In his speech, Jensen Huang discussed physical AI, the Rubin AI chips, the GeForce 5090, the future of agentic AI, and robots, among other ground-breaking developments in artificial intelligence.
The CEO of NVIDIA Corp., Jensen Huang, gave a powerful keynote speech at the company’s much awaited GTC 2025. With his trademark wit, technological expertise, and famous leather jacket, the CEO paved the way for the next phase of artificial intelligence.
To a crowded audience full of developers, corporate executives, and other stakeholders, the two-hour keynote address emphasized the speed at which the AI revolution is unfolding and the crucial role that NVIDIA has played in this success story. The presentation provided an overview of what is to come, including details about the chipmaker’s past products, pushing the boundaries of hardware, how it is helping big tech grow AI, and how it is changing AI-driven businesses.
Huang was addressing the crowd at what has been called the AI Super Bowl. There were 25,000 people that attended the event in person in San Jose, California, which was a huge turnout. This TLDR version of the keynote, which featured the development of AI along with some surprises and futuristic presentations, is available in case you missed it.
The main conclusions drawn from Jensen Huang’s speech
The next stage of physical AI
Huang discussed the emergence of physical AI in his speech, claiming that it is the next significant advancement in AI. AI that understands physical notions such as friction, inertia, cause and effect, and object permanence is referred to as “physical AI,” he explained. According to him, “a new era of AI, which we call Physical AI, will be enabled by the ability to understand the physical world, the three-dimensional world, and it will enable robotics.”
According to the CEO, it is a revolutionary stage of AI that comes after agentic, generative, and perceptual AI. Robotics will be built on physical AI, resulting in a significant change in AI applications. Unlike previous models, which mostly processed data and produced content, physical AI will enable machines to travel and function in real-world settings. Huang thinks that by developing robotics, physical AI would attract new partners and sectors, hence growing the AI ecosystem. The path ahead is not simple, though. The CEO also emphasized how difficult it is to develop physical AI due to the enormous processing challenges.
These problems include training AI without human-in-the-loop constraints to speed learning, efficiently leveraging data to educate AI how to interact with the physical world, and scaling AI models to interpret real-world physics. Huang’s confidence regarding the combination of AI and robotics is demonstrated by his classification of physical AI as the next natural step in AI development.
The Blackwell Architecture and GeForce 5090
The focus of last year’s GTC conference was Nvidia’s Blackwell AI chips, and Huang’s speech this year included them as one of the primary points of interest. Blackwell Ultra, an upgraded version of the chip, was introduced by the CEO. The newest GPU from the firm is the Ultra.
With the help of the Blackwell GPU architecture, the firm unveiled the GeForce 5090. The new flagship GPU is 30% more efficient than the previous 4090, and Huang emphasized that it offers noticeably better AI-powered graphics and performance.
Regarding the development of Gen AI
“AI is about creation, not just perception anymore.” According to Huang, the origins of artificial intelligence fundamentally shifted from perception (computer vision and voice recognition) to generative AI (text-to-image, image-to-text, protein synthesis, etc.). “This shift from retrieval-based computing to generative computing signified a fundamental shift in our approach to problem-solving,” he noted.
Rubin Vera
Additionally, Huang introduced the next generation of artificial intelligence processors, called Rubin, after the astronomer who made the discovery of dark matter. A CPU known as Vera and a novel GPU design named Rubin make up the GPU, which is anticipated to launch in 2026. In 2024, Vera, the company’s first proprietary CPU design, is anticipated to double the speed of the CPU found in Grace Blackwell processors. Together, Vera and Rubin can handle 50 petaflops during inference, surpassing the 20 petaflops on the present-day Blackwell processors. The speed at which a computer processes information is measured in petaflops, or PF. Rubin has 288 GB of fast RAM as well.
The emergence of artificial intelligence
The majority of 2025 has been devoted to the introduction of agentic AI into the workforce, which will spark a wide range of applications in many sectors. Huang also appeared optimistic about the current pattern. “AI that sees, thinks, and acts on its own is the next big thing,” he stated. He claimed that the world was at a turning point in computing worth $1 trillion. According to the CEO, the need for AI computing has been rising quickly due to the development of agentic and reasoning AI.
During the conversation on the future of agentic AI, the CEO also revealed the open Llama Nemotron family of reasoning models. These models are intended to give developers and businesses a ready-to-use platform on which to build AI agents that can complete complicated tasks either individually or in coordinated teams. On-demand AI reasoning capabilities can be provided by the models based on Llama models, according to Huang. In order to increase complicated decision-making, reasoning, multistep math, and coding, the company has improved the new reasoning model family during post-training.
CUDA-X GPU and self-driving cars
The CEO claimed that CUDA has reached a tipping point in accelerated computing and that CUDA-X GPU-accelerated libraries and microservices are now driving industries across the spectrum. Additionally, he disclosed that the company’s cuOpt decision optimization technology will be made publicly available. He did, however, point out that General Motors has adopted NVIDIA AI, simulation, and accelerated computing for next-generation vehicles and automation, demonstrating how AI’s expanding influence is spreading to robotics, self-driving cars, and factories. The CEO also unveiled NVIDIA Halos, a safety system that combines NVIDIA’s AI research, automotive hardware, and software for the protection of self-driving cars.
The future of robotics and AI
At the conclusion of his speech, Huang displayed a number of robotics-related announcements. The CEO stated that there is a 50 million job deficit and that robotics is a $10 trillion market. The first open and customizable foundation model for humanoid reasoning, Isaac GR00T N1, was unveiled by NVIDIA to speed up robotics. In order to provide users more control over AI-driven world development, the company also unveiled new NVIDIA Cosmos world foundation models.
Infinite, controlled data creation is being made possible by NVIDIA with Cosmos and the Omniverse platform. The Newton open-source physics engine, created in collaboration with Disney Research and Google DeepMind, was also unveiled by him. His speech came to a close with a live demonstration of Blue, a little AI-powered robot that looks like Wall-E from Disney. Applause and appreciation were shown for the friendly robot.
