Companies like Waymo and Tesla have been working on self-driving for more than a decade, but their advantage might not be as much as some people believe.
That’s the striking claim from NVIDIA CEO Jensen Huang, who recently outlined his perspective on the evolution of autonomous vehicle technology on the No Priors podcast. In a wide-ranging discussion, Huang broke down the development of self-driving cars into four distinct eras—and suggested that much of the industry’s early work may have been rendered obsolete by recent advances in AI. Perhaps most provocatively, he argued that starting from scratch just three years ago would have yielded nearly identical results to a decade of development.
“Self-driving cars really have four eras,” Huang explained. “The first era was smart sensors connected into a car—the Mobileye era. Even the earliest days of Waymo, you’re using smart sensors, a lot of human-engineered algorithms, and extreme mapping.”
He described this first generation as essentially creating “a car that is driving on digital rails. It’s no different than the rails at Disneyland, except they are digital rails.”
According to Huang, the second generation maintained a modular approach with distinct systems for perception, world modeling, and planning. “Each one of these modules had the limits of their technology,” he noted. “Perception was first affected by deep learning, and then it propagated through the pipeline. But that system was too brittle. It only knows how to perform what you taught it.”
The third era, Huang explained, is characterized by end-to-end models—integrated systems that handle the entire driving task holistically rather than through separate modules. Looking ahead, he predicted a fourth era: “End-to-end models with reasoning.”
Then came his most eyebrow-raising assertion: “If we started self-driving cars probably three years ago, we would probably be exactly the same place. And I don’t mind it. I’ve been working on it for 10 years.”
Huang’s comments suggest that the recent revolution in AI—particularly the emergence of large language models and advanced neural networks—has fundamentally changed the autonomous vehicle landscape. The implication is that traditional approaches, no matter how refined over years of development, may be less valuable than cutting-edge AI architectures that didn’t exist until recently. This echoes broader trends in the AI industry, where foundation models and end-to-end learning have displaced older, more rigid approaches across numerous domains.
NVIDIA has positioned itself as a central player in this transformation, providing the computing infrastructure that powers AI development across industries. Huang pointed out that “NVIDIA’s self-driving car stack, by the way, is number one rated in safety in the world today. Number two is Tesla.” NVIDIA had also released an open-source Alpamayo model last week that would enable automakers to add self-driving to their own vehicles. And if NVIDIA, better known for its GPUs than for autonomous vehicles, ends up being a big player in the self-driving space space, it would be a massive achievement — and would validate Huang’s thesis that recent technological advances can rapidly overtake years of earlier work.