The United States government has formalized accusations against the Chinese startup Moonshot AI, alleging that it used data extracted from the Fable artificial intelligence model, developed by Anthropic, to create its new system version, named Kimi K3.
The complaint was released by Michael Kratsios, who serves as the principal technology official in the White House, via a post on the social media network X. US authorities also pointed out that the Chinese company allegedly obtained servers equipped with advanced Nvidia chips, specifically GB300 units, which were reportedly used outside of China for training artificial intelligence models.
The accusations were disclosed in Washington on Wednesday (22), and the White House maintained that such practice is part of a strategy aimed at the improper acquisition of technologies created by American AI companies.
Michael Kratsios, responsible for the White House's technology division, detailed that Moonshot AI allegedly performed a procedure known as model distillation. This technique allows smaller systems to be trained using responses generated by more sophisticated models. Kratsios publicly stated: 'We have information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model. Large-scale clandestine industrial distillation intended to steal proprietary technology from the United States and weaken American research is unacceptable.'
Distillation is a method used in the field of artificial intelligence to reduce training costs by leveraging the results of more complex systems. According to reports from the US Department of State, this method was allegedly applied by Chinese companies to generate models with similar performance in certain tests, without fully replicating the functionalities of the original systems.
The US government had previously expressed concern over such practices. A diplomatic document cited by Reuters indicated that the State Department had directed an international mobilization to warn about alleged efforts by Chinese companies, including Moonshot AI, to obtain intellectual property from American AI laboratories.
This dispute unfolds amid the rapid advancement of China's artificial intelligence sector. Moonshot AI promoted Kimi K3 as an open-source model containing 2.8 trillion parameters, describing it as one of the largest systems of its kind globally with performance close to that of Anthropic's Fable.
In addition to the allegations regarding model development, the American administration stated that the Chinese company acquired servers equipped with Nvidia GB300 chips and used them in Thailand, possibly for training its systems.
The suspicion regarding the use of advanced Nvidia hardware arises after previous concerns raised by US authorities about the Chinese startup DeepSeek. A former Trump administration official, who preferred not to be identified, mentioned that a recent model from this company was trained with advanced Nvidia chips, which could constitute a violation of export restrictions imposed by the US.
Anthropic, responsible for Fable, is at the center of this conflict for developing some of the most advanced artificial intelligence models. The base text also reports that the Fable and Mythos models were temporarily removed from circulation by the US government due to security concerns.
The United States government is considering imposing restrictions on the use of artificial intelligence models developed in China, motivated by security issues and the need to protect domestic companies.
Although previous attempts at a direct ban on Chinese AIs failed, the possibility of such a measure has resurfaced in recent weeks, driven by the emergence of more advanced tools from Asia. Sources consulted by Axios indicate that while past movements aimed for an explicit prohibition, the current approach may focus on the claim that these Chinese models are not safe.
The relaunch of the Kimi K3 model, created by the Chinese startup Moonshot AI, was the factor that reignited the debate within the Donald Trump administration. This model features a context window of one million tokens, allowing for the processing of very extensive commands. In blind tests, where users evaluate the quality of the response to the same prompt, Kimi surpassed the most recent models offered by Anthropic and OpenAI.
Kimi K3 operates under an open-weights model, meaning the developer discloses the numerical parameters but keeps the code and training process confidential. This raises fears that China could offer more accessible AI solutions with performance comparable to those of American corporations.
There is no consensus among American federal authorities regarding a ban. While one faction advocates combating Chinese AIs, another expresses concern that a restriction could harm free competition and inhibit innovation in the sector. David Sacks, a White House advisor on artificial intelligence matters, argued that the US already has a duopoly in the segment and that certain companies seek the government to eliminate open competitors, implicitly referring to OpenAI and Anthropic.
However, this perspective may be weakening in the Trump administration. Axios points out that advisors linked to the technology and venture capital sectors have lost some of their influence, giving more space to individuals focused on national security in discussions.
Although a direct ban is not seen as the main path by the White House, Axios sources suggest that more likely methods include implementing licensing rules, the threat of inclusion on the notorious Entity List, and conducting public pressure campaigns. Such actions would be directed at American companies, aiming to discourage them from using Chinese AI technologies.
In parallel, China is also analyzing how to position itself in this race for AI development. Beijing is studying the possibility of banning the export of AI models, in addition to planning to create stricter legislation against code leaks and limiting foreign investment in local startups.
