The United States government has accused the Chinese artificial intelligence company Moonshot Ai of illegally using Fable, the most advanced language model from the American company Anthropic, to create its Kimi K3. Not only that: Moonshot Ai is also said to have obtained advanced Nvidia microchips – the company that dominates the global market for artificial intelligence processors – which are subject to trade restrictions.
STATEMENTS FROM THE TRUMP ADMINISTRATION
Michael Kratsios, scientific and technological advisor to President Donald Trump, wrote on X that Washington has obtained information that would prove that Moonshot Ai “distilled” Anthropic’s Fable for the development of K3. “A large-scale, secret industrial distillation aimed at stealing US proprietary technology and undermining American research is unacceptable,” he added.
The Treasury Secretary declared that he is considering placing Moonshot Ai on a sort of “blacklist” of entities subject to trade restrictions and imposing sanctions.
WHAT DISTILLATION IS AND WHAT CHINESE ARTIFICIAL INTELLIGENCE COMPANIES DO
Usually, Chinese artificial intelligence models excel in areas – such as mathematics or programming – where answers are clearly right or wrong. They have more difficulty dealing with open-ended problems or those requiring autonomous and prolonged judgment.
This characteristic is linked to the export controls applied by the United States towards China, which prevent the latter from accessing advanced microchips. Consequently, lacking high computing power and thus unable to compete with Americans in model training, Chinese artificial intelligence companies focus on the inference phase, where the model responds to user requests.
Even in inference, however, Chinese models tend to use data extracted from American systems through a process called “distillation.” Distillation, moreover, was the accusation that OpenAi (the company behind ChatGpt) made against DeepSeek (the Chinese startup widely discussed in early 2025): OpenAi claimed that DeepSeek had improperly used the output of its models to achieve roughly comparable results but at lower costs.
THE ANTHROPIC PRECEDENT
Last February, Anthropic said that three Chinese artificial intelligence companies had illegally used its Claude model to extract features useful for improving their systems: DeepSeek, MiniMax, and Moonshot Ai had generated more than sixteen million interactions with Claude using over twenty thousand fake accounts, violating Anthropic’s terms of service and circumventing regional access restrictions to the model.
THE ARTIFICIAL INTELLIGENCE RACE BETWEEN AMERICA AND CHINA
The Trump administration considers artificial intelligence – due to its transformative potential on the economy and defense – a critical technology for the United States and believes it is engaged in a “race” with China: “if we lose it,” said Interior Secretary Doug Burgum, “it will have a direct impact on our national security.”
Not surprisingly, Anthropic’s public policy lead Sarah Heck, referring to Moonshot Ai’s conduct, wrote on X that “illegal and hostile distillation constitutes intellectual property theft and a form of industrial espionage that strengthens the military and intelligence capabilities of adversary powers. It is,” she added, “a national challenge that poses serious risks to US national security.”
NOT ONLY DISTILLATION: MOONSHOT AI ALSO OBTAINED NVIDIA CHIPS
Kratsios also wrote that Moonshot Ai used servers containing Nvidia artificial intelligence processors (GB300 systems) in Thailand “probably to train its models.” Already in February, the Trump administration had claimed that DeepSeek’s latest model was trained with Nvidia’s most advanced microchips, violating US export controls.
In February, the US State Department had warned to be alert to alleged efforts by Chinese technology companies to steal American intellectual property. The communication stated that “artificial intelligence models developed through clandestine and unauthorized distillation campaigns allow foreign actors to release products that appear to offer comparable performance on certain benchmarks at a much lower cost, but do not fully replicate the performance of the original system.”
