Finished Humanoid Robots

What it is and how big it is

A humanoid robot is a general-purpose robot built on a human frame, aimed first at factory and logistics work. Industry data cited by Inovance put China's humanoid shipments at about 20,000 units in the first half of 2026, up 314% and already above all of 2025 (2026 interim report, p. 18). A robot uses many joints, so the number of motors and reducers sold is far higher than the number of robots (see servo and frameless motors).

Where China stands

Three companies show three different stages.

• Unitree (宇树科技, 688836.SH) is the volume leader by its own account. It says it shipped more than 5,500 humanoids in 2025 and recorded 5,215 sales, earning ¥867.8m from them, and reported a 60.1% gross margin on its main business and ¥278.2m net profit for the year. Its unaudited first-half 2026 revenue was ¥1.15bn, up 48.5% (listing announcement, p. 36). Foreign sales were 43.7% of its main business revenue (listing prospectus, pp. 33, 34, 36, 44).
• UBTech (优必选, 9880.HK) sold 921 full-size humanoid units in the first half of 2026 for ¥590.3m, up 1,445.0%, but still made a loss of ¥338.8m. Its full-size count includes wheeled models (2026 interim results, pp. 1, 2, 6, 12).
• XPeng (小鹏汽车, XPEV and 9868.HK) is a carmaker with a robotics unit that raised over US$900m in August 2026. It expects its IRON robot to reach mass production by the end of 2026 and be delivered from 2027, and has not published unit sales (XPeng announcement).

Who makes the joints

The robot makers do not simply buy finished joints. Unitree says it designs its own joint motors, reducers and drives (prospectus, p. 136). UBTech lists its own servo drives and integrated harmonic joints among its core technology (2026 interim results, pp. 5, 7). Neither filing names an outside joint supplier. Unitree's five largest raw-material suppliers were 22.5% of its purchases in 2025, a low concentration, and the suppliers it names for 2023 to 2025 are small electronics and metal-parts firms, none of them a company on this map (prospectus, p. 150). Most of its top-five slots in 2025 are anonymised, so the prospectus cannot rule out larger names there.

That sets a ceiling on what the parts makers earlier in the chain can sell to these robot makers. Robot actuator makers such as Tuopu and Sanhua have been linked by media to Tesla, but none of the Chinese humanoid makers on this page lists them, or any other company, as an actuator supplier.

Computing, hands and sensing

Two more inputs sit beside the joints: the robot's computing, covered on the robot compute chips page, and its hands and force sensing, covered under dexterous hands and force sensors. XPeng says IRON runs on three of its own Turing AI chips, and UBTech says its Cruzr Y1 uses domestic computing chips (2026 interim results, p. 7). Neither Unitree nor UBTech names its chip suppliers.

What a humanoid costs to make

Only one maker discloses enough to answer this. Unitree's prospectus gives prices, a cost split and a purchase breakdown, all for 2025 unless stated:

• Price. It sold 5,215 humanoids at an average of ¥166,400 each, down 36.1% from ¥260,400 in 2024, which it attributes to a cheaper mix (the mid-size G1 replacing the full-size H1) and price cuts (prospectus, p. 208). It says it shipped more than 5,500, so shipments and recorded sales differ.
• Margin and unit cost. Humanoid gross profit was ¥548.3m on ¥867.8m of revenue, a margin of 63.2%, down from 69.3% in 2024 (our calculation from pp. 33, 212). That implies a cost of about ¥61,300 a robot, down from about ¥80,100 (our calculation). The company says unit cost fell as processes improved (p. 208).
• Cost mix. Across its business, 81.7% of cost of sales was direct materials, 8.1% direct labour and 10.2% manufacturing overhead, up from 72.9% materials in 2024. It says humanoids pushed the materials share up because they are more material-heavy (p. 211).
• What it buys. Purchases of ¥793.8m were 50.7% mechanical parts (machined, die-cast, fastener and plastic parts), 22.4% electronic components (including chips and circuit boards) and 22.2% electrical materials (wire, batteries and similar) (pp. 148, 149). The category definitions name no motors, reducers or screws, and it outsources processes such as rotor winding rather than buying finished joints (p. 151).

The split does not say how much of a robot goes to motors, reducers, screws or chips, and no other robot maker in this chain discloses one. For scale, Leaderdrive's harmonic reducer and metal parts revenue was ¥476.5m on 425,158 reducers sold in 2025, so a reducer sold for at most about ¥1,100 including the metal parts (our calculation from its 2025 annual report, pp. 37, 38). Even a robot with 20 such reducers, an assumed count, would carry about ¥22,000 of them, around a third of Unitree's estimated cost per robot.

Taken together, the figures show a robot maker whose average price fell by about a third in a year while its margin only slipped from 69% to 63%, because its own unit cost fell by about 23% by our estimate. The filings do not say whether any of that cost reduction came from lower prices paid to suppliers.

Margins and risks

Unitree's reported gross margin of 60.1% is well above UBTech's 44.7% and the margins reported at component makers, such as 14.6% on Zhongda Leader's precision reducers or 27.0% on Tuopu's robot actuators. The figures are not like for like, since they cover different products and periods.

Unitree also warns that US trade or certification rules could stop it selling its current models in the US, which was 13.3% of its 2025 revenue (prospectus, p. 23).

Our view: both makers that disclose a margin, Unitree and UBTech, also design their own joint drives, and we have no comparison from a maker that buys complete joints. So the filings cannot show whether in-house design is what earns the margin. What they do show is that neither names an outside joint supplier, and that Unitree buys from a long tail of small suppliers, which leaves little room for any single parts maker to claim a large share of a robot.

What to watch

• Volumes against targets. UBTech describes a target of 10,000 humanoids a year (2026 interim results, p. 8), and XPeng has given no volumes.
• UBTech's loss. Its gross profit was ¥566.9m but its operating loss was ¥279.1m (pp. 1, 2), so operating costs are larger than gross profit.
• Named parts suppliers. None of the three companies names its actuator suppliers. A disclosure from a robot maker would be the first direct evidence of share for any supplier.
• Trade controls. Unitree's prospectus names US rules as a risk.