Jensen Huang: NVIDIA's moat is 10x token cost cuts yearly

Jensen Huang used the Dwarkesh Podcast to define NVIDIA as the layer between electrons and tokens—the layer that converts physics into intelligence. This self-definition underpins his rejection of the view that AI kills software; instead, he argues that as AI agents proliferate, demand for the tools supporting them will surge.

Huang backed up NVIDIA's position with concrete numbers: Blackwell delivers 50x the performance of Hopper, and models optimized directly by NVIDIA run 3x faster—which he identifies as the essence of the CUDA moat. But the deeper point is sustainability: the real moat is a roadmap that reliably cuts token cost by 10x every year.

He also offered rare candor on strategy, saying NVIDIA avoids the cloud business as a matter of philosophy and admitting he did not think deeply enough about investing in Anthropic. Such openness from a CEO at the top of the AI chip market is notable, and it fits the picture of a leader confident enough to discuss mistakes and competitor dynamics openly.