Physical AI and Humanoid Robots in 2026: Shipments, Market Leaders and What OpenAI, Google, Anthropic and Mistral Are Building

After two years of demo videos, humanoid robots have started to ship in volume. Between 19,000 and 25,000 humanoids were delivered in the first half of 2026, three to five times more than a year earlier, and more than 97% came from Chinese manufacturers. At the same time the AI labs are moving into the physical world: OpenAI will build its own humanoid, Google sells robot reasoning models through the Gemini API, Mistral released its first robotics model and Anthropic measures how fast Claude can program a robot. This overview puts numbers on the market and separates what works today from what is still a promise.

What "physical AI" means

Physical AI is the umbrella term for AI that perceives and acts in the real world: humanoid and wheeled robots, robot arms, autonomous vehicles and drones. What is new compared with classic industrial automation is the software. Instead of programming every movement, robots now run robot foundation models, usually vision-language-action (VLA) models that turn camera images and an instruction in natural language into motor commands. They are trained on large datasets of teleoperated demonstrations and simulation, which is why "data production" has become a business of its own (see the application chart below).

The same building blocks as in language models are used: transformers, pre-training at scale, reasoning models and GPU clusters. That is why the AI labs see robotics as the next market, and why robot makers buy compute on the scale of AI labs. Figure, for example, signed a deal with Nscale in September 2026 for access to up to 100,000 Nvidia Vera Rubin GPUs for training. Our article on AI compute infrastructure explains the economics behind that.

Shipments: the market tripled in six months

There is no official count of humanoid robots, and the three market trackers disagree by about 6,000 units. They do agree on the direction. Smart Analytics Global (SAG) counts about 19,100 units in H1 2026, up 272% from 5,100 a year earlier. Counterpoint counts over 22,000 (+~300%), IDC almost 25,000 (+432%). For the full year Counterpoint expects over 50,000 units and SAG close to 60,000, with industry revenue of about $1.6 billion in 2026 and almost $3 billion in 2027.

Put in perspective: around 50,000 humanoids a year is still a niche compared with industrial robots, of which several hundred thousand are installed every year. The long-term forecasts are bold. SAG expects 500,000 units a year by 2030, IDC over 750,000, and Berg Insight 26 million a year by 2040. Such numbers depend on prices falling and on robots becoming useful in real work, not only in demos.

Who leads: AGIBOT overtakes Unitree

The top five vendors hold about 86% of the market, and all five are Chinese. AGIBOT (Shanghai) grew 562% year on year according to SAG, produced its 20,000th robot in September 2026 and started deploying more than 300 robots at the Chimelong theme parks. Unitree (Hangzhou), the former leader, completed its IPO on Shanghai's STAR Market in August 2026, raising about RMB 6.1 billion ($904 million); it is the first humanoid maker listed in mainland China and had produced about 18,000 bipedal humanoids by July. UBTECH raised its 2026 guidance for the Walker S industrial humanoid from 2,000–3,000 to 5,000 units.

Western companies barely show up in the shipment data. Tesla has not published any Optimus production figures; it was still installing the first production lines in Fremont at the Q2 call in July, and the Optimus V3 unveiling has slipped several times (Q3 results are due on 21 October). Figure AI and Agility Robotics run pilot fleets with industrial customers. Agility presented Digit 5 in September with a 50 lb (about 23 kg) payload and nine-minute charging, with general availability planned for the end of 2027.

What humanoids are actually used for

This is the most important chart for anyone planning an investment. Six out of ten humanoids shipped in H1 2026 went into entertainment and performance or into data production and research, meaning shows, trade fairs, universities and labs that collect training data. Service and guidance (receptions, retail, museums) take 19%. Only 12.8% went into manufacturing and 4.9% into warehousing and logistics. SAG counts more broadly and puts industrial and commercial uses at over 70% of shipments, up from about 50% a year earlier, so the trend points towards real work. But the factory humanoid is still in the pilot phase.

The funding race

Investors are betting on two different things. Figure AI ($39 billion) builds the whole robot, hardware and its Helix model; Helix 2.5 generalised zero-shot to 30 unseen homes in an internal evaluation in September. Skild AI ($14 billion) and Physical Intelligence ($5.6 billion) build the "brain" only: foundation models meant to run on many different robots. Skild released S1 in August 2026, a model that the company says transfers across robot bodies without retraining. In China, the capital comes increasingly from the stock market, as Unitree's IPO shows.

What the AI labs are building: OpenAI, Google, Anthropic, Mistral, Nvidia

The pattern is clear. Google and Nvidia sell models and tools that robot makers can build on today. OpenAI wants the full stack, including hardware. Mistral targets the European industrial niche with small, efficient models. Anthropic's contribution is indirect but practical: Claude is already used to write robot software, ROS integrations and test code, which shortens development for any robotics team.

Europe and the region

Europe will not win the volume game against China, but it has strong players in industrial robotics and a growing physical AI scene. NEURA Robotics (Metzingen) announced a Series C of up to $1.4 billion and is buying companies such as ADLATUS and ACTIVE Shuttle. Agile Robots (Munich) builds the Agile ONE humanoid in series at its plant in Fürstenfeldbruck. Hexagon placed its AEON humanoid in Schaeffler's "Humanoid Gym", with a planned rollout of at least 1,000 units (a commitment, not an installed fleet). Close to Croatia, the Slovenian postal service, Pošta Slovenije, has a partnership with the humanoid maker MagicLab. And Waymo has started testing and mapping in Munich.

For companies in Croatia, Austria and Germany this means two things. First, physical AI is coming through existing suppliers and system integrators, not as a separate purchase decision. Second, the EU rules apply: the Machinery Regulation (EU) 2023/1230 applies from January 2027 and explicitly covers machines with self-evolving behaviour, and AI used as a safety component can be high-risk under the AI Act.

What we recommend

The bottom line

2026 is the year humanoids moved from prototypes to the first tens of thousands of units, almost all made in China and mostly used outside the factory. The money and the AI labs are betting that robot foundation models will do for physical work what language models did for text. Whether that happens by 2030 depends on cost and reliability in real operations, not on demos. For European companies the sensible path is to watch, pilot narrowly and build the data and integration skills now, while drawing the larger, faster gains from digital AI automation.