
Jensen Huang Calls AI Model Distillation Competition
Published by AINave Editorial • Reviewed by Ramit
Nvidia CEO Jensen Huang says AI model distillation is competition, even as U.S. officials and AI companies have described particular alleged uses of model outputs as theft or violations of service terms. The distinction matters: distillation describes a training technique, while the allegations concern how some developers may have accessed models.
Distillation as a form of competition
In the account of Huang’s CNBC interview, distillation means training smaller models using outputs generated by larger ones. Asked whether using a competitor’s outputs amounts to robbery, Huang replied, “That’s called competition.” He compared learning from model outputs with rivals taking apart Nvidia chips to understand how they work, arguing that competition can push the industry to innovate faster. His comparison and comments make a case for treating the practice as part of ordinary rivalry, rather than as inherently improper.
That argument describes distillation in general. It does not settle whether a particular effort respected a model provider’s terms or the conditions under which its outputs were accessed.
The dispute is also about access
The reported counterposition is not simply that training on another model’s outputs is always theft. Treasury Secretary Scott Bessent branded distillation as theft and raised the prospect of sanctions against foreign entities using U.S.-developed models. CISA accused Chinese companies of running large-scale distillation operations that breached terms of service, while Anthropic flagged Alibaba and DeepSeek for allegedly distilling its systems illicitly. These are accusations, not established findings in the account; it also says China denies the claims. The reported positions and allegations point to a practical dividing line: how outputs were obtained can matter as much as how they were used in training.
Huang’s proposed response focuses on that access. He said companies that want to prevent others from using their outputs should identify who is accessing their services and restrict access. He did not specify particular controls, so the suggestion is a general direction, not a detailed technical solution.
For model providers, that leaves a tension between making a service available and limiting use they consider abusive. Calling distillation competition does not resolve allegations of terms-of-service breaches; calling alleged conduct theft does not make every form of distillation equivalent to it.



















