The Physical AI Landscape 2026: A Market Map by Element Technology

01How to read this map: what counts as physical AI

Physical AI is AI that closes a loop with the physical world: it senses an environment, builds some understanding of it, decides on an action, and moves something — then senses the result. Every product on this map exists to make one part of that loop work: sensors feed it, models think inside it, actuators close it, and simulation and fleet software wrap around it.

That loop is why a map organized by company category alone ("humanoid startups", "chip makers") misleads. The competitive battles are happening at the level of element technologies — who supplies the strain-wave gear inside a shoulder joint, whose tactile skin covers a fingertip, whose world model evaluates a policy before it touches real hardware. So this map slices the field into five layers and 23 element-technology categories, and places companies wherever they genuinely compete, even if that's three boxes at once.

02Layer A — models: the race for a generalist robot brain

The model layer is where physical AI most resembles the LLM boom, and where the largest rounds are being written. Vision-language-action (VLA) models treat robot control the way GPT treats text: one large model, trained on enormous and varied demonstration data, meant to transfer across bodies and tasks. Three approaches now compete. US startups sell the brain itself. Chinese players pair models with open datasets and high-volume hardware. And big tech sells the model layer as infrastructure anyone can buy.

World models are the layer's second half: generative models that predict how a scene evolves, used to imagine, plan, and — increasingly — to evaluate a robot policy in a synthetic world before it ever touches hardware. The category runs from NVIDIA's Cosmos and Google DeepMind's Genie to Wayve's driving-specific GAIA and 1X's humanoid world model.

03Layer B — data & simulation: the field's scarcest resource

Language models had the web to learn from. Robot models have nothing equivalent — there is no internet-scale archive of robot actions — so an entire industry has formed to manufacture that missing corpus. It splits in two. Simulation companies (NVIDIA's Isaac stack, the open MuJoCo and Genesis engines, synthetic-data specialists) generate physics-accurate experience cheaply. Demonstration-data companies collect the real thing: teleoperation services, wearable capture rigs, and — a distinctly Chinese development — city-scale data-collection programs like JD.com's, staffed by human operators piloting robots through millions of task hours.

The two halves are converging into a single pipeline: pre-train broadly in simulation, fine-tune on real demonstrations, deploy, and feed what the fleet sees back into training. Owning any stage of that flywheel is a defensible position, which is why data infrastructure attracts specialist startups and robot makers alike.

04Layer C — components: where the value hides

Strip any robot to its bill of materials and the value concentrates in a handful of element technologies — each of them its own market, with its own leaders, price curves, and challengers. This is the layer where Japan and Europe still hold some of their strongest positions, and where China's supply chain has advanced furthest beyond final assembly. The eight component categories on the chaos map deserve individual treatment.

05Layer D — embodiments: eight ways to put AI in a body

Humanoids get the headlines, and the category genuinely did change character in 2025–26: from staged demos to factories, price lists, and public markets. But the embodiment layer is eight categories wide, and most of the deployed physical AI in the world today doesn't walk on two legs. Warehouse robotics is the most commercially mature market — Amazon passed one million deployed robots in July 2025, and Geek+, Exotec, and AutoStore sell goods-to-person systems globally. Autonomous vehicles are among the largest categories by fleet size and cumulative funding. Surgical robotics, long a near-monopoly of Intuitive's da Vinci, finally saw credible competition clear regulatory gates in 2025–26.

The incumbents are not standing still either: the industrial-arm giants of the last half-century — FANUC, Yaskawa, ABB, KUKA — are retrofitting their installed base with AI, and the sector's ownership is shifting with the technology. SoftBank's agreement to acquire ABB's robotics division for $5.375B (announced October 2025) put one of the 'big four' arm makers inside a portfolio explicitly built around physical AI.

06Layer E — operations: the unglamorous layer that decides profitability

A robot that works in a demo and a fleet of robots that makes money are separated by this layer. Development still runs on the open ROS 2 ecosystem, with commercial platforms building industrial-grade tooling above it. Once fleets deploy, the problems become telemetry, remote intervention, and multi-vendor orchestration — a category where robot-data specialists are becoming the observability stack of the industry. And between development and deployment sits validation: scenario-based testing, functional-safety hardware, and the certification bodies preparing standards for robots that share space with people.

This layer is small in company count and easy to overlook on the map. It is also where every other layer's promises get audited — which is why its players show up as partners in almost every serious deployment story in Layers A and D.

07The regional picture: no country owns the whole stack

Count the companies on this map by region and a structural pattern appears. North America dominates the model and software-infrastructure layers. China is the only country with volume players in nearly every hardware category — from reducers and LiDAR to humanoids and logistics — and its embodiment companies compete on shipped units. Japan's strength sits in Layer C: precision reducers, motors, force sensors, and image sensors, plus logistics robotics — with a government-backed push (METI committed roughly ¥387B in aid) to build a domestic physical AI model layer it currently lacks. South Korea concentrates in batteries, dexterous hands, and manufacturing-backed humanoids; Europe holds niche champions across cobots, inspection quadrupeds, warehouse systems, and the safety-certification complex.

Two caveats keep this honest. The bars below count curated, representative entries — not the full population of each country's industry. And HQ location increasingly understates how cross-border the sector is: American humanoids run Korean batteries and Japanese gears; Chinese arms makers are owned by European groups and vice versa.

08How this map is maintained

A landscape map is only as useful as it is current, so this one is maintained as living data, not a one-off graphic. Every company entry requires verifiable activity — a shipping product, documented deployments, or a disclosed funding round; we don't list stealth companies on rumor. Companies that wind down or lose independence come off the map (Monarch Tractor's 2026 shutdown and Covariant's 2024 absorption into Amazon are recent examples of exits you won't find above). Categories marked "fast-moving" — foundation models, world models, demonstration data, humanoids, hands and tactile sensing, fleet software — are re-checked monthly against primary sources; slower incumbent categories are reviewed as events warrant.

The as-of date on the map is the honest signal: it tells you when this data was last reviewed, and it changes only when a review actually happened.

09FAQ