White House Alleges Moonshot Used Claude Fable and NVIDIA GB300 Systems to Build Kimi K3
Michael Kratsios, Director of the White House Office of Science and Technology Policy and President Donald Trump’s chief science and technology advisor, has accused Chinese artificial intelligence developer Moonshot AI of using outputs from Anthropic’s Claude Fable to develop its recently released Kimi K3 model. Kratsios said the US government has information indicating that Moonshot created a sophisticated internal platform capable of conducting large scale distillation against American models while rapidly switching between different access methods to avoid detection.
We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model.
— Director Michael Kratsios (@mkratsios47) July 22, 2026
To do this they developed a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of…
Kratsios also alleged that Moonshot acquired servers equipped with NVIDIA GB300 accelerators and accessed additional GB300 infrastructure located in Thailand, likely for training its artificial intelligence models. However, the official did not identify the server operators, disclose where Moonshot’s reported systems were located, or release technical documentation showing how the hardware was used. The statement described the government’s findings as information rather than publicly verified evidence, leaving the central claims unconfirmed outside government and industry sources.
The accusation builds on an earlier Anthropic investigation published in February 2026. Anthropic claimed that Moonshot generated more than 3.4 million exchanges with Claude through hundreds of fraudulent accounts and multiple access pathways. According to Anthropic, those requests targeted agentic reasoning, coding, data analysis, computer use, and computer vision, with later activity attempting to reconstruct Claude’s reasoning traces. Anthropic attributed the activity to Moonshot through request metadata that reportedly matched public profiles belonging to senior company employees.
That earlier campaign provides context for the White House allegation, but it does not independently prove that Claude Fable outputs were specifically incorporated into Kimi K3. Fable 5 launched on June 9, 2026, became unavailable on June 12, returned globally on July 1, and Kimi K3 was officially introduced on July 16. The narrow timeline has led some researchers to question how much Fable specific distillation could have influenced a model as large as K3, although an existing extraction platform could theoretically redirect traffic toward a newly released model.
Moonshot has not disclosed the complete training data or computing infrastructure used to develop Kimi K3. The company describes K3 as a 2.8 trillion parameter open weight model with native vision, a 1 million token context window, and a Mixture of Experts architecture that activates 16 of 896 experts for each token. Moonshot says Kimi Delta Attention and Attention Residuals improve scaling efficiency, but a complete technical report covering the model’s training process has not yet been published.
The GB300 claim could also create new scrutiny around US export controls. NVIDIA’s Blackwell Ultra systems represent some of the company’s most advanced AI infrastructure, with each GB300 NVL72 rack combining 72 GPUs, 36 Grace CPUs, and high bandwidth NVLink connectivity. Anthropic and NVIDIA’s Claude deployment on Microsoft Azure, demonstrating the scale of infrastructure involved. Accessing computing capacity in Thailand does not by itself establish that export rules were violated, but ownership, control, customer identity, and the final destination of the workloads could become central to any government investigation.
US Treasury Secretary Scott Bessent has warned that sanctions and Entity List restrictions could be considered when covert industrial scale distillation crosses into intellectual property theft. Moonshot had not issued a detailed public response to the latest allegations at the time of publication.
The most important distinction is between an allegation supported by undisclosed government information and a conclusion demonstrated through publicly available evidence. Kimi identifying itself as Claude during an isolated conversation would not prove distillation because language models can repeat incorrect identity statements, imitate training text, or produce outputs influenced by prompts.
Anthropic’s earlier discovery of 3.4 million Moonshot related exchanges presents a more substantial foundation for investigating capability extraction, but connecting that campaign directly to Fable 5 and Kimi K3 requires additional technical evidence. Training records, request timelines, account metadata, infrastructure logs, model fingerprints, and documented access to GB300 systems would provide a stronger basis for evaluating the claim.
The wider issue extends beyond Moonshot. Advanced models can now be accessed through international cloud providers while high performance GPUs operate inside data centers across multiple jurisdictions. This makes controlling the movement of physical chips only one part of the challenge. Governments must also determine how export rules apply when restricted companies remotely access powerful infrastructure located outside China.
Should the US government release more technical evidence before considering sanctions against Moonshot AI over Kimi K3?
