Nvidia CEO Jensen Huang is a big celebrity in Las Vegas this week. His CES keynote at the Fontainebleau Resort was harder to get into than any of the sold-out Vegas shows. Journalists who canceled their plans for the event waited for hours outside the 3,600-seat BleauLive Theater. Many who arrived on time—after navigating the maze of convention centers and, in one case, flying in from abroad to see the tech king of the day—were turned away by the sheer volume and redirected to the outdoor viewing party, where some 2,000 attendees gathered with a mixture of frustration and reverence.
Shortly after 1 p.m., Huang ran onto the stage, wearing a shiny, bright leather jacket, and wished the crowd a Happy New Year. He opened with a quick history of AI, tracing the past few years of clear progress—from the growth of large-scale language models to the development of OpenAI in thinking systems and the explosion of the so-called AI of AI All built on the theme that dominated most of his 90-minute presentation: physical AI.
Physical AI is a concept that has gained momentum among leading researchers over the past year. The goal is to train AI systems to understand intuitive laws that humans take for granted—such as gravity, causality, motion and permanence—so that machines can communicate and safely interact with real environments.
Nvidia is entering the self-driving race
Huang launched Alpamayo, the world’s basic model designed to enable autonomous driving. He called it “the world’s first AI to think independently”
To demonstrate, Nvidia played a one-shot video of a Mercedes car fitted with Alpamayo driving around the busy city of San Francisco. The car swerved, stopped at lights and traffic, gave way to pedestrians and changed lanes. The human driver was behind the wheel while driving but did not intervene.
One very interesting thing Huang discussed is how Nvidia trains portable AI systems—a very different challenge for training language models. The greatest examples of language are found in literature, where humanity has produced enormous amounts. But how do you teach AI Newton’s second law of motion?
“Where does that data come from?” Huang asked. “Instead of languages—because we’ve created a bunch of text that looks at basic facts for AI to learn—how do we teach AI basic facts of physics? There are lots and lots of videos, but not enough to capture the variety of interactions we need.”
Nvidia’s answer is artificial data: information generated by AI systems based on real-world data samples. In Alpamayo’s case, Nvidia’s alternative world model—called Cosmos—uses limited real-world inputs to produce more complex, physically realistic videos. The basic traffic scenario becomes a series of realistic camera views of vehicles interacting on congested roads. A static image of a robot and vegetables turns into a dynamic kitchen space. Even textual information can be turned into video with precise physical movements.
Nvidia said the first fleet of Alpamayo powered robots, built on 2025 Mercedes-Benz CLA vehicles, is expected to be launched in the US in the first quarter, followed by Europe in the second quarter and Asia later in 2026.
For now, Alpamayo is still a Level 2 autonomous driving system—similar to Tesla’s Full Self-Driving—that requires a human driver to stay behind the wheel at all times. Nvidia’s long-term goal is Level 4 autonomy, where cars can operate without human supervision in certain, time-delayed environments. That’s one step below full autonomy, or Level 5.
“ChatGPT’s moment of physical AI is about to come,” Huang said in a voiceover accompanying one of the videos shown during the keynote.
