U.S. Accuses China of Stealing K3 AI Model from Anthropic

What happened?

On July 23, 2026, Michael Kratsios, science advisor to Donald Trump, posted a direct accusation on X (formerly Twitter): the Chinese company Moonshot AI allegedly created its Kimi K3 model—an open model with 2.8 trillion parameters—by distilling Anthropic's Fable model, a technique that involves making massive queries to one model to train another. Kratsios claimed that Moonshot AI developed a sophisticated internal platform to conduct large-scale distillation against U.S. models, rapidly switching between access methods to avoid detection. Furthermore, he accused the company of accessing servers with Nvidia GB300 accelerators, which are banned for China, and using data centers in Thailand to bypass restrictions.

Treasury Secretary Scott Bessent seconded the accusations with a message on X: “We support open source, but open source is not open season on American intellectual property. When PRC companies conduct covert distillation attacks that cross the line into intellectual property theft, sanctions and entity list additions are on the table.”

Why does it matter?

The K3 case is not an isolated incident. Anthropic had already accused Moonshot AI, DeepSeek, and MiniMax of distillation in February 2026. However, the novelty is that the accusation now comes directly from the White House, elevating the conflict to a diplomatic and national security level. Kimi K3 has caused concern among U.S. investors: following its launch on July 16, shares of AI companies fell amid evidence that China is catching up—or even matching—U.S. frontier models.

The case raises fundamental questions about the limits of distillation, a technique that can be legitimate for creating smaller, more efficient models, but which according to Kratsios has become “covert industrial theft” when used on a large scale and clandestinely. It also exposes the porosity of tech sanctions: despite restrictions, China has found ways to evade them, such as renting GPU farms abroad.

What will be the consequences?

Immediate consequences could include economic sanctions against Moonshot AI and possibly its inclusion on the U.S. Department of Commerce's entity list. In the medium term, it is likely that export controls on AI hardware, including the GB300, will be tightened, and measures against the use of data centers in third countries to circumvent bans will be reinforced. It could also accelerate the fragmentation of the global AI ecosystem, with separate technological blocs and divergent standards.

For companies and users, the K3 case underscores the need to review open model usage policies and licenses. If large-scale distillation is considered illegitimate, developers using distilled models could face legal risks. Additionally, Kratsios's accusation that Moonshot AI accessed prohibited hardware could lead to greater surveillance of GPU supply chains in Asia.

What should readers know?

First, that distillation is not inherently illegal; in fact, it is a common industry practice to optimize models. The difference lies in the scale and secrecy: using distillation to replicate a competitor's model without authorization and by evading access controls is what the U.S. considers unacceptable. Second, that China has systematically denied these accusations, and so far there is no conclusive public evidence. The Australian Strategic Policy Institute ranks China as a leader in 66 out of 74 technologies, suggesting its progress does not depend exclusively on espionage. Finally, that this case could set a legal and regulatory precedent for the use of open models and distillation, affecting startups and large companies alike.

“Large-scale, covert distillation intended to steal proprietary U.S. technology and undermine U.S. research is unacceptable,” stated Michael Kratsios.

In summary, the U.S. accusation against Moonshot AI regarding the K3 model is a turning point in the technological rivalry between the two powers. It opens a debate on the ethical and legal limits of distillation, exposes cracks in the sanctions regime, and anticipates a tightening of restrictions. For readers of TheVortiq, the lesson is clear: in the race for AI, the line between open collaboration and intellectual property theft is becoming increasingly blurred, and the consequences for the global ecosystem will be profound.