The buyers behind humanoid robots' biggest orders are not factories

In the fourth quarter of 2025, a leading humanoid robot company signed six big orders in a row, worth about RMB 829 million in total.

Not one of the buyers was a factory. All six were data collection centers set up by local government platforms.

The company’s 2026 interim report lists industrial names such as Airbus, BYD and Foxconn among its partners. But it does not break the number down: of the 921 full-size humanoid robots sold in six months, how many went onto production lines and how many went into data collection centers.

1. First, who is actually buying humanoid robots

• Shows and research. According to IDC, about 18,000 humanoid robots shipped worldwide in 2025, and entertainment, education and research, and guiding accounted for more than 85%
• Data collection centers. According to industry media, 64 were in operation across China by April 2026, with more than 2,500 robots deployed
• State-owned R&D. In 2025 China Mobile ran a tender worth about RMB 124 million; the buyer was its subsidiary that does embodied-AI research

Demand for shows is real, but rental prices keep falling: the daily rate went from RMB 10,000 to 20,000 in early 2025 to RMB 3,000 to 5,000 after the 2026 Spring Festival. By this October, reports put it below RMB 1,000.

2. On the factory floor, the real output is data

On the carmaker side, what can be found in public is letters of intent. The carmakers’ own description is that robots on the shop floor are still in small-scale on-site training and data accumulation. I have not seen any carmaker reorder a hundred units or more.

My reading: a real production line is, for now, closer to a data mine. Capacity is a by-product.

The data does sell. Real-robot data goes for hundreds to over a thousand RMB per hour. The teleoperation operator earns about RMB 30 an hour, standing behind the robot and guiding it through each motion. But one report describes operators working 10 to 12 hours a day and ending up with only 4 to 5 hours of usable data. A full-scale training site costs tens of millions to over a hundred million RMB to build, and by the estimate in that reporting, even if every hour of data were sold, data sales alone would need more than three years to pay it back.

Compare a mature category: the prospectus of a warehouse mobile robot company shows a customer repurchase rate of about 70%. Repurchase is what tells you customers are paying for capacity.

3. Why the productivity math does not close yet

• Speed. According to reporting on a fixed-station test at Xiaomi’s car factory, short moves reach 70 to 80% of human speed, and tasks longer than one minute only about 30%
• Success rate. Industrial sites generally require 99% or higher. If the previous station delivers a part every 15 seconds, the robot cannot take 30
• Consumables. A dexterous hand costs over RMB 100,000 and lasts only a few weeks to two or three months in continuous work
• Form. Factory floors are flat. Rodney Brooks, who has built robots for decades, expects successful “humanoids” to have wheels instead of legs. A factory in Changchun already uses wheeled robots for depalletizing: 80 cases an hour, running 24 hours

There are signs of progress too. At one loading station in Xiaomi’s die-casting shop, the robot ran 3 hours straight at about 90% success in March 2026, and reached 98% by July; a human at the same station is at 99%. But that is one station, after a quarter of tuning.

4. So how to judge: watch five signals, not videos

① At one standard station, success rate stays above 99% for a full shift.

② No teleoperation, no on-site engineers. How to check: the acceptance document states how many robots and how many people watching them.

③ Hourly cost per robot is below local labor.

④ The same customer reorders a hundred units or more.

⑤ The financial report splits out revenue from industrial customers, excluding government data collection centers.

Until then, what a factory can use is the robot’s eyes. I once compiled a list of scenarios in warehouse logistics: of 28 scenarios, 23 are about taking a picture, recognizing and judging, and only 4 need hands.

So my advice: in the pilot contract, write these five signals in as acceptance clauses. If the vendor will not sign them, treat the project as a marketing and data budget, and do not count it as capacity.

---

Of the five signals, which one would your shop floor hit first?