WASHINGTON – The accusation arrived before the evidence. Michael Kratsios, the White House science and technology adviser, said Wednesday that China’s Moonshot AI had covertly extracted capabilities from Anthropic’s Claude Fable 5 through a process called distillation, using outputs from a more powerful AI model to train a weaker one without authorization. Treasury Secretary Scott Bessent followed with a threat: sanctions and export-control blacklisting were on the table. The technical case for either action, based on what the US government has made public, is thinner than the official statements suggest.
Moonshot’s Kimi K3, the model at the center of the accusation, was released in early July and immediately drew benchmarks that make American AI companies nervous. TechCrunch reported that K3 outperforms every major model except Anthropic’s Fable 5 and OpenAI’s GPT-5.6, and that it is open-weight, meaning anyone can download its underlying parameters and run it freely. That last fact is central to the dispute: the US government now appears to be arguing that making a distilled model publicly available compounds the IP harm.
The technical timeline that Kratsios’s accusation depends on is difficult to reconcile with how large AI models are developed. Anthropic made Fable 5 publicly available on July 1. Kimi K3 appeared weeks later. For Moonshot to have conducted a large-scale distillation in that window, running millions of queries against Fable 5, incorporating the outputs into training data, and retraining a model competitive enough to worry the White House would require a compression of the standard development timeline that AI researchers describe as implausible. Distillation at the claimed scale typically takes months, not weeks.
None of this confirms the claim is false. What Kratsios offered publicly was an allegation, not a technical reconstruction of how the model was built. The White House has presumably reviewed intelligence that did not enter the public record. The distillation accusation is also separate from a second claim Kratsios raised: that Moonshot had acquired Nvidia GB300 servers, restricted from export to China, and accessed them in Thailand, a country that does not face the same controls but has drawn scrutiny from US enforcement officials tracking chip diversion. That accusation, if substantiated, carries legal consequences independent of the IP question.
Bessent’s statement on the platform X offered the administration’s clearest articulation of its legal theory. “Open source is not open season on American IP,” Bessent wrote. “When firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table.” The argument is that publishing Kimi K3 as an open-weight model does not resolve the underlying IP violation if the model was built on unauthorized Fable 5 outputs. That legal theory has not been tested in court. Moonshot had no immediate comment.
This is the second major IP dispute of the year linking a US AI company to a Chinese competitor. Eastern Herald reported in June that Anthropic accused Alibaba of running more than 28 million fraudulent exchanges against Claude to extract its software engineering capabilities, a pattern documented in a letter Anthropic sent to the Senate as lawmakers drafted sanctions legislation. In that case, the evidence was specific. The Moonshot case, based on what has been publicly disclosed, does not yet have an equivalent evidential foundation.
The pace of the escalation is notable. Bessent warned on Tuesday that sanctions against Chinese AI companies over IP theft were under active consideration, and within 24 hours Kratsios had named a specific company and a specific model. The speed suggests the administration had already identified Moonshot as its target before the press statements were issued.
What the White House framing does not address is what happens to the researchers and developers outside China who have already downloaded Kimi K3’s parameters and built applications on them. An Entity List designation for Moonshot would restrict American companies from using the model commercially, but the weights themselves are already distributed globally. The enforcement problem that open-weight models pose for export controls, that you cannot recall a publicly available file, is one the US government has not resolved.
Moonshot is simultaneously pursuing a reported $30 billion valuation in a Hong Kong IPO, and Kimi K3’s launch demand was severe enough that the company paused new subscriptions after its GPU infrastructure was overwhelmed within days. A sanctions designation or Entity List addition would move that listing date and damage the valuation, regardless of whether the IP theft accusation is ultimately substantiated. Whether that outcome is a side effect of US policy or part of its purpose is a question neither Kratsios nor Bessent answered on Wednesday.
The broader pattern is one the US government has been building toward since China’s DeepSeek demonstrated earlier this year that competitive AI models could emerge outside the American research establishment at speed. The response to DeepSeek was largely rhetorical. The response to Kimi K3 is a specific accusation of IP theft backed by a Treasury sanction threat. Whether the escalation reflects a genuine assessment of competitive threat, or a policy instrument designed to slow a rival Washington cannot otherwise outrun, will depend on whether the evidence eventually matches the accusation.
