What Is Physical AI? The 3 Computers, Real Costs and Limits

What Physical AI Is and How It Differs From Chatbots and Robots

Physical AI is the use of AI models that sense an environment, predict what will happen next and control a machine that changes that environment. It is broader than humanoid robots. Robotaxis, warehouse robots, factory arms and drones all count.

NVIDIA's physical AI glossary gives the clearest vendor definition. It says what the machines are for, not how they work.

"Perceive, understand, and perform complex actions in the physical world." (NVIDIA, Physical AI glossary, 2026)

The term has no legal or standards-body definition, and every vendor stretches it a little differently. That is the first reason to be careful with headlines about it. The table below separates the concepts that get mixed up.

Embodied AI and physical AI overlap heavily. Researchers use embodied AI when the body is the point, and companies use physical AI when they also want to include cars and whole factories.

Why a chatbot is not physical AI

A chatbot can describe how to pick up a cup. A physical AI system has to find the cup, estimate its shape and position, account for friction, move an arm, apply the right force and recover if the cup slips. That demands low-latency control and a safety case, and it has to work in a world that never repeats itself exactly.

Physical AI does not replace classical robotics either. Motion planning, feedback control and safety systems stay in the loop. The learned model decides what to grab, and older, duller software keeps the arm from hurting anyone. Our guide to AI robots shows how the two layers combine in real machines.

How Physical AI Works: Sense, Predict, Plan, Act

A physical AI system runs a loop with five parts, and the model is only one of them.

• Perception: cameras, lidar, radar, microphones and force sensors turn the world into data.
• World understanding: the system builds a picture of objects, people, geometry and motion.
• Prediction and planning: it estimates what will change and picks the next steps.
• Action: motor commands drive steering, braking, grasping or walking.
• Learning: demonstrations, trial and error and simulation feed back into the model.

Anyone can get one impressive action on video. Getting the same action right thousands of times in a messy room is the hard part.

Vision-language-action models

A vision-language-action (VLA) model takes camera images and a written instruction and outputs robot movement. It is the physical counterpart of a language model. NVIDIA's Isaac GR00T N1.6, released in January 2026, is an open reasoning VLA for humanoid robots. It draws on Cosmos Reason 2, a separate open vision-language model that helps the robot understand a scene before it acts.

A VLA is not the whole robot. The sensors, motors, safety limits and control loops around it still decide whether a movement is allowed.

The planner and the actor

Google DeepMind splits the job in two. In its September 2025 announcement, DeepMind described Gemini Robotics-ER 1.5 as a high-level brain that builds multi-step plans, and Gemini Robotics 1.5 as the model that carries out each step. The actor "can now think before taking action," writing out reasoning in natural language before it moves.

"Can transfer motions learned from one robot to another, without needing to specialize the model to each new embodiment." (Google DeepMind, September 2025)

Access shows how early this is. DeepMind put the planner in the Gemini API for any developer and kept the acting model for select partners.

Where world models fit

A world model predicts what a scene will look like after something happens. It is not the controller that drives motors. Its job is to imagine consequences, so a robot can test a plan in its head and so developers can generate training scenes without breaking real hardware. NVIDIA's Cosmos is the best-known example, and the next section shows where it sits in the compute stack.

The Three Computers Behind Every Physical AI System

Every physical AI system uses three kinds of compute, and only one of them is on the machine. NVIDIA calls this its three-computer framework: one stage to train and fine-tune models, one to simulate and validate them (including digital twins), and one that runs the trained model on the machine in real time.

Two of the three live in data centers. Only the third travels with the machine.

Training: the Cosmos example

NVIDIA launched Cosmos in January 2025 as an open platform of world models. According to NVIDIA's own account on Hugging Face, the models were trained on 20 million hours of video, equal to about 9,000 trillion tokens, using 10,000 H100 GPUs over three months. The current generation is larger. NVIDIA's Cosmos page lists Cosmos 3 in three sizes of 64 billion, 16 billion and 4 billion parameters, released under the OpenMDW 1.1 license from the Linux Foundation.

Simulation: why synthetic data matters

Real robots collect data slowly. A machine has to move to produce an example, hardware breaks, rare failures almost never show up, and dangerous cases are unsafe to repeat. NVIDIA's page says world models and simulation can "help create varied scenarios and synthetic data" for that reason. The weak spot is the sim-to-real gap: simulated friction, lighting and sensors never match reality exactly, so a policy that wins in simulation can still fail on a slightly slippery floor.

Runtime: the chip on the machine

NVIDIA's Jetson Thor page lists 2,070 FP4 TFLOPS (sparse), 128 GB of memory and a 40 to 130 W range for the developer kit and T5000 module, and claims 7.5 times the performance and 3.5 times the energy efficiency of the previous AGX Orin. A smaller Jetson T4000 module, announced in January 2026, offers 1,200 FP4 TFLOPS and 64 GB at $1,999 in volumes of 1,000. Those are vendor peak ratings. A robot's real draw also includes motors, sensors and cooling. For the wider edge story, see our guide to edge AI.

Who Is Building Physical AI and What Is Actually Deployed

The physical AI in service today is mostly industrial arms and robotaxis, not humanoids. The table lists the main builders with the best verified figure for each.

Waymo is the clearest proof that physical AI works at scale. TechCrunch reported in March 2026 that Waymo was providing about 500,000 paid rides a week across 10 cities, up from about 50,000 a week in May 2024. That is tenfold growth in under two years, and every ride is a vehicle perceiving, predicting and steering with no driver. A ride someone paid for is a harder number to flatter than a demo video, so we weight it more heavily than any humanoid clip.

Factories are the other large base. The International Federation of Robotics (IFR) counted more than 600,000 industrial robots installed in 2025, with about 5 million at work. Most of those arms run fixed programs, so they are physical AI candidates rather than proven physical AI. Humanoids are tiny by comparison, at about 7,000 full-size units sold that year. Our guides to humanoid robots and AI robots cover those numbers in detail.

The vendor figures above come from NVIDIA's own announcements, such as the one on March 16, 2026, where Jensen Huang said that "every industrial company will become a robotics company." A name on a vendor's partner list says nothing about how many robots or how much work is involved. Read the list as who is experimenting, not who is shipping.

Physical AI by the Numbers: Compute, Power and Market Forecasts

The compute is measurable and the market forecasts are scenarios. The table puts the best-known forecasts side by side, and they do not measure the same thing.

McKinsey's event page on physical AI value pools is the only one of the three to put a number on humanoids' share, and it is small: about 2% of the sector by 2045. So the humanoid headline and the physical AI headline are different bets. One is a forecast for a robot shape, and the other is a forecast for an industry. Treat anything that runs out to 2050 as an analyst's belief about costs and adoption, not as a measurement.

The Number Most Guides Don't Show

NVIDIA says it trained Cosmos on 10,000 H100 GPUs for three months. Nobody turns that into energy, so here is our arithmetic from NVIDIA's disclosed figures. It is an estimate, not a measured result.

The H100 rating of 700 W comes from our H100 specifications guide. Real draw is lower than the rating, and the sum leaves out cooling, networking and storage, so treat 15 GWh as a ceiling for the chips alone.

That training run happens once. The 130 W chip runs inside every robot, every day. If a million robots each ran a Thor-class chip flat out, they would draw 130 MW together and use the same 15 GWh in under five days. That is a hypothetical, not a forecast. It shows why the balance of physical AI compute should shift from training toward the machines as fleets grow.

Why Physical AI Is a Data Center Story

Physical AI creates data center demand at three points, and none of them is the robot itself. The model has to be trained, the training data has to be produced, and the result has to be tested before anyone puts it near people.

• World models and VLAs are trained on GPU clusters, as the Cosmos run shows.
• Video has to be cleaned and labeled before training. NVIDIA says its NeMo Curator pipeline can process, curate and label about 20 million hours of video in 14 days on Blackwell GPUs, against over three years on CPUs alone. That is a vendor claim, but it shows why even the preparation step needs accelerators.
• Policies run through huge numbers of virtual trials before anyone tests them on a physical machine.

The cloud providers have noticed. In March 2026, NVIDIA said Microsoft Azure and Nebius are offering a Physical AI Data Factory blueprint for synthetic data, with CoreWeave supporting Isaac Lab for robot learning and Alibaba Cloud adding NVIDIA's physical AI stack to its platform. Those names are the same companies that build AI data centers, so robotics adds a new tenant to the same buildings. Our explainer on what AI data centers are covers the facilities themselves.

The split between cloud and machine follows latency. A robot cannot wait on a round trip to a server before it steps around a person, so control runs on the chip. Planning, fleet coordination and training can wait, so they run in data centers. Our guide to AI training vs inference explains why those two workloads need different hardware.

What Physical AI Cannot Do Yet

Physical AI works best in narrow, supervised jobs, and four limits explain most of the gap between a demo and a deployment.

• The sim-to-real gap: a policy trained in simulation can fail on a surface, object or lighting condition the simulation got slightly wrong.
• Hidden human help: the IFR says most humanoid applications still need some human teleoperation, and companies rarely publish how often a robot asks for help.
• Where robots actually go: IDC says 69% of humanoids shipped in the first half of 2026 went to research, education, exhibitions, entertainment and government data centers, not paid factory or warehouse work.
• Safety rules: the humanoid-specific standard, ISO 25785-1, was still a working draft in early 2026, so pilots rely on older industrial robot rules.

Coverage keeps making three mistakes. It treats physical AI as a synonym for humanoids, when cars and factory arms are far more numerous. It treats a VLA as the whole robot, when the model is one layer of five. And it treats a smooth demo as proof of reliability, when a demo never shows how often the machine needed help.

The sentence that changed in a year

Jensen Huang said at CES on January 6, 2025, in NVIDIA's Cosmos launch release: "The ChatGPT moment for robotics is coming." At CES on January 5, 2026, NVIDIA's release quoted him saying: "The ChatGPT moment for robotics is here." Fortune's report of the keynote, cited in our humanoid robots explainer, has him saying it is "nearly here." Either way, the humanoid numbers do not look like a moment of that size yet. Waymo grew and the models got better, but about 7,000 full-size humanoids sold in 2025 is a rounding error next to 600,000 industrial arms. Our read: the moment is real for models and for robotaxis, and still a prediction for general-purpose robots.

What to watch

1. Whether Waymo reaches its target of more than 1 million paid rides a week by the end of 2026, a goal co-CEO Tekedra Mawakana has stated. 2. NVIDIA's GR00T N2, previewed in March 2026 for release by the end of the year, which the company says succeeds at new tasks in new environments more than twice as often as leading VLAs. That is a vendor claim until someone tests it. 3. Whether any company publishes how often its robots need human help per hour of work. 4. Whether the standards bodies finish a humanoid safety rule that pilots can point to.

If we had to pick one number to track, it would be the third: how often a robot needs help. It decides whether the economics of a physical AI deployment work, and almost nobody reports it.

Frequently Asked Questions

Related Articles