sanctions or an outright embargo, global developers could be caught in the crossfire. Many teams and startups rely on open-weight models like Kimi K3 to escape high API fees and deploy on private infrastructure. Banning these weights would create immediate compliance headaches for multinational software stacks.

For the broader tech landscape, this signals a major shift in the geopolitical tech war. Restricting high-end hardware failed to stop Chinese labs from matching state-of-the-art reasoning capabilities. Now that Chinese firms are distributing frontier-grade weights for free, they are effectively turning the model layer into a low-margin commodity, starving proprietary U.S. labs of pricing power.

In our view, attempting to police open-source code under the guise of IP protection sets a dangerous precedent. Relying on AI watermarks as legal proof remains technically unproven, and American labs have hardly been transparent about their own training corpora. Rather than a defense of intellectual property, this looks like a political attempt to shield domestic monopolies from open-source pricing pressure.