Jensen Huang: 'US companies should be able to use Chinese AI models'

While the United States accuses Moonshot AI of having distilled Kimi K3 from Claude Fable of Anthropic, NVIDIA CEO Jensen Huang fully opens up to the entry of Chinese artificial intelligence models in the USA. In an interview with Axios during the inauguration of a new phase at the Wistron plant in Fort Worth, Texas, dedicated to AI infrastructure production for NVIDIA, the manager stated that American companies should "absolutely" be able to use Chinese models.

For the US AI ecosystem, the debut of Kimi K3, the new open-weight model developed by Moonshot AI, was almost like the "DeepSeek moment" of early 2025. According to initial independent assessments, Kimi K3 indeed ranks at the top of benchmark rankings, being inferior only to the most advanced frontier models such as Claude Fable 5 and OpenAI GPT-5. In some benchmarks, while in other tests dedicated to programming, it reached the first position.

Although it does not represent the most powerful model available, it is noted to be comparable to GPT-5.5 and Claude Opus 4.8, offering similar performance at an estimated cost of about one third compared to competitors. The announcement of Kimi K3 had an immediate impact on financial markets, but as it was with DeepSeek, Huang claims investors are misinterpreting the significance of the arrival of increasingly economical and efficient open models. The CEO contests the thesis that cheaper models reduce the demand for infrastructure; rather, they increase it, as they allow a greater number of companies and developers to adopt artificial intelligence.

The manager argues that a broader diffusion of AI inevitably ends up increasing the need for accelerators, servers, and data centers. In other words, the availability of open models would represent an advantage for the entire hardware supply chain, including NVIDIA, rather than a threat to the sector. Huang also dismissed concerns that Chinese models could contain hidden mechanisms exploitable by Beijing. According to the NVIDIA CEO, this is an idea based on a false assumption, as open models can be downloaded, analyzed, modified, refined, and even run completely isolated from the Internet. This transparency, he argues, makes it possible to identify potential vulnerabilities and correct them more quickly.

According to Huang, the same principle should also be applied to models developed in the United States. The manager criticized the limitations imposed by the U.S. government on some advanced models, such as Anthropic Mythos and Fable 5, as well as the pressures exerted on OpenAI regarding the release of GPT-5.6. In his view, instead of initially limiting their availability, it would be preferable to make them accessible and progressively improve their security through public audits and continuous fixes.

Huang also believes that open models have a strategic role in areas such as scientific research and cybersecurity. "If everything became a single model, a single point of attack, and a single source of error, the world would be much more vulnerable," he stated, arguing that the availability of more independent models makes the entire ecosystem more resilient.

In any case, in the coming days it will be possible to evaluate the real capabilities of Kimi K3 more accurately when Moonshot AI makes the complete weights of the model public and the community can independently verify its performance and characteristics. According to some analyses, the gap between the best open models and proprietary ones has now shrunk to about three months, a scenario that could make it increasingly difficult to effectively limit their dissemination through regulatory bans.