NVIDIA founder and CEO Jensen Huang used his keynote at GTC Taiwan on Sunday to make a sweeping declaration: the age of physical AI has arrived. Speaking at the Taipei Popular Music Center, Huang argued that artificial intelligence is no longer a screen-bound service — it is becoming a tangible presence embedded in the machines, vehicles, and industrial systems that shape everyday life.
From Digital to Physical: What Agentic AI Really Means
At the heart of Huang's address was the concept of agentic AI — systems that can understand context, reason through problems, plan sequences of action, and execute them autonomously. Huang described this class of AI as, in essence, a digital robot. The distinction that matters, he argued, is where that intelligence lives: when agentic capabilities move from cloud servers and laptop screens into cars, factory floors, satellites, agricultural equipment, and heavy industrial machinery, AI crosses a threshold from software service to physical actor.
The implications, Huang said, extend well beyond consumer technology. Robots, autonomous vehicles, and smart infrastructure are set to become nodes in a vast, distributed network of AI agents — one that operates continuously across both virtual and physical environments.
Why Data, Not Compute, Is the Real Bottleneck
Huang identified the data problem as the central challenge standing between current AI capabilities and functional physical AI. Language models advanced rapidly, he explained, because the internet offered an almost unlimited supply of human-generated, human-readable text. Physical AI requires something fundamentally different: first-person, embodied data captured from the perspective of a machine actively performing a task in the real world.
The gap is stark. Most video recorded globally is shot from a third-person vantage point and carries limited value for training robotic systems that need to understand the world as they will actually experience it. To address this, NVIDIA is scaling its Omniverse simulation platform alongside teleoperation systems, human demonstration pipelines, synthetic data generation tools, and world-model capabilities — giving robots a way to learn in virtual environments before they ever operate in the real one.
Cosmos 3: A Foundation Model Built for the Physical World
The headline announcement was Cosmos 3, which NVIDIA positions as a foundation model for physical AI. Huang described it as capable of interpreting real-world scenes, generating physically accurate synthetic video, and functioning as a simulator for training and evaluating robot decision-making.
Crucially, Cosmos 3 is being released as an open model system. NVIDIA plans to publish the model weights, training data, and methodology, allowing developers to fine-tune it for specific domains — robotics, factory automation, and autonomous driving among them. Huang acknowledged the large language model space is now intensely competitive, but said NVIDIA occupies a front-line position in physical AI, and that Cosmos 3 is designed to serve as the foundational layer for a wide range of robotic systems yet to be built.
Autonomous Vehicles Enter the Reasoning Era
NVIDIA also unveiled Alpamayo 2, an open model for autonomous vehicles. Huang argued that the next generation of self-driving systems must move beyond perception and object recognition toward genuine reasoning — the capacity to interpret road conditions, anticipate pedestrian and vehicle behavior, and make sound decisions in unpredictable environments.
During the keynote, NVIDIA demonstrated a self-driving system capable of explaining its own decisions in real time: when to yield, when to avoid an obstacle, and how to manage spacing with surrounding traffic. Huang acknowledged — with evident humor — that a car narrating its every thought aloud would likely try passengers' patience, but said that continuous internal reasoning is precisely what defines capable autonomous driving.
Isaac Groot: Lowering the Barrier to Humanoid Robotics
On humanoid robotics, NVIDIA introduced the Isaac Groot reference platform. Huang described the challenge plainly: building a humanoid robot is extraordinarily difficult because research teams must integrate sensors, motors, simulators, data pipelines, training systems, and computing hardware before substantive research can even begin — a process that can consume months of engineering effort.
Isaac Groot consolidates the full NVIDIA robotics stack — open models, simulation and training libraries, data generators, Isaac Lab, Omniverse, Cosmos, and Jetson Thor — into a single reference design. The platform is intended to help universities, research institutions, and companies accelerate their entry into frontier robotics without having to rebuild the infrastructure from scratch.
A Billion Agents — and a Surprise Finale
Huang closed his remarks with a broad vision: billions of agentic systems operating globally, from cloud data centers to personal computers, from factory floors to autonomous machines. AI, he said, will no longer be confined to a chat window on a screen. It will take physical form — visible, tangible, and capable of directed action.
In a characteristically unexpected finale, Huang treated the audience to a short film produced entirely by AI. It opened with scenes of Taiwanese night markets and moved to an upbeat rap soundtrack whose lyrics captured NVIDIA's ambitions for AI development. The moment brought the event to its highest energy point — and set the stage for what promises to be a consequential Computex season in Taipei.
