Physical AI in one definition
Physical AI is artificial intelligence embodied in a machine that senses its environment, makes a decision and produces a physical action. A robot, autonomous vehicle or drone needs a model, but it also needs sensors, real-time compute, power, connectivity, motion control, safety systems and a manufactured body. That full dependency chain is the Physical AI stack.
The definition matters for investors because value does not have to accrue to the company whose logo is on the finished machine. Suppliers can earn durable economics when their component is difficult to qualify, their software owns a recurring workflow or their manufacturing capacity is hard to replace.
Physical AI versus generative AI
Generative AI produces digital outputs such as text, images, software or audio. Physical AI closes a loop with the real world: sense, interpret, plan, act and measure the result. The physical loop adds latency, energy, reliability and safety constraints. A plausible answer is not enough when a machine is moving through a factory, road or operating room.
The categories overlap. Generative models can help a robot understand instructions, plan tasks or learn from simulation. Physical AI adds the hardware and control systems required to execute those decisions under real-world uncertainty.
The ten-layer Physical AI stack
PXS Research maps the system across AI training infrastructure; grid, power and thermal systems; critical materials; manufacturing and industrial automation; edge compute and control; connectivity and positioning; perception and sensing; actuation and motion control; simulation; and autonomy platforms and end markets. The layers are connected, so a bottleneck in power, sensing or manufacturing can limit adoption even when model capability improves.
This stack is a research map, not a claim that every supplier has meaningful robotics revenue. Company pages separate direct disclosed exposure from optional exposure and identify the evidence that would strengthen or invalidate each thesis.
How investors can get exposure
Public exposure exists through compute, industrial automation, machine vision, sensing, power, precision components and workflow software. Private companies dominate several headline categories, including general-purpose humanoids, so public-market research should not invent a direct proxy where none exists. The better question is which public suppliers sell qualified tools across several potential winners.
PXS Research treats current signals as dated research states rather than recommendations. The companion stock hub applies inclusion rules, shows the live research set and links each company to its thesis, risks and last-reviewed date.