Nvidia CEO Jensen Huang unveils new Rubin AI chips at GTC 2025

Hey there, time traveller!
This article was published 18/03/2025 (539 days ago), so information in it may no longer be current.

Nvidia founder Jensen Huang kicked off the company’s artificial intelligence developer conference on Tuesday by telling a crowd of thousands that AI is going through “an inflection point.”

At GTC 2025 — dubbed the “Super Bowl of AI” — Huang focused his keynote on the company’s advancements in AI and his predictions for how the industry will move over the next few years. Demand for GPUs from the top four cloud service providers is surging, he said, adding that he expects Nvidia’s data center infrastructure revenue to hit $1 trillion by 2028.

Huang’s highly anticipated announcement revealed more details around Nvidia’s next-generation graphics architectures: Blackwell Ultra and Vera Rubin — named for the famous astronomer. Blackwell Ultra is slated for the second half of 2025, while its successor, the Rubin AI chip, is expected to launch in late 2026. Rubin Ultra will take the stage in 2027.

In a talk that lasted at over two hours, Huang outlined the “extraordinary progress” that AI has made. In 10 years, he said, AI graduated from perception and “computer vision” to generative AI, and now to agentic AI — or AI that has the ability to reason.

“AI understands the context, understands what we’re asking. Understands the meaning of our request,” he said. “It now generates answers. Fundamentally changed how computing is done.”

The next wave of AI, he said, is already happening: robotics.

Robotics fueled by so-called “physical AI” can understand concepts like friction and inertia, cause and effect, and object permanence, he said.

“Each one of these phases, each one of these waves, opens up new market opportunities for all of us,” Huang said.

The key to that physical AI, and many of Huang’s other announcements, was the concept of using synthetic data generation — AI or computer-created data — for model training. AI needs digital experiences to learn from, he said, and it learns at speeds that make using humans in the training loops obsolete.

“There’s only so much data and so much human demonstration we can perform,” he said. “This is the big breakthrough in the last couple of years: reinforcement learning.”

Nvidia’s tech, he said, can help with that type of learning for AI as it attacks or tries to engage in solving a problem, step by step.

To that end, Huang announced Isaac GR00T N1, an open-source foundation model designed to assist in developing humanoid robots. Isaac GR00T N1 would be paired with an updated Cosmos AI model to help develop simulated training data for robots.

Benjamin Lee, a professor of electrical and systems engineering at the University of Pennsylvania, said that the challenge in training robotics lies in data collection because training in the real world is time-consuming and expensive.

A simulated environment has long been a standard for reinforcement learning, Lee said, so researchers can test the effectiveness of their models.

“I think it’s really exciting. Providing a platform, and an open-source one, will allow more people to learn on reinforcement learning,” Lee said. “More researchers could start playing with this synthetic data — not just big players in the industry but also academic researchers.”

Huang introduced the Cosmos series of AI models, which can generate cost-efficient photo-realistic video that can then be used to train robots and other automated services, at CES earlier this year.

The open-source model, which works with the Nvidia’s Omniverse — a physics simulation tool — to create more realistic video, promises to be much cheaper than traditional forms of gathering training, such as having cars record road experiences or having people teach robots repetitive tasks.

U.S. car maker General Motors plans to integrate Nvidia technology in its new fleet of self-driving cars, Huang said. The two two companies will work together to build custom AI systems using both Omniverse and Cosmos to train AI manufacturing models.

The Nvidia head also unveiled the company’s Halos system, an AI solution built around automotive — especially autonomous driving — safety.

“We’re the first company in the world, I believe, to have every line of code safety assessed,” Huang said.

At the end of his talk, Huang an open-source physics engine for robotics simulation called Newton, which is being developed with Google DeepMind and Disney Research.

A small, boxy robot named Blue joined him on stage, popping up from a hatch in the floor. It beeped at Huang and followed his commands, standing beside him as he wrapped up his thoughts.

“The age of generalist robotics is here,” Huang said.