A White House science adviser has directly accused China’s Moonshot AI of running a “large scale” campaign to steal capabilities from Anthropic’s frontier models and of gaining access to restricted Nvidia servers — the highest-level US government allegation yet against a Chinese AI company for intellectual property theft.
Michael Kratsios, President Donald Trump’s Science and Technology adviser, said in a social media post on X on July 23 that the US government has information showing Moonshot AI “distilled” capabilities from Anthropic’s Fable model to develop its Kimi K3 system. The model, unveiled last week, stunned the tech industry by appearing to narrow the gap between Chinese and Western frontier AI.
🔍 THE BOTTOM LINE
This is not a think-tank paper or a leaked intelligence assessment. This is a named White House official, on the record, accusing a specific Chinese company of both IP theft and export control evasion. The allegation comes one day after Treasury Secretary Scott Bessent threatened sanctions against overseas models that “steal from our great companies,” and weeks before Trump is expected to meet Xi Jinping in September. The US-China AI war has moved from industry accusations to government-level naming.
What is distillation? It’s a technique where a smaller, weaker AI model extracts knowledge from a larger, more capable one by querying it millions of times and learning from the responses. It is not hacking — it uses the target model’s own API. The legal question is whether this constitutes theft, legitimate model compression, or something the law has not yet caught up with. Anthropic calls it theft. China calls it “misguided” to complain about. The US government now appears to be calling it sanctionable.
What Kratsios Actually Said
According to BBC reporting, Kratsios said Moonshot carried out the distillation campaign against Anthropic’s Fable AI — the same model Anthropic has been fighting to protect through export controls since June. He also alleged that Moonshot gained access to restricted cutting-edge Nvidia servers to train its models, a direct violation of US export controls.
The BBC reports that Kratsios posted: “We have information that Moonshot AI distilled capabilities from Anthropic’s Fable AI for the development of its K3 model.” The White House, Anthropic, Moonshot, and Nvidia have all been contacted for comment.
The Bessent Follow-Up
The Kratsios allegation did not come in isolation. Treasury Secretary Scott Bessent told Fox Business on July 22 that the US would examine whether Chinese AI models have stolen capabilities from American companies, BBC reports.
“We’ve seen a lot of talk about open-source models coming and threatening the large language models in the US,” Bessent said. “If we see, especially, that overseas models are stealing from our great companies, we have the ability to sanction them.”
This is the first time a Treasury Secretary has publicly floated sanctions against a Chinese AI company for distillation. The Treasury Department’s sanctions authority — typically used against nation-states, terror groups, and drug cartels — would represent a dramatic escalation if applied to a Chinese AI startup.
How This Connects to the Broader Distillation Fight
The accusation builds on months of escalating tension. Anthropic’s June letter to the Senate Banking Committee accused Alibaba of running 25,000 fake accounts and extracting 28.8 million dialogue exchanges from Claude over 44 days. OpenAI, Anthropic, and Google have been sharing intelligence through the Frontier Model Forum to coordinate a clampdown since April.
China’s government formally rejected the accusations on July 18, calling them “misguided” and framing distillation as legitimate model compression. The Global Times argued that US complaints about distillation are “technological hegemony anxiety” — an attempt to kick away the ladder after Western labs built their own capabilities on permissive data regimes.
What changes today is the source of the accusation. When Anthropic alleged distillation, it was a company protecting its IP. When Kratsios alleges it, it is the US government building a case for sanctions. The shift from corporate complaint to government posture is the story.
The Nvidia Chip Allegation
The second prong of Kratsios’s accusation — that Moonshot accessed banned Nvidia servers — is arguably more serious than the distillation claim. Distillation occupies a legal grey area. Export control violations are unambiguous crimes under US law.
Bloomberg reported that Kratsios said Moonshot improperly used US AI models and Nvidia chips to create the Kimi K3 system. The US has maintained a complex regime of chip export restrictions targeting China since 2023, with restrictions tightened and eased multiple times depending on diplomatic conditions.
If Moonshot trained Kimi K3 on restricted Nvidia hardware, it would represent a failure of the export control system at exactly the moment the US government is trying to demonstrate that system works. It also raises the question: if a Chinese startup can access banned chips to train a model that rivals US frontier systems, what exactly are the controls controlling?
The September Trump-Xi Factor
The BBC notes that the increased scrutiny comes as Trump is expected to meet Chinese President Xi Jinping in September. The timing of the Kratsios allegation — days after Bessent’s sanctions threat, weeks before a summit — suggests a deliberate escalation strategy. The US may be building a public case to bring to the negotiating table: we know what Moonshot did, we have the evidence, and we have the authority to sanction.
Whether that strategy produces results or provokes retaliation is the open question. China has its own levers — rare earth export controls, its own AI models that now rival American ones, and a domestic market large enough to sustain its AI industry regardless of US sanctions. The September meeting may determine whether the AI cold war stays cold.
NZ Angle
New Zealand sits on the periphery of this conflict, but the implications are real. If the US sanctions Chinese AI companies, NZ companies using models like Kimi K3 — or any model built on distilled US technology — could face compliance questions. NZ’s sovereign AI strategy depends on understanding which models are clean and which carry legal risk. The distillation fight is not just a US-China story. It is a supply chain integrity story that every AI-consuming country needs to track.
❓ FAQ
What exactly is Moonshot accused of? Two things: (1) running a “large scale” distillation campaign to extract capabilities from Anthropic’s Fable model, and (2) accessing restricted Nvidia servers to train its Kimi K3 model. Both allegations come from White House science adviser Michael Kratsios.
Is distillation actually illegal? The legal status is contested. Anthropic calls it theft. China calls it legitimate model compression. There is no settled international law on whether querying a model’s API millions of times to train a competitor constitutes IP infringement. The US Treasury Secretary is now suggesting it could trigger sanctions.
What is Kimi K3? It is Moonshot AI’s flagship model, unveiled in July 2026. It gained global attention for apparently narrowing the gap between Chinese and Western frontier AI models. If it was built using distilled Anthropic capabilities and banned Nvidia chips, as Kratsios alleges, its competitive position becomes legally and diplomatically untenable.
Could this lead to sanctions? Treasury Secretary Bessent explicitly said the US “has the ability to sanction” overseas models that steal from US companies. Whether the government follows through depends on the evidence, the diplomatic calculus around the September Trump-Xi meeting, and whether the threat itself is enough to change Chinese behaviour.
What does this mean for NZ? NZ companies adopting Chinese AI models need to consider whether those models carry legal risk. If the US sanctions Chinese AI firms, using their models could create compliance exposure. NZ’s sovereign AI strategy needs to account for which models are clean and which are tainted by distillation or export control violations.
