At CES 2025’s opening keynote, Nvidia co-founder and CEO Jensen Huang unveiled what it calls “Cosmos,” a set of AI technologies that forms a new software platform. Cosmos includes “state-of-the-art generative world foundation models,” as Huang called them.
Cosmos taps generative AI to fill the biggest gap that’s keeping robots from becoming more useful: training data. As Huang says, “In order to train a robot, you need a lot of data.” This technology has the potential to enable robotics and autonomous vehicle makers to make important leaps forward in the years ahead.
For that reason, ZDNET and the rest of the CNET Group awarded Nvidia Cosmos two official Best of CES awards for 2025. You can see all 12 of the Best of CES winners selected by the CNET Group (ZDNET, CNET, PCMag, Mashable, and Lifehacker) in partnership with the organization that runs CES.
Also: CES 2025: The most impressive products we’ve seen
Cosmos works in conjunction with Nvidia’s physics simulation tool, Omniverse. Omniverse generates simulations, and Cosmos then turns them into photo-realistic video imagery to train robots and automobiles. “Take thousands of drives and turn them into billions of miles” is how Huang characterized the interplay between Omniverse and Cosmos.
Huang was surrounded onstage by a number of existing models of general-purpose humanoid robots.
Huang called Cosmos “the world’s first world foundation model,” noting it is trained on 20 million hours of video. “It’s really about teaching the AI to understand the physical world.”
Also: Nvidia’s Omniverse: The metaverse is a network, not a destination
Huang compared the Cosmos project to Meta Platforms’s wildly popular Llama large language model, saying, “We really hope it will do for the world of robotics and AI what Llama has done for enterprise AI.”
Cosmos with Omniverse can be used for applications such as training a robot for a warehouse by having the robot perform hours of training in a simulation of the warehouse environment. “The ChatGPT moment for robotics is just around the corner,” said Huang.
The Cosmos code is available under an open-source license on GitHub, said Huang.
Huang Emphasized Nvidia’s automotive partners in his CES 2025 keynote.
Nvidia has inked a deal for Toyota, the world’s largest car manufacturer, to use the company’s autonomous driving chips and software in multiple different models of car, Huang also announced during the opening keynote.
“The AV revolution has arrived,” said Huang, meaning autonomous vehicles.
Also: Nvidia Project Digits: A Linux-powered desktop for AI developers
“Today, Toyota and Nvidia are going to partner together to create their next-generation AVs.” Self-driving cars will be the “first trillion-dollar robotics market,” Huang predicted.
Nvidia has had a relationship with Toyota for several years now. The company’s DGX computers have been used by Toyota for training artificial intelligence models for self-driving vehicles. Monday’s announcement is an expansion of that relationship, said Nvidia’s head of automotive products, Ali Kani, in a briefing with reporters, where the carmaker will also use the company’s “AGX” onboard AI computer. The latest version of that chip, Huang announced, is called the “Thor AGX.” It is twenty times more powerful than its predecessor Orin model.
Huang was surrounded onstage by a number of existing models of general-purpose humanoid robots.
Huang also unveiled additions to its suite of AI software. The updates include a group of AI models based on Meta Platforms’s Llama model, called Llama Nemotron. Huang told the audience that Llama is “the reason every organization has been activated to work on AI.” The Nvidia versions are meant to “fine-tune” Llama for enterprise use.
Huang also talked at length about the rising prominence of “agentic AI,” where large language models, or multi-modal AI models, can call upon outside programs to carry out tasks.
Also: AI data centers are becoming ‘mind-blowingly large’
“There’s this whole world of agentic AI, all these amazing new startups building frameworks like LangGraph, Llama Index, and Crew AI,” said Justin Boitano, president of enterprise AI software products at Nvidia, in a briefing we had.
Those startups are “changing the programming model of how do you write applications: you write an AI, you give it a role, which is like a persona, you give it a goal, you can create it with just a prompt.” Nvidia works extensively with the startups, said Boitano.
Huang said that agentic AI, combined with self-driving cars and robots, are “three types of robots we are working on.”
A new version of the Grace-Blackwell combined CPU and GPU chip, called GB10, is the brains of the new DIGITS desktop computer.
Other announcements in the keynote included GEFORCE “Blackwell,” the latest version of the company’s gaming GPU, which is slashed in price from its predecessor 4090 to $549 from $1,599, available starting this month, with laptop versions coming in March; and Project DIGITS, a compact personal computer optimized for AI development, running a new version of the Grace-Blackwell combined CPU and GPU chip, called GB10.
+ There are no comments
Add yours