Jensen Huang on the National Security Risks of AI Compute

In an interview published on April 15, Nvidia CEO Jensen Huang addressed the strategic risks associated with the sale of high-end artificial intelligence compute to strategic adversaries and the resulting responsibilities of the sellers.

The discussion, conducted by technology podcaster Dwarkesh Patel, focused on the intersection of American-made hardware and the development of AI models with significant cyber-offensive capabilities. The conversation specifically referenced the capabilities demonstrated by Anthropic’s Mythos Preview as a benchmark for the potential risks posed by such technology.

During a forty-minute segment of the interview, Patel repeatedly questioned Huang on the ethical and security obligations of companies that provide the compute necessary to train these models. The central inquiry focused on what responsibility a seller bears if the compute they provide is used by a strategic adversary to develop AI with serious cyber-offensive capabilities.

Strategic Competition and the “Five-Layer Cake”

Huang responded to these questions by framing the AI ecosystem as a "five-layer cake." He argued that the competition between the United States and China involves multiple layers of the technology stack, and that ceding any single layer to China would result in "industrial suicide."

According to analysis from War on the Rocks, Huang’s responses avoided a direct answer regarding the specific responsibility of the seller, instead focusing on the broader industrial and strategic implications of the U.S.-China competition in the AI sector.

Huang further asserted that China already possesses sufficient compute resources to carry out certain activities, though he did not elaborate on the specific capabilities in the provided text.