Nvidia Would Stop Chip Sales If It Knew AI Was Harming People

On whether Nvidia would stop selling chips to an AI lab developing something he considered too dangerous, Jensen Huang said yes—but he framed the answer as a hypothetical and set limits on what Nvidia could judge. Before stopping sales, he said, the company would have a conversation. Nvidia does not run its customers’ companies, and its technology is general-purpose; imposing Nvidia’s values on another company, institution or country is, in his words, largely out of bounds.

The boundary he did draw was knowledge of harm: if Nvidia knew for certain that an application of its technology was bringing harm to people, Huang said, it would stop selling. He did not describe a broader system for reviewing customers or determining where to draw the line. His answer pairs a willingness to act in that specific circumstance with caution about claiming authority over others’ uses of general-purpose technology.

That distinction matters to his wider position on safety. Huang argues that developers have agency over how AI advances, but says responsibility for safe development cannot be separated from the work of building more capable systems.

Responding to Emily Chang’s question about prominent figures calling for a slowdown, Huang called safety “paramount.” He argued that AI’s potential depends on developing it safely and responsibly, but rejected the idea that safety and capability should be treated as separate conversations.

His prescription was to accelerate several kinds of work alongside AI development: what he called “open shell,” the integration of containment technology into AI systems, and monitoring technologies, including AI systems that monitor other AI. Huang’s argument is that people build and engineer the technology, and therefore have the ability to improve it and influence how it develops. In his account, safety measures are not a reason to stop advancing AI; they are part of advancing it responsibly.

That emphasis on agency also informs his definition of sovereign AI. Rather than requiring a country to build every component itself, Huang defines sovereignty by what a country can control and whether it can depend on continued access.

Asked whether AI can be sovereign when it runs on American chips, Huang said yes. He defines sovereignty as control over a society’s own data: people’s life experiences, a company’s knowledge, a country’s culture and values, and the ability to build more data, experience and intelligence from them. “The computer is just an instrument,” he said.

He added a second condition: no one should be able to deny a country service. Huang compared AI to electricity, the internet and water, describing them as critical infrastructure for societies and companies. A country may need collaboration, partnerships and support from others, he said, while retaining agency and the ability to determine its own direction.

In this account, dependence on outside technology does not by itself rule out sovereignty. The question is whether a country controls its own data and can rely on access to the infrastructure it needs. Huang’s confidence in continued AI development also rests on a commercial claim: he says the technology has moved from potential toward useful, profitable work.

Huang described OpenAI and Anthropic—both Nvidia customers and companies in which he is an investor—as the fastest-growing technology companies in history. He said their revenues and profits were real, and that AI had changed over the preceding six months: from “unprofitable, unuseful” and full of potential to technology being used in real work and generating profits.

His explanation for the labs’ demand for more computing power was a reinforcing cycle. Profitable tokens and strong demand for them make more compute valuable, he said, because compute can produce more tokens, which can generate more profits. He called a public offering an important step in each company’s journey and said the money raised could fuel growth. But he described the two labs as only the “tip of the iceberg,” arguing that the AI industry is broader and deeper than its most visible companies.

On concerns about AI’s expense, Huang said the cost of a token was falling rapidly. He connected lower token costs to a declining cost of intelligence and, in turn, to greater benefits for customers.

The commercial case sits alongside visible obstacles. Chang pointed to data-center projects blocked or delayed in the United States, pushback in Europe, and AI IPOs being postponed or scrapped, asking whether these were cracks in the boom. Huang’s answer was that AI’s capabilities would not recede now that people had seen what it could do. He called the near-term challenges, in some ways, self-imposed.

In particular, he said the industry was doing a “horrible job” associating AI with the fact that it is built and engineered by people—and can therefore be improved. He described the industry’s public narrative as out of control and compared the development of young AI companies to the internet industry’s growing up. These companies are expanding quickly, he said, while still growing into the accountability and responsibility that come with building companies.

Taken together, Huang’s claims link safety to both agency and growth: developers can shape the technology, and they should build containment and monitoring as they advance it. But his answer on Nvidia’s own responsibility remains narrower. He said the company would stop sales if it knew an application was harming people, while cautioning that general-purpose technology limits its authority to judge how other customers use it.