Nvidia Kicks Off GTC 2025 with Focus on AI Advancements
Nvidia founder Jensen Huang announced pivotal developments in artificial intelligence at the company’s AI developer conference, GTC 2025, which he dubbed the “Super Bowl of AI.” Huang emphasized AI’s current “inflection point” and detailed Nvidia’s advancements and strategic direction for the coming years.
Huang stated that demand for GPUs from the top cloud service providers is soaring. He projected that Nvidia’s data center infrastructure revenue will reach $1 trillion by 2028.
Next-Generation Graphics Architectures Unveiled
During the keynote, Huang spotlighted Nvidia’s next-generation graphics architectures: Blackwell Ultra and Vera Rubin. Blackwell Ultra is slated to launch in the second half of 2025, with its successor, the Rubin AI chip, expected in late 2026. Rubin Ultra is planned for a 2027 release.
Huang elaborated on the “extraordinary progress” of AI over the past decade. He noted AI’s evolution from perception and computer vision to generative AI and, most recently, to agentic AI, which can reason and generate answers. “AI understands the context, understands what we’re asking. Understands the meaning of our request,” he explained. “It now generates answers. Fundamentally changed how computing is done.”
The Future of AI: Robotics and Physical AI
The next wave of AI, according to Huang, is robotics, particularly with the rise of “physical AI,” designed to understand concepts like friction, inertia, cause and effect, and object permanence. “Each one of these phases, each one of these waves, opens up new market opportunities for all of us,” Huang said.
Synthetic data generation emerged as a key element of both physical AI and many of Huang’s other announcements. AI needs digital experiences to learn, and it learns so quickly that human involvement in the training loops becomes obsolete. “There’s only so much data and so much human demonstration we can perform,” Huang stated. “This is the big breakthrough in the last couple of years: reinforcement learning.” Nvidia’s technology is designed to support reinforcement learning for AI as it engages in solving problems step by step.
Isaac GR00T N1 and Cosmos AI Model
To further these advancements, Huang introduced Isaac GR00T N1, an open-source foundation model designed to aid the development of humanoid robots. It will be paired with an updated Cosmos AI model to develop simulated training data for robots. Benjamin Lee, a professor at the University of Pennsylvania commented about the implications. Lee noted that providing an open-source platform will allow more researchers to explore reinforcement learning using synthetic data.
At CES earlier this year, Huang introduced the Cosmos series of AI models for generating cost-effective photo-realistic video to train robots and automated services. This open-source model works with Nvidia’s Omniverse to create realistic video, which is projected to be cheaper than traditional training forms.
Partnerships and Future Solutions
During the presentation, Huang announced that General Motors plans to integrate Nvidia technology into its new fleet of self-driving cars. The partnership entails building custom AI systems using Omniverse and Cosmos models for training AI manufacturing models.
Nvidia also unveiled Halos, an AI solution focusing on automotive safety, especially in autonomous driving. He also announced Newton, an open-source physics engine developed in partnership with Google DeepMind and Disney Research, for robotics simulation. To conclude his talk, a small robot named Blue joined Huang on stage, showcasing the arrival of “the age of generalist robotics.”
