US Treasury Secretary Bessent Threatens Sanctions on Chinese AI Models: Cites Detection of US LLM "Watermarks" as Kimi K3 Becomes Focus of Discussion

In an interview with Fox, Bessent clearly articulated the administration's stance: "We have seen a lot of discussion about how open-source models are threatening American large language models. This administration supports open-source models, but what we do not support is intellectual property theft. If we find—especially that overseas models are stealing the fruits of our great companies—we have the power to impose sanctions on them."

He revealed that the US government has already detected the "watermarks" of American large language models on many Chinese AI models, stating, "This is unacceptable, and we will thoroughly investigate this matter in the coming days or weeks."

Bessent also floated a potentially more potent regulatory direction: Should US companies using Chinese AI models be required to disclose this fact to their customers? He added a telling remark—"You cannot use counterfeit goods."

The theft referred to by Bessent is technically known as "distillation"—an AI training method that uses a smaller, less capable model to build itself using the outputs of an existing, more powerful model.

He alleged that Chinese models have made "hundreds of millions of calls" to U.S. large language models, "attempting to reconstruct the models' code and thought processes, which is unacceptable."

This allegation is not without foundation; rather, it aligns with recent collective statements from U.S. AI companies. OpenAI and Anthropic have been accusing Chinese competitors of systematically and without authorization using the outputs of leading U.S. models to develop rival chatbots at an extremely low cost—a practice known as "adversarial distillation."

This April, the Trump administration vowed to take further measures to address this issue, including holding foreign actors accountable. A more specific case is: Anthropic wrote to the U.S. Senate Committee on Banking, Housing, and Urban Affairs last month, accusing Chinese tech company Alibaba of launching the "largest known distillation attack" against it to date.

The trigger for this AI sell-off was the new model, Kimi K3, released last week by Chinese AI startup Moonshot AI. This open-weight model outperformed products from OpenAI and Anthropic in certain industry benchmarks, directly sparking market panic over whether Silicon Valley's astronomical AI infrastructure spending can pay off as scheduled—driving a sharp sell-off in US tech stocks last week.

According to data from Artificial Analysis, the cost per "Intelligence Index Task" for Kimi K3 is only $0.95, compared with $1.04 for OpenAI's GPT 5.6 Sol (max) and as high as $2.75 for Anthropic's Claude Fable 5 (including fallback mechanisms). In other words, the cost per task for Anthropic's flagship model is nearly three times that of Kimi K3.

Next, there are three key dates worth watching closely:

Whether the results of the "watermark" review promised by Bessent will point to specific targets of sanctions.

The summit between the heads of state of China and the US in about two months—where AI-related topics are highly likely to be put on the negotiating table.

The special US-China AI talks in September—where Bessent will represent the US side, according to a Reuters report.

From a market perspective, last week's panic selling of tech stocks has proven that the narrative of "low-cost Chinese models" poses a very real threat to the US AI sector. Now that the regulatory hammer has been raised, regardless of whether sanctions are ultimately implemented, this battle over the intellectual property of open-source models will become one of the core variables driving valuation volatility for US and Chinese tech stocks in the second half of the year.

This content was translated using AI and reviewed for clarity. It is for informational purposes only.