The Chips Behind "Physical AI" Aren't Data-Center GPUs — What's Actually Inside a Robot's Brain

You've likely heard the term "physical AI" more often since around 2025. It was coined by NVIDIA: CEO Jensen Huang, in his January 2025 CES keynote, described AI's progression from "perception AI" (understanding images and words) to "generative AI" (creating text and images) and now to physical AI. This article lays out what the term actually means and, more concretely, what chips are really moving the robots built around it — based on primary sources only.

What is physical AI?

Physical AI is a term NVIDIA CEO Jensen Huang introduced in his January 2025 CES keynote, describing it as AI "that can perceive, reason, plan and act." NVIDIA frames it as the next stage beyond "perception AI" (understanding images and language) and "generative AI" (creating text and images).

That three-stage narrative, however, isn't consistent even within NVIDIA's own materials. A separate NVIDIA release published the same day as the CES keynote described the stage after generative AI as "agentic AI" (AI that autonomously judges and acts) rather than physical AI, and the company's own glossary page skips the staged narrative entirely, defining physical AI only as "generative AI extended with an understanding of spatial relationships and physical behavior in the 3D world." In short, physical AI is not a fixed, academically settled category — it's a framing NVIDIA uses to talk about its own product line. For how AI chips fit into the broader semiconductor picture, see what AI semiconductors are (in Japanese).

Why does the chip stack behind physical AI split into three layers?

Getting a robot to "see, think, and move" requires three fundamentally different kinds of computation. NVIDIA builds its product line around exactly this split: training, inference, and actuation.

Layer 1: Training and simulation in the data center

The AI model that becomes a robot's "brain" is first trained on large GPUs in a data center. You can't let a real robot break itself millions of times during training, so NVIDIA instead uses virtual-world software — Omniverse, Cosmos, and Isaac Sim — to let robots practice inside a simulation. Ming-Yu Liu, vice president of research at NVIDIA, wrote in an August 2025 company blog post that "physical AI needs a virtual environment that feels real, a parallel universe where the robots can safely learn through trial and error." That statement comes from the team selling the simulation products, so it shouldn't be read as proof that simulation training alone guarantees real-world performance.

Layer 2: The edge-inference SoC mounted on the robot

Once trained, the AI model runs on a chip — an SoC (system-on-chip, a single chip that integrates multiple processing functions) — mounted on the robot itself. This on-robot computation is called inference: using what the model learned during training to decide "what to do right now." NVIDIA's Jetson Thor is the flagship example. In a March 2026 announcement, NVIDIA described Jetson Thor as enabling developers "to move from simulation training to real-world deployment with greater speed, intelligence and reliability," and its product page calls it a "supercomputer for humanoids."

The top module, the T5000, ships with 128GB of memory and a configurable power range of 40-130W, according to NVIDIA's official specs. Compute performance varies sharply with conditions: at FP4 precision (a lower-precision calculation mode that trades accuracy for throughput), using sparse (pruned) computation, in the MAXN power mode that draws the module's full power budget, it reaches up to 2,070 trillion calculations per second (2,070 TFLOPS). Under the same sparse setting but higher (dense) precision, that drops to roughly 1,035 TFLOPS; under the standard 120W power mode it's roughly 1,820 TFLOPS. The headline number matters less than the fact that it swings widely depending on the conditions attached to it. Jetson Thor handles perception and motion planning — the motors themselves are typically driven by a separate, real-time control chip. For the CPU/GPU division of labor more broadly, see CPU vs. GPU for AI (in Japanese).

Layer 3: Power semiconductors that drive the joints

Even after the inference chip decides what to do next, actually moving a joint or motor requires power semiconductors that control large currents with precision. This is where GaN (gallium nitride, a compound semiconductor that handles higher voltage and current in a smaller footprint than silicon) is gaining ground. Texas Instruments (TI) offers GaN-based motor drivers for humanoid joint actuation, stating that "GaN technology reduces the size of the power stage by more than 50% compared to MOSFET alternatives, thanks to integrated drivers." Since a humanoid needs compact, efficient motor control at every individual joint, smaller power semiconductors translate directly into more design freedom for the whole robot.

What is GR00T actually doing inside a robot's head?

The AI model that actually runs on the inference chip is, most prominently, NVIDIA's Isaac GR00T. According to the paper describing the original GR00T N1, released in March 2025, the model has a "dual-system" architecture combining two AI components. A vision-language model ("System 2") slowly interprets its surroundings and instructions, while a diffusion transformer ("System 1", an AI technique well-suited to generating continuous motion) quickly converts that judgment into fluid motor actions — the two are trained jointly, end to end. NVIDIA frames the split as inspired by principles of human cognition, likening it to reflexive versus deliberate human decision-making.

How big is the physical AI / humanoid robot market?

Market-size expectations shape how much chip demand actually materializes. In a January 2024 report, Goldman Sachs projected the total addressable market (TAM) for humanoid robots at $38 billion by 2035, with cumulative shipments of 1.4 million units — more than six times its own $6 billion projection from roughly two years earlier. That figure, however, is now more than two and a half years old as of August 2026, and other research firms' estimates vary substantially, so no single number should be treated as settled.

Japan has its own broad push to grow its semiconductor and AI industries, though we could not confirm a physical-AI-specific government target with primary sources during this reporting. For the national semiconductor targets, see Japan's Growth Strategy numeric targets (in Japanese). We also intentionally left out the semiconductors behind onboard cameras and sensors, and the production plans of individual manufacturers such as Tesla, Figure, and Unitree, since we could not verify either area with primary sources during this reporting.