Nvidia CEO urges US-China dialogue on AI safety after Anthropic's Mythos

Nvidia CEO Jensen Huang used the discussion of Anthropic's Mythos breakthrough not for another argument over AI leadership, but to make a broader political point: the US and China need direct dialogue on how to develop and deploy increasingly powerful models safely. In his view, new advances in generative AI and research systems have reached a level where competition alone is not enough. If the major players fail to agree on even basic safety principles, the pace of progress will begin to outrun the ability of governments and companies to manage the risks.

The trigger for the statement was Mythos, an Anthropic breakthrough that Huang cited as an example of how quickly the upper limit of AI capabilities is shifting. He effectively linked a technical leap to a diplomatic agenda: the more powerful models become, the more important it is to discuss not only computing power but also rules of use. These are questions that can no longer be kept inside one country or one company: how to test systems before release, what restrictions dangerous scenarios require, how to prevent abuse, and where the line falls between open research and sensitive technology.

Huang's idea is simple: if the US and China remain stuck in a logic of mutual deterrence, agreeing on such rules will be much harder.

The topic is especially sensitive for Nvidia. The company sits at the center of the global AI boom, because its accelerators and software ecosystem remain key infrastructure for training and running large models. Nvidia has also worked with customers around the world for many years, including in China, and the relationship between Washington and Beijing over chips, exports and access to advanced computing has long been part of a larger geopolitical game.

Against that backdrop, Huang's words can be read in two ways. On one hand, this is the position of a tech executive who needs predictable rules. On the other, it is a signal that a purely coercive approach to AI regulation may prove too narrow if the technologies themselves go global faster than national restrictions can keep up.

The practical point of such a dialogue is not to create a single global regulator, but to try to lock in at least a minimum of compatible approaches. These could include shared principles for assessing dangerous capabilities, an exchange of stress-testing methods, rules for labeling synthetic content, and incident-response procedures. Even a partial overlap in standards could reduce the risk that companies rush releases without adequate checks simply because they fear falling behind in the international race.

It also matters that the statement centered not on an Nvidia product but on an achievement by Anthropic, a company that has actively built its reputation around AI safety. That shifts the emphasis from commercial competition to the question of trust in the next generation of systems. When major market players start talking not only about performance but also about shared safety rules, it means the industry increasingly sees risk less as an abstraction for researchers and more as a practical task for business and governments.

For the US and China, this conversation is particularly hard: they compete for leadership, restrict each other's access to critical technologies, and yet face the same challenges, from disinformation and autonomous code to the use of AI in defense, education and public administration.

The core of Huang's statement is that the next stage of the AI race will be determined not only by who shows the next breakthrough first, but by whether the major powers can agree on minimal rules of coexistence. For the market, this is an important signal: safety is ceasing to be an add-on to the product and is becoming part of international politics. If such a dialogue does begin, the winners will include not only model developers and chip suppliers but also users, who will have to live with the consequences of increasingly autonomous AI systems.