Nvidia CEO Jensen Huang outlines three-computer architecture for physical AI

Jensen Huang thinks building a useful robot takes three computers. Not one supercomputer, not a clever chip, but three distinct machines, each with its own job.

The Nvidia CEO says physical AI needs one computer to train the model, one to test it in simulation, and one to run the robot itself. Conveniently, Nvidia sells all three.

One brain, three bodies

Physical AI is the term for systems that perceive, reason and act in the real world. Think warehouse robots, autonomous machines and factory arms, rather than chatbots answering emails.

The training computer comes first. This is where the AI model learns, and Nvidia points to its DGX and Grace systems for the job.

The simulation computer sits in the middle. Nvidia’s Omniverse platform handles this layer, building digital twins where robots can practice before touching anything real.

The runtime computer lives inside the machine. Nvidia’s example here is Jetson Thor, an edge computer embedded in the robot that makes decisions in real time.

Huang shared these views in an October 2026 interview. He leaned hard on digital twins as the key to preparing robots and agents for real-world work.

The reasoning is about data. Large language models had the entire internet to learn from. Robots navigating a specific factory floor or hospital corridor have far less to draw on, and Huang expects that data scarcity to be a real constraint in many physical environments.

The tools behind the pitch

Nvidia has paired the architecture with a growing set of software. The Isaac GR00T robotics foundation model is designed as a general starting point for robot intelligence.

Cosmos is the company’s tool for generating synthetic data. It produces the artificial training material that fills the gap left by scarce real-world footage.

Then there is Newton, a physics engine developed in collaboration with Google DeepMind and Disney.

Nvidia also claims it is the only company building complete hardware and software stacks across training, simulation and deployment. All of it runs on CUDA, the company’s long-standing programming platform for its chips.

The money and the labor gap

Physical AI reportedly brought in around $10 billion in revenue during the first half of 2026, according to Nvidia’s claims.

Huang has also tied the push to a labor problem. He projected a global shortage of 30-50 million workers by 2030 across sectors including manufacturing, logistics and healthcare.

What this means

For investors, the three-computer framing is really a revenue map. Each layer represents a separate purchase, and Nvidia is positioned to collect at every stage of a robot’s life cycle.

A company building robots would buy training hardware for its data center, simulation tools for testing, and edge chips for every unit it ships.

The CUDA angle is the moat. Developers who build on Nvidia’s software at the training stage have strong incentives to stay with it through simulation and deployment, making competitors’ lives harder at each step.

Watch how partners like Google DeepMind and Disney use Newton, and whether the reported physical AI revenue holds up in future disclosures.