--- title: "Humanoids Went to Work in China First. That's the Whole Strategy" date: 2026-10-19 category: China AI Decoded site: NeuroAI canonical: https://neuroai.site/a/na-decode-humanoids-factory-first language: en --- # Humanoids Went to Work in China First. That's the Whole Strategy > China's humanoid push targets factories, labs and training centres rather than living rooms, and the price curve explains why that order matters. The line runs at walking pace. At the end of it, a machine the height of a short adult lifts a finished door lock, turns the face toward a camera, and holds it still while a vision system decides whether the plating has run. Nobody applauds. The cell is bolted to the side of a working plant, and the technicians treat it the way they treat the conveyor behind them: something that has to keep cycling. That is the shift worth noticing this year. The humanoid (人形机器人) has left the stage and joined the shift roster. ## Key takeaways - China's humanoid output is up 94% in 2026, with Unitree and AgiBot taking about 80% of global shipments (TrendForce, April 2026). - Unitree shipped roughly 5,500 units in 2025 and plans 10,000–20,000 in 2026; AgiBot rolled out its 10,000th unit in late March (company disclosures). - Unitree's H2 lists at about US$16,000–30,000 and the R1 starts near US$4,900 (company and distributor listings). - More than **70%** of Unitree's humanoid revenue comes from research and education; only 9% comes from industrial use (company filings). - MIIT and SASAC target 10,000-unit-scale deployment capability and more than 100 high-value scenarios by the end of 2026 (SCIO). ## The price is the argument Most writing about humanoids treats cost as a downstream variable: capability first, scale second, price third. Chinese vendors have run the sequence backwards. A factory can now buy a machine to test rather than a programme to justify. The public market has priced that bet. Unitree's STAR Market listing was approved in July 2026 at a valuation of about RMB 61 billion (≈ US$8.6bn / HK$67bn), raising RMB 4.202 billion (≈ US$592M / HK$4.6bn) (company filings). The accounts behind the valuation show 2025 revenue of RMB 1.71 billion (≈ US$241M / HK$1.9bn) and net profit of RMB 287.6 million (≈ US$40.5M / HK$316M). The low price is not a subsidy. It comes from suppliers that already serve electric vehicles: motor control, battery packs, actuators, and the contract manufacturers that assemble both. In bill-of-materials terms, a humanoid is an electric car that stands up. That inheritance is why a Chinese vendor can quote a five-figure price while a Western integrator still quotes a project. What the market is valuing is capacity, not output. Unitree plans for 75,000 humanoids and 115,000 quadrupeds a year (company filings). Whether orders exist to fill that line is a question the filings do not answer. ## Where the units actually go Shipments are real and concentrated. TrendForce (9 April 2026) puts China's humanoid output up 94% year on year, with Unitree and AgiBot together taking about 80% of global shipments. Unitree shipped roughly 5,500 units in 2025 and plans 10,000–20,000 this year; AgiBot rolled out its 10,000th unit in late March 2026. Capacity is arriving behind them. UBTech's Liuzhou plant, designed for more than 10,000 industrial humanoids a year, went into production in September 2026 (China Economic Net, 13 October 2026). Then comes the number that should govern how all the others are read. More than 70% of Unitree's humanoid revenue comes from research and education. Industrial use accounts for 9% (company filings). The factory-first story is, at present, an institution-first business: universities, vocational schools, training centres and demonstration halls buy the machines, and factories test a few. The market-size figures are just as easy to misread. China's embodied intelligence (具身智能) market is estimated at RMB 1.09 trillion (≈ US$153bn / HK$1.2tn) in 2026, from RMB 213.3 billion (≈ US$30bn / HK$235bn) in 2018, a compound growth rate of 22–23% (China Embodied Intelligence Industry Development Report 2026). That category is far wider than humanoids. It measures an ecosystem, not demand for a biped. ## The software is the gap In June 2026, MIIT and SASAC launched a real-scenario training programme for humanoids and embodied intelligence, targeting 10,000-unit-scale deployment capability by the end of 2026 and more than 100 high-value scenarios (SCIO, 4 August 2026). Read that as an engineering admission. If generalisation worked, a ministry would not need to organise scenarios. A unit that has learned to inspect door locks has learned the reflectance of chrome under one lamp. It has not learned to dress a cable harness, and it will not transfer without new demonstration data. The programme also does something quieter. It makes the owners of factories, hospitals and campuses the buyers of last resort. Deployment targets are easier to hit when the state can ask its own sites to open a bay. ## What a buyer should test - Cycle time on your part, not the demonstrator's part. - Interventions per hour, and whether anyone nearby can stop the machine. - Whether the gripper, the tooling and the data port are open or proprietary. - What teleoperation costs when the model fails; a remote human is a wage, not a fallback. - Calibration interval, spares lead time, and who trains the on-site technician. A purchaser who runs those five tests will learn quickly which vendors are selling a product and which are selling a position in a supply chain. ## Honest limitations Revenue, capacity and customer mix are company-disclosed and not audited at the unit level. Shipment shares come from TrendForce estimates, and the market-size figure comes from an industry report whose definition includes far more than humanoids. No company here discloses an order book, deployed hours or intervention rates, so the utilisation of the installed base is unknown. The opening scene describes a demonstrator cell of the kind vendors now build, not a named customer line. ## What readers can do now - **Operators:** insist on a paid two-week trial in your own cell, and count interventions rather than watching the demo. - **Investors:** separate capacity announcements from revenue mix; the 9% industrial share is the honest indicator of where demand sits. - **Researchers:** the transfer problem is under-measured. Publish intervention rates and failure taxonomies, not only success videos. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-decode-humanoids-factory-first Free to quote with attribution and a link to the original.