Treasury Secretary Scott Bessent threatened sanctions against Chinese AI developers on Tuesday, saying the Trump administration is investigating whether their models were built with intellectual property taken from US rivals through distillation.
The immediate test case is Moonshot AI, the Beijing lab led by Yang Zhilin that released Kimi K3 on July 16th and plans to publish the model's full weights on July 27th. Bessent did not name Moonshot AI, DeepSeek, MiniMax or any other potential target during the Fox Business interview described by Business Insider. He also offered no technical evidence or legal authority behind the sanctions warning.
Yang, a Carnegie Mellon University Ph.D. and former Google Brain and Meta AI researcher, co-authored the Transformer-XL and XLNet research before founding Moonshot AI in 2023. His lab has built its distribution strategy around downloadable models that developers can inspect, modify and host themselves. Kimi K3, if Moonshot AI ships the weights on schedule, would bring that strategy closer to the capabilities sold through closed systems from Anthropic and OpenAI.
Bessent's intervention gives those closed-model developers a federal policy tool against a competitor that is pressuring their pricing and control over distribution. It also compresses several distinct issues - model distillation, terms-of-service violations, copyright, trade-secret theft and sanctions law - into a single allegation of theft.
A corporate complaint becomes federal policy
The factual foundation for Bessent's warning comes largely from US AI labs. On February 23rd, Anthropic said it had identified campaigns by DeepSeek, Moonshot AI and MiniMax that generated over 16 million Claude exchanges through roughly 24,000 accounts.
Anthropic attributed over 3.4 million exchanges to Moonshot AI. According to Anthropic, Moonshot AI used hundreds of fraudulent accounts and multiple access routes to collect outputs involving coding, agentic reasoning, data analysis, computer use and vision. Anthropic said request metadata matched public profiles of senior Moonshot AI employees.
Those are detailed allegations from an interested party, rather than findings issued by a court or regulator. Anthropic has not published the underlying account records, network indicators or model comparisons needed for outsiders to reproduce its attribution. Its February report also did not publicly establish that Kimi K3, which Moonshot AI announced nearly five months later, was trained using the Claude outputs at issue.
Distillation itself is a standard training method. Anthropic acknowledges using it legitimately when a developer trains a smaller model from its own larger system. Anthropic's complaint concerns the means of access: fake accounts, proxy services, regional restrictions and systematic extraction that allegedly violated Claude's terms.
Bessent erased that distinction in his television formulation, describing distillation as a technical term for theft. He said investigators were finding US model "watermarks" in Chinese systems, without identifying the models, explaining what constituted a watermark or publishing a detection method. He said Treasury could impose sanctions, though no specific sanctions program or designation standard was identified.
A recent Justice Department prosecution shows the evidence normally associated with AI trade-secret theft. In January, a federal jury convicted former Google engineer Linwei Ding after prosecutors presented evidence that he had taken over 2,000 pages of confidential AI infrastructure material from Google's network. The government has disclosed no comparable evidence involving Moonshot AI or Kimi K3.
Kimi K3 raises the commercial stakes
Moonshot AI says Kimi K3 has 2.8 trillion total parameters, native vision and a 1-million-token context window. The model is already available through Moonshot AI's app, coding product and API, with full weights due on July 27th. Moonshot AI charges $3 per million cache-miss input tokens and $15 per million output tokens through its official API.
The benchmark story is less definitive than the political reaction suggests. Moonshot AI reports competitive results against Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol across coding and agentic tasks, while its own launch post says Kimi K3 still trails those proprietary models overall. Several results use different agent harnesses, internal tests or special evaluation configurations. The weights and technical report have yet to be published, limiting independent verification.
Kimi K3 still creates a distribution problem for closed labs. A capable downloadable model can be hosted by multiple providers, adapted without seeking permission from Moonshot AI and deployed inside infrastructure that never sends customer data to an external model vendor. That weakens the pricing and platform control available to API-only developers.
The market has already treated Kimi K3 as a threat to the economics of US frontier labs. RuntimeWire reported on July 20th that an IG analyst estimated Kimi K3 had erased $314 billion from private valuation estimates for OpenAI and Anthropic, although the analyst did not disclose the methodology behind that figure.
A Washington Post editorial published before Kimi K3's release warned that closed AI labs have a commercial interest in encouraging Washington to restrict open-weight competitors. Bessent named Meta, NVIDIA and Reflection AI as evidence that the administration supports American open models. Those projects have yet to displace the most capable Chinese open-weight releases.
The domestic IP record complicates the message
The administration's theft framing arrived one day after a federal judge approved Anthropic's $1.5 billion settlement with authors whose books were obtained from pirate libraries. The court had previously ruled that using lawfully acquired books to train Claude qualified as fair use, while leaving Anthropic exposed over the acquisition and retention of millions of pirated copies. The settlement resolved that remaining dispute.
That case drew a line between a training technique and the conduct used to obtain training material. Anthropic makes a related argument against Chinese labs: distillation can be legitimate, while industrial extraction through deceptive accounts and prohibited access can violate contracts and potentially other laws.
Bessent has jumped ahead of the public evidence required to apply that distinction. No Chinese developer was named. No affected US model was identified. No watermark analysis was released. Moonshot AI's Kimi K3 weights, training report and full evaluation details are still scheduled for publication six days after the sanctions threat.
The timing gives proprietary US labs protection before independent researchers can examine Moonshot AI's release. It also recasts a competition problem as an enforcement question. If Treasury acts, its case will need to show which protected material was taken, how it entered a Chinese model and which sanctions authority converts that conduct into a block on US dealings. Bessent's interview supplied none of those elements.