Chinese factories can now build humanoids by the thousand. Whether anyone can put them to work is a separate problem, and a harder one.
Off the Line Is Not on the Floor
On , AgiBot’s 15,000th robot came off the line in Shanghai and went straight to a phone ODM’s plant in Nanchang. It was a G2, a wheeled dual-arm machine that some trackers count as a humanoid.
Eighty-eight days later the 20,000th came off the line and went to a theme park in Hengqin. Both were delivered. Neither figure says how many hours either machine has worked since.
The word “humanoid” covers more ground in this industry than it does in a robotics lab. AgiBot’s own announcement calls the 15,000th unit a robot, not a humanoid, and its catalogue runs from bipedal machines to quadrupeds to floor scrubbers. Some market trackers count wheeled dual-arm machines like the G2 in their humanoid totals; AgiBot’s production milestones count the whole catalogue. Neither count is wrong. Both cover something wider than the word suggests.
The ramp is real. AgiBot began volume production in and took five months to reach a thousand units. The next 4,000 units took another eleven months. Since December it has added five thousand every quarter, a rate no other humanoid maker has managed.
The market data points the same way. Counterpoint Research counts more than 22,000 humanoids shipped worldwide in the first half of 2026, up almost 300 percent from a year earlier. The five largest vendors were all Chinese and took 86 percent of the total between them. IDC, working from a different sample, puts the total closer to 25,000. This article uses Counterpoint throughout. The gap between them is a reminder that nobody is counting the same way yet.
Four things get counted in this business, and they are confused constantly. Production is units coming off the line. Shipment is units sent to customers or distributors, invoiced or not. Deployment is units installed at a site, which includes the one in the lobby that has never done a day’s work. Productive use is hours of paid work. The first three arrive with press releases and quarterly reports behind them. The fourth has almost no public data at all.
That gap is the story now, and it has moved. The constraint is no longer how fast China can build humanoids. It is whether anyone can put them to work at a price that holds.
Where the Units Go
Counterpoint’s application breakdown covers part of it. Entertainment and performance, together with data production and research, account for more than 60 percent of shipments. Smart manufacturing takes 13 percent. Warehousing and logistics take 5 percent.
Chimelong is what the entertainment tier looks like on the ground. When Chimelong’s Spaceship park in Hengqin reopened on after an upgrade, more than 300 AgiBot machines went across roughly 100 interaction points: shows, guided tours, retail, a hotel desk, a science-education hall. It is a substantial deployment. It is not a factory.
The industrial tier is thinner and better documented. AgiBot’s 15,000th unit went to Longcheer, a phone ODM whose Nanchang plant inspects tablets and which already had eight of the same model on the line. The milestone unit was the ninth.
Those eight ran the line’s full inspection stage during a six-day livestream in late June. AgiBot says they completed 64,828 task cycles over more than 64 operating hours, against 17,625 units of output, a 99.99 percent success rate. That works out to roughly a thousand cycles per operating hour. The company has not said how success was defined or how human interventions were counted.
Longcheer’s own account is more measured. Zhang Long, who runs the company’s pilot with AgiBot at the Nanchang plant, told reporters days later that the robots’ efficiency and first-pass rate had already met the factory’s standard, and that the six days had needed almost no intervention. He also said the arms already doing the job were still better at the inspection itself.
Research and data collection form the least visible tier and, in some ways, the most consequential. Robots in training halls repeat tasks for months, and the footage they generate becomes the raw material for the next model.
AgiBot publishes a large slice of its own. AgiBot World 2026 is out under CC BY-NC-SA 4.0, which asks for attribution, limits use to non-commercial purposes and requires derivatives to carry the same terms. Free to read. Not, on the licence’s plain terms, free to train a product on.
Whether that reading survives a court is another matter, and the uncertainty is probably part of why publishing feels affordable.
What limits all of it is locality. Data gathered in Chinese factories encodes Chinese line layouts, Chinese part numbers, Chinese shift patterns. Whether any of that transfers remains an open question, and the people best placed to answer are the ones least likely to publish.
Three Kinds of Numbers
The three companies that matter here are usually lined up as competitors. They are more usefully read as three different kinds of disclosure.
AgiBot reports production. Its milestones are units off the line, and two of those units happened to be handed to a customer the same day, which is not the same as a fleet running.
Unitree publishes list prices. Its store sells the R1 from $4,900, the G1 from $13,500, the H2 at $29,900 and the H2 Plus at $100,000. Those are published numbers, no negotiation required, and they are why the name turns up in procurement conversations from Detroit to Düsseldorf. They are also prices for machines, not for deployments.
UBTECH reports orders. In it said orders for the Walker series had passed RMB 800 million since the start of that year. That is a contract figure, not revenue and not a unit count, and the space between those three things is where most of the industry’s optimism lives.
UBTECH has deployed the Walker S2 in factories, where its tasks have measurable cycle times. What it has not published is how many of those machines were on shift this week.
A buyer trying to size this market from those disclosures is pricing a house from the listing photo.
The number that would settle it is unglamorous: paid operating hours per unit per month. No vendor publishes it, and no widely cited benchmark tracks it across companies. Until someone publishes comparable operating-hour data, the figures above describe production, prices and commitments far better than they describe demand that persists.
None of the three is a number about work. One counts machines leaving a building, one posts prices, one counts money promised. None of them counts hours.
What $4,900 Does Not Include
The cost of a humanoid is mostly a story about parts. Harmonic reducers, frameless torque motors, planetary roller screws, dexterous hands, force and tactile sensing, depth cameras, lidar. Five years ago the precision end of that list came largely from Japan and Germany. Chinese suppliers now compete across the whole list, at levels that still vary by component, and the change shows up in lead times before it shows up in price.
The price is the part that gets quoted. Unitree’s $4,900 R1 is a real product you can order, and it is the number that made the category feel close. It is also the smallest line on the invoice. Tax and freight are extra. The base R1 is not a turnkey industrial system. Payload, duty cycle and safety validation all have to be measured against the task it is meant to do. Integration, tooling, spares and maintenance are quoted rather than listed, and the parts that wear first tend to be the ones that were hardest to make.
| Line item | What it covers | Notes |
|---|---|---|
| Unit price | Unitree R1, from $4,900 | Published; excludes tax and freight |
| Tax and freight | Varies by destination | Quoted per order |
| Integration | Commissioning, mapping, task setup | Quoted per day |
| End-of-arm tooling | Grippers, suction, fixtures | Quoted per project |
| Spares | Reducers, hand tendons, batteries | Quoted per contract |
| Maintenance | Scheduled service, repairs, remote support | Quoted per contract |
| Safety certification | Required before a line starts | Scope varies by site |
That is where the money goes, and no vendor publishes an average. Until two real quotes are in hand, any payback calculation is invented.
Morgan Stanley’s work on Tesla’s Optimus points to substantial reliance on Chinese components, which complicates a clean China-versus-West framing: a robot designed in California can depend on a supply chain anchored in Shenzhen.
The bottleneck sits in the hand. Tendons, tactile sensors and actuators all have to fit into the same small volume, and all of them wear out. Getting a five-fingered hand through a year of factory work is where the arithmetic stops being friendly.
The Factory Test
Buyers do not measure robots in demonstrations. They measure four things.
Uptime: whether the machine runs a shift, then a week, then a quarter without someone standing beside it.
Generalization: what happens when a part arrives at a different angle, or the bin is a different size, or the light changes.
Cycle time and intervention: how long the task takes, how often a person has to step in, how the cell recovers after a failure.
Cost: against a fixed arm, against dedicated automation, against a wage.
Reuters reported in that Chinese humanoid investment and capacity have been running ahead of both commercial demand and the machines’ demonstrated ability. Read carefully, that is not a verdict on the technology. It describes a manufacturing sector that got very good at building something before anyone worked out what it costs to run.
The questions a plant manager asks are duller than the ones journalists ask. What is the mean time between failures? Who resets it at three in the morning? What does a replacement hand cost, and how long does it take to arrive? None of that appears in a launch video, and all of it decides whether the second order happens.
A fair assessment cuts both ways. A fixed arm is cheaper, faster and more reliable for anything that repeats in one place all day. Two legs and two hands only pay off where the layout moves, the parts vary or the work travels. That is a narrower set of jobs than the shipment figures imply, and a wider one than the skeptics allow.
At Longcheer the calculation is different again. Li Long, who heads the company’s robotics division, says the inspection station runs two shifts and is hard to staff, while a robot lasts two to three years and its speed is closing on a person’s. The case for a humanoid there is not that it beats an arm at the task. It is that the shift cannot be filled.
The Policy Risk Is Now a Line Item
Three separate U.S. mechanisms now stand between Chinese robots and American buyers, and they do different things.
On , the FCC put foreign-produced advanced robotic devices on its Covered List, which governs equipment authorization. New models cannot be certified for sale in the United States. Anything authorized before that date can stay on the market. Manufacturers can apply to the Pentagon for conditional approval for robots, and on the FCC announced that certain devices had received conditional approval and were exempt from the list.
Separately, Unitree appears on the Section 1260H list of companies tied to China’s military. The listing carries procurement and compliance consequences. It is not a ban on commercial sales.
Export controls are a third track, aimed at components and technology rather than finished machines.
For a buyer, the practical question is not whether Chinese humanoids will be prohibited. It is whether the model quoted today will still receive firmware updates and a successor model. Category rules bite on product cycles, not on the installed base. The exemption route means the answer depends on case-by-case approvals, and a buyer has no way to know which way a given case will go.
The next milestone worth watching is not another factory coming online. It is a customer who has run a fleet long enough to know what it costs, deciding to buy a second one.
That is a duller headline than a production record, and a better test of whether any of this is an industry.
AgiBot’s 20,000th unit went to a theme park in Hengqin. Nobody has published how many hours it worked last month. When someone does, the argument about Chinese humanoids will finally have a number in it.
