Physical AI

Physical AI is artificial intelligence embodied in machines that sense and act in the real world. Where generative AI produces text or images on a screen, physical AI perceives its surroundings through sensors and moves through them, powering robots, humanoids, and autonomous systems that do physical work.

What is physical AI?

The split is between thinking and doing. A chatbot reads a prompt and writes an answer; nothing it produces touches the physical world. Physical AI closes that loop. It takes in the world through cameras, depth sensors, and touch, decides what to do, then acts through motors and actuators, so the output is a grasp, a step, or a weld rather than a paragraph. NVIDIA, which has built much of the compute and software for this shift, frames it as a new industrial era. At its GTC keynotes, CEO Jensen Huang has drawn the line directly:

“Physical AI has arrived — every industrial company will become a robotics company.” — Jensen Huang, NVIDIA CEO, NVIDIA and Global Robotics Leaders Take Physical AI to the Real World

The reason the term arrived now is that the hard part, generalization, started to work. Earlier robots were hand-programmed for one cell. Physical AI runs on robotics foundation models, large neural networks trained on demonstrations and simulation that learn skills they can transfer to new tasks. NVIDIA’s Isaac GR00T N1, announced in March 2025, is pitched as the first open foundation model for humanoid robots and uses a dual-system design: a fast “System 1” for reflex-like action and a slower “System 2” for deliberate planning (NVIDIA).

How does sim-to-real training work?

Robots cannot practice millions of times on a real floor without breaking things, so physical AI leans on simulation. The idea, called sim-to-real, is to train the model in a physically accurate virtual world, then transfer the learned policy to the metal. NVIDIA’s Isaac and Cosmos frameworks generate synthetic worlds and world-model predictions for this; in one example, the company generated 780,000 synthetic trajectories, the equivalent of nine months of human demonstration data, in 11 hours, and reported a 40% performance gain from blending that synthetic data with real data versus real data alone (NVIDIA). Simulation is the cheat that turns a data-starved problem into a compute problem.

That last word is the connection to silicon. Training these world models and running the policy on a robot are both compute-hungry. The same AI infrastructure story that powers data centers, custom accelerators like the TPU and the optical links described in silicon photonics, is the substrate physical AI draws on to train models in the cloud and, increasingly, to run them onboard. Physical AI is what that compute does once it leaves the screen.

Why does physical AI matter for investors?

Physical AI is the thesis that AI’s next leg is embodied, and its clearest expression is the humanoid robot. The pitch is automating physical labor, which NVIDIA ties to a global labor shortage it estimates at more than 50 million people (NVIDIA). The category is still early relative to the installed base of industrial robots; the IFR’s position paper stresses that humanoids are expected to complement existing robots, not replace the 4.6 million already at work (International Federation of Robotics). The money follows the long-run size. Goldman Sachs Research projects a humanoid market reaching $38 billion by 2035, with shipments of 1.4 million units, a forecast it raised more than sixfold as AI progress beat its expectations (Goldman Sachs). For an investor, physical AI routes spending across a layered chain: the compute and foundation models, the sensors and actuators, and the robot makers that integrate them. The risk is that the field is forecast-rich and deployment-thin, so reliability on real floors, not demo videos, is the number that matters.

Related terms & concepts

FAQ

How is physical AI different from generative AI?

Generative AI lives on a screen, producing text, images, or code from a prompt. Physical AI is embodied: it perceives the world through sensors and acts on it through motors and actuators, so its outputs are movements in real space, not pixels. NVIDIA's Jensen Huang put the shift bluntly: "Physical AI has arrived — every industrial company will become a robotics company" (NVIDIA).

What is NVIDIA's role in physical AI?

NVIDIA supplies a cloud-to-robot stack: the Isaac and Cosmos simulation frameworks for training robots in virtual worlds, the Jetson Thor onboard computer, and the open Isaac GR00T foundation models. GR00T N1, announced in March 2025, is pitched as the first open foundation model for humanoid robots (NVIDIA).

What is sim-to-real in physical AI?

Sim-to-real means training a robot's AI in a physically accurate simulation, then transferring the learned policy to the physical robot. NVIDIA generated 780,000 synthetic trajectories, the equivalent of nine months of human demonstration, in 11 hours, and reports a 40% performance gain from blending synthetic with real data (NVIDIA).

Sources & references

  1. NVIDIA and Global Robotics Leaders Take Physical AI to the Real World · NVIDIA, 2026-03-16
  2. NVIDIA Announces Isaac GR00T N1 — the World's First Open Humanoid Robot Foundation Model · NVIDIA, 2025-03-18
  3. New IFR position paper on humanoid robots published · International Federation of Robotics, 2025-07-17
  4. The global market for humanoid robots could reach $38 billion by 2035 · Goldman Sachs, 2024-01-08