China Humanoid Robot Output Heads for 100,000 Units in 2026

Factories open faster than buyers sign

Chinese humanoid robot makers are now building at industrial scale. The open question is whether they will find enough paying customers to keep that capacity busy. China's Ministry of Industry and Information Technology (MIIT) used a July press briefing at the World Artificial Intelligence Conference (WAIC) to say that domestic output of complete humanoid robots should top 100,000 units this year. That compares with about 14,000 Chinese-built machines out of roughly 18,000 shipped worldwide in 2025.22 Chinese firms took more than 97% of the roughly 19,100 humanoids shipped worldwide in the first half of 2026, a 272% jump from a year earlier.21

The clearest sign of the shift is the plants themselves. On September 12, UBTECH opened a 14,000-square-meter factory in Liuzhou, Guangxi. It is designed to turn out about 10,000 industrial humanoids a year, and its line is paced to finish one robot every ten minutes.2530 Four days earlier, XPeng said its first Iron humanoid had come off a production line in Guangzhou that it describes as 80% autonomous. XPeng plans mass production by the end of 2026, with deliveries in 2027.26 AgiBot, the Shanghai-based shipment leader, says it has built its 20,000th robot.21

The numbers vary depending on who counts

Everyone covering the story agrees that China leads on volume. The figures behind that lead disagree more than the headlines admit. Smart Analytics Global puts first-half global shipments at about 19,100 units. Counterpoint Research counted more than 22,000 for the same six months, a gap of roughly 3,000 machines.29 Forecasts for the full year also vary. Counterpoint expects more than 50,000 units worldwide, while Smart Analytics Global expects about 60,000.29 Morgan Stanley's latest forecast for Chinese shipments is 50,000.21

The gap between those forecasts and MIIT's 100,000 figure is worth noticing. MIIT is talking about production. The research firms are counting shipments. When a government output target is roughly double the private shipment estimates, robots are probably being built faster than they are being sold. That is our reading, not a reported fact, but the capacity data points the same way. UBTECH's own results show the gap. It sold 1,079 full-size humanoids in all of 2025 and 921 in the first half of 2026, yet it has just added 10,000 units of annual capacity.28 Its humanoid revenue did jump to about 590 million yuan in the first half.28

AgiBot gives a clear picture of how fast production is growing. Its supply-chain chief says it took more than two years to go from zero to 5,000 units of capacity, but only six months to get from 5,000 to 15,000. He credits new suppliers, about 60% of which come from the auto industry or are listed companies.31 The company expects to ship 40,000 to 50,000 units this year.31 By Smart Analytics Global's count, it led the first half with about 8,400 units (44% of the market), ahead of Unitree's roughly 5,900 (31%).29

Public markets get a first look

Unitree's listing on Shanghai's STAR Market gave investors their first audited look at the sector. The Hangzhou company filed its prospectus on March 20. Its listing committee approval came 73 days later, which Chinese financial media called a record, and subscriptions opened on August 10.5 The filing shows something rare among humanoid makers: real profits. Revenue was about 1.7 billion yuan in 2025, up 335%, with a gross margin of about 60%.6 Unitree shipped more than 5,500 humanoids last year. By comparison, Figure AI and Agility Robotics shipped about 150 units each.4

The prospectus also shows who is actually buying. Research and education account for 74% of Unitree's humanoid revenue and shipments. Commercial and consumer buyers account for 17%, mostly using the robots for promotion and display. Industrial uses make up just 9%.9 Unitree plans to spend its roughly 4.2 billion yuan in proceeds on model research, robot development, new products and a manufacturing base.5 That base is meant to reach 75,000 humanoids and 115,000 quadrupeds a year.8 Forbes noted the contrast with UBTECH, the first humanoid stock, which has been listed in Hong Kong since 2023 and still loses about 700 million yuan a year on similar revenue.10

Funding piles into startups with little revenue

Private funding is growing even faster than factory floor space. Chinese embodied AI companies raised 93.5 billion yuan (about $13.86 billion) across 322 deals in the first half of 2026.19 Another tally found 15 startups crossing a 10-billion-yuan valuation in the same period, as disclosed funding topped 46 billion yuan.17 Some rounds were huge. Shanghai-based TARS AI raised a $455 million Pre-A round in April, which it billed as the largest single round in the sector's history in China. State-backed funds in Beijing and Shanghai joined that deal.16 X Square Robot closed six rounds between January and June, and Spirit AI raised about 4.5 billion yuan across four rounds.17

The money is still coming. Home-robot startup Ola Dimensions closed its fourth round on October 13, only seven months after it was founded. It has not announced a launch date for a product.20 Estimates of the totals disagree, and some trackers publish figures that look inflated. Still, the reporting points the same way: investors are paying for early positions rather than proven revenue. One analysis estimates that Galaxy General brings in about $2 million a year against a $2.8 billion valuation.14

The real contest is over robot software

Much of the new money is going into robot software rather than robot bodies. In August, a cluster of rounds went to companies building foundation models, world models and data systems. Noin Intelligence raised $74 million for its GLOW world model, and Discover Robotics raised $100 million less than a month after its first round.19 Unitree has open-sourced two general-purpose models: UnifoLM-WMA-0 in September 2025 and UnifoLM-VLA-0 in January 2026.4 AgiBot runs a 4,000-square-meter "data factory" in Shanghai's Lingang district, where about 100 robots repeat tasks with more than 3,000 objects to produce training data. Nvidia has named AgiBot a partner in its Isaac GR00T robotics ecosystem.39

The buyers are paying for flexibility. Tech Times reports that UBTECH, Galbot and AgiBot hold factory orders from BYD, Foxconn and Airbus. It credits vision-language-action (VLA) models, which let a robot take on a new task from a spoken or written instruction without being reprogrammed.35 The limits are real, though. One industry report says US companies lead China by roughly 12 to 18 months on foundation models. In that view, China's advantage is cost, local parts supply and specific use cases, not intelligence.32 One critical analysis cites a 75% lab success rate on a VLA loading task, a gap a factory manager would notice.27

Warehouses come before homes

Logistics is where these robots are most clearly doing real work. RobotToday called AgiBot's G2 Max deployment with JD Logistics the first confirmed use of a humanoid in a working third-party logistics warehouse. The robot handles round-the-clock palletizing with up to 50 kilograms across both arms.33 Robotera has robots in more than ten logistics centers run by China Post and SF Holding.34 Inside UBTECH's Liuzhou plant, the company's wheeled Cruzr robots handle depalletizing and moving materials. Finished humanoids go into an automated warehouse that holds 112 of them.27

AgiBot says its first factory workstation took about four months to go live, and that later copies took one to two weeks.31 Longcheer, which assembles tablets, plans to grow its AgiBot G2 fleet to 100 units by the third quarter.38 AgiBot's figures have not been independently audited, including 140 continuous hours on a production line at a 99.9% task-success rate.35 Neither AgiBot nor JD has released uptime or cost figures for the warehouse project, so its economics can't be checked against conventional automation.33

What happens next

China has won the first phase, which is building robots in volume. It has not yet won the second, which is getting customers to put them to productive, paid work. A useful test is MIIT's program, run with the state assets regulator, to move humanoids from testing into real factory work at a scale of 10,000 units by the end of 2026.24 Observers describe the deployments so far as real but narrower than the funding headlines suggest. Most are pilots and partial-line trials.34

As we read it, building capacity ahead of demand is a deliberate strategy, not a mistake. When demand arrives, Chinese makers will have the supply chains, lower costs and real-world training data to meet it. Analysts at Counterpoint expect China's 100-plus humanoid makers to consolidate to a few dozen within 12 to 24 months.6 The companies that last will be the ones whose robots prove useful at work. The number they build will matter less.