The United States warns of potential sanctions on Chinese AI, while Huang responds defiantly.

      On Tuesday, Treasury Secretary Scott Bessent appeared on television to issue a warning to China. Meanwhile, the CEO of the company driving the AI boom took a contrary stance on the same day.

      The threat

      Bessent indicated that the US would examine Chinese open-source models for potential intellectual property theft, as reported by Bloomberg. This could lead to sanctions against their creators. “This administration supports open-source models, but what we do not tolerate is IP theft,” he stated.

      He cited evidence from watermarks. “We are discovering watermarks from our US large language models on numerous Chinese models,” he remarked. The technical term for this is distillation, which involves training one model using outputs from another. Bessent labeled it as theft and assured that action would be taken within weeks.

      What set it off

      The catalyst for this reaction was Kimi K3, an open-weight model from China’s Moonshot AI. It competes with or surpasses the best models from OpenAI and Anthropic on several benchmarks, all while being significantly cheaper. Its introduction triggered a selloff in chip stocks and renewed anxiety in Washington. For months, US labs have accused Chinese competitors of extracting their models.

      Huang’s rebuttal

      Nvidia’s Jensen Huang, speaking to Axios at a new chip facility in Texas, deemed this panic unfounded. “These Chinese models are impressive,” he stated. “Excellent open-source models should be utilized.” He insisted that American companies should “absolutely” have the freedom to operate them.

      His reasoning is rooted in business. Cheaper, free models increase accessibility to AI. Greater AI usage leads to heightened demand for the chips, data centers, and energy that Nvidia sells. “Free AI should benefit chip sales,” he asserted. He also dismissed concerns that a downloaded Chinese model poses a “backdoor” for Beijing, explaining that companies can secure them in isolated environments.

      The awkward part

      Not everyone agrees with the theft accusation. Distillation is a widespread practice, and US labs engage in it as well. “We acknowledge that distillation is a minimal factor,” remarked Hugging Face CEO Clem Delangue in a TechCrunch interview. He added that it is “a practice shared by all, including American firms.”

      There’s also a degree of irony. Anthropic, one of the fiercest critics, recently had a $1.5 billion settlement approved for using pirated books to train its model, Claude. Huang went even further, encouraging the government to lift restrictions on Anthropic’s own Mythos model. “Let Anthropic operate freely,” he urged.

      What happens next

      Bessent is set to lead the US delegation in AI discussions with China in September. The issue of distillation will be addressed during these talks. As of now, the divide is clear: the government aims to exclude Chinese models, whereas the leading chip manufacturer advocates for their inclusion.