WASHINGTON — Scott Bessent, the US Treasury secretary, said Tuesday that the Trump administration was prepared to impose sanctions on China if evidence confirmed that Chinese artificial intelligence developers had trained their models on stolen American intellectual property. The warning arrived three days after Moonshot AI’s Kimi K3 model attracted international attention for performance metrics that independent researchers say outpace what its publicly disclosed computing resources should theoretically produce.
Bessent made the remarks in an interview with CNBC, saying the administration was scrutinizing open-source AI models for intellectual property violations. The specific concern, as articulated by Treasury, is that Chinese AI companies may have used weights, training outputs, or proprietary fine-tuning derived from American frontier models, effectively incorporating years of American research investment into a competing product that is then released freely, undercutting the commercial position of the firms whose work was absorbed without authorization.
The accusation follows a pattern of complaints from American AI companies about what they describe as systematic extraction of their work through derivative training methods. A common technique, sometimes called distillation, involves feeding a proprietary model’s outputs into a new training pipeline, then presenting the resulting system as an independent development. The legal standing of this practice is contested. No court has definitively ruled whether training a model on the outputs of a proprietary system constitutes copyright infringement, trade secret misappropriation, or competitive reverse engineering that existing law does not prohibit.
What is new in Bessent’s statement is the administration’s willingness to connect AI intellectual property concerns to the sanctions toolkit. This marks the first time a senior US Treasury official has publicly done so. Sanctions are among the most powerful economic instruments available to Washington, used to restrict access to dollar-clearing systems, freeze assets, and cut institutions from the global financial network. Extending that mechanism to AI companies would represent a significant escalation beyond the chip export controls the Commerce Department has maintained since 2022.
Kimi K3 has become the focal point of these concerns without Moonshot AI appearing in any official US government document by name. The model, launched by the Beijing startup last week, is described by the company as the world’s first open-weight 2.8-trillion-parameter model with a 1-million-token context window. Its performance on coding and reasoning benchmarks placed it competitively against closed frontier models from American companies. Demand surged so quickly after launch that Moonshot paused new subscriptions, citing GPU compute limitations.
The concern centers on a gap between performance and compute. Independent researchers noted that achieving Kimi K3’s benchmarked capabilities through standard training from scratch would typically require more computational infrastructure than Moonshot is known to operate. The gap has prompted discussion that the model may have incorporated training signal derived from competitor models or other materials not disclosed in Moonshot’s documentation. Moonshot AI has not addressed the IP questions directly, attributing the model’s performance to architectural innovations and efficiency improvements in its training process.
Bessent said the administration would scrutinize such models and act if evidence of theft is confirmed. He did not specify what evidence threshold would trigger sanctions, which agency would make that determination, or what timeline a decision would follow. The statement leaves substantial room for both escalation and retreat, and the administration’s track record on technology policy suggests that warnings precede action by variable intervals depending on diplomatic conditions.
The broader US-China technology confrontation has already reshaped the global semiconductor supply chain through successive rounds of export restrictions on Nvidia and other chipmakers. Those restrictions have been only partially effective. Nvidia has tightened its Asia buyer list under US government pressure over suspected chip leakage to Chinese-linked entities, yet Chinese institutions have continued accessing restricted hardware through third-country routes that compliance reviews have caught only unevenly. A sanctions regime targeting AI models would be conceptually different, aimed at the software rather than the hardware, and would raise immediate questions about how model weights, which can be downloaded and redistributed digitally, could practically be restricted.
The broader question Bessent’s statement raises is what a sanctions determination would actually require. Establishing that a specific commercial AI model incorporated specific proprietary data demands technical investigation that government agencies have not previously conducted at scale. Whether Treasury or any other department has the analytical capacity to make that determination at the rigor a formal sanctions designation would require is unclear. The administration has not described an investigative process or named a lead agency for the review.
China’s government has consistently framed American AI restrictions as technological protectionism dressed in security language. Bessent’s remarks will do little to alter that characterization. Whether the threat of sanctions represents a genuine policy direction or a negotiating position in the broader US-China economic relationship is a question Tuesday’s statement left open. What is not in question is that a US Treasury secretary placing AI intellectual property on the sanctions agenda is a development without precedent in the technology competition between the two countries. Washington has added a new instrument to the table. What it does with it remains to be seen.
