White House Accuses Moonshot AI of Stealing Anthropic’s Model Capabilities

Michael Kratsios, President Trump's Science and Technology adviser, publicly accused Beijing-based Moonshot AI on Wednesday of running a large-scale campaign to steal capabilities from Anthropic's Fable AI model through a technique called distillation. The accusations target Moonshot's Kimi K3 model, unveiled last week to significant attention for reportedly closing the gap with leading Western AI systems. Kratsios also alleged Moonshot trained its models on restricted Nvidia GB300 Grace Blackwell servers, hardware the US barred from export to China in 2022.

What happened

Detail	Fact
Accused party	Moonshot AI, Beijing-based developer of the Kimi K3 model
Alleged victim	Anthropic’s Fable AI model
Technique alleged	Distillation: extracting answers from a stronger model to train a weaker one
Hardware alleged	Nvidia GB300 Grace Blackwell servers, restricted from China export since 2022
Accuser	Michael Kratsios, Trump’s Science and Technology adviser, via a post on X
Kimi K3 open-source release	Scheduled for 27 July 2026

Kratsios stated on X that the US government holds information confirming Moonshot AI distilled capabilities from Anthropic’s Fable AI during development of its Kimi K3 model. Distillation, for those unfamiliar, is a process where a less capable model is fed large volumes of outputs from a more capable one, effectively teaching itself using the stronger model’s knowledge without paying for or licensing it.

Kimi K3 attracted global attention when Moonshot unveiled it last week. The company claimed the model can rival top US AI systems, and independent observers noted it appeared to meaningfully narrow the performance gap between Chinese and Western AI.

The hardware accusation

Kratsios went further than the IP theft allegation. He said Moonshot likely trained its models on servers powered by Nvidia’s GB300 Grace Blackwell computing platform. The US restricted the export of Nvidia’s most advanced chips to China in 2022, citing concerns they could support Chinese military applications. Governments worldwide have since cracked down on chip smuggling to enforce those controls.

The BBC has contacted Moonshot, Anthropic, the White House, and Nvidia for comment. None had responded publicly at the time of reporting.

Why it matters

This is not an isolated accusation. In June 2026, Anthropic accused Alibaba, the Chinese e-commerce and tech giant, of illicitly extracting capabilities from its Claude AI models using the same distillation method. Anthropic subsequently urged Congress to penalise companies behind such attacks and strengthen protections against US technology theft.

The White House signalled its intent back in April, when Kratsios issued a memo stating the administration would work more closely with US AI companies to combat what it called “industrial-scale campaigns” by foreign rivals to access proprietary information.

Now Treasury Secretary Scott Bessent has added financial pressure to the warnings. On Tuesday he said the US would examine whether Chinese AI models have stolen American capabilities. A day later, he confirmed sanctions are under consideration for companies that “cross the line” into IP theft. His framing was pointed: “We support open-source AI and the innovation it unlocks. But open source is not open season on American IP.”

The timing matters politically too. Trump is expected to meet Chinese President Xi Jinping in September, and these public accusations could shape the agenda of those talks. For context on how US-China tensions around AI spending and competitive pressure are evolving, see our coverage of Bessent’s broader sanctions threat against Chinese AI model makers.

Our take

Distillation as a form of capability theft is a genuinely difficult problem to police. A company running queries against a public API or a released model and using the outputs as training data looks, on the surface, like ordinary usage. The line between legitimate benchmarking and systematic extraction is blurry, and proving intent or scale requires the kind of internal telemetry that only the model provider has access to. Anthropic clearly has some of that data, which is why both they and the US government feel confident enough to name names publicly.

For businesses evaluating AI vendors or building on top of AI APIs, this matters in one practical sense: the legal and compliance environment around AI model usage is tightening fast. Using a third-party model’s outputs as training data for your own system without a clear licence to do so could expose you to the same accusations, even at a small scale. If you are building any kind of AI integration that feeds one model’s outputs into another, review your terms of service now.

Moonshot releasing Kimi K3 as open-source on 27 July also creates a different problem. Once it is freely downloadable, any further accusations about how it was built become largely academic. The model is out. The question becomes whether the US escalates with sanctions before or after that date.

What to do about it

1. Review the terms of service for any AI APIs your business calls, specifically the clauses on using outputs to train other models.
2. If you use a Chinese-developed AI model or plan to after Kimi K3 drops on 27 July, flag the compliance risk with your legal team before integrating it into a product.
3. Watch the Bessent sanctions announcement closely. If sanctions land before 27 July, the open-source release timeline may shift.
4. Contact your AI vendor to ask whether their models include provenance logging, so you can demonstrate your own usage is legitimate.

If the US does impose sanctions on Moonshot or similar firms, any business that has built workflows on top of their APIs will need a contingency model. Start mapping alternatives now.

Source: Bing News · Anthropic