The humanoid supply chain is already Chinese: 18,500 of 19,100 units in six months

Subtract 18,500 from 19,100 and the entire non-Chinese humanoid industry fits inside about 600 units for the half year ^[1]. That figure is generous, because SAG states the Chinese number as a floor rather than a point estimate ^[1]. AgiBot alone averaged roughly 1,400 units a month over the same period, against about 100 a month for every non-Chinese vendor combined ^[6]. On SAG's numbers the ratio of Chinese to non-Chinese output is close to 31 to 1 ^[5].

The sourcing problem does not stop at the platform vendor. SAG's own explanation for the gap is component depth: China already dominates global manufacture of the electric motors, sensors, batteries and precision actuators that go into these machines, so a builder in Shenzhen or Shanghai has suppliers a short drive away rather than a freight container away ^[8]. A buyer who insists on a non-Chinese badge on the chassis is still buying a bill of materials assembled from that base. Domestic assembly changes the label and the lead time; it does not change the dependency.

The caveat SAG attaches to its own headline is the interesting part, and it cuts in an awkward direction. The firm says the overwhelming majority of units shipped are going into pilots, research labs, demonstrations and early deployments, with very few running autonomously on work that justifies their cost against a robotic arm or a wheeled platform ^[12]. So the 97% measures placements, not productive hours. But pilots return data, and 18,500 fielded machines return more failure modes than 600. The qualifier that softens China's lead in value creation quietly compounds its lead in learning.

The forecast deserves the same arithmetic scrutiny. SAG projects roughly 60,000 units for full-year 2026, which implies about 40,900 in the second half, or a little over twice the first-half run rate ^[3]. Its 2030 figure of 500,000 is framed as nearly 100x growth from the 2025 baseline ^[7][14]. The baseline doing that work is 5,100, and 5,100 is the first-half 2025 number, not the year ^[3]. Dividing a full-year 2030 projection by a half-year 2025 actual gets you 98 ^[4]. Against a genuine full-year 2025 base, the multiple is materially smaller. Anyone sizing capacity or a component contract off that slope has inherited a denominator choice, not a measurement.

Then there is the capital question. Both AgiBot and Unitree are reportedly preparing for possible IPOs ^[11], and Beijing has already designated humanoid robotics a strategic technology, with subsidies and research grants at national level and industrial parks and tax breaks from provincial and municipal governments ^[9]. Two vendors holding about 75% of shipments ^[2] will be raising public money against a growth story that Western pilot orders help validate.

Tesla's Optimus is the cleanest illustration of the position Western buyers are in. SAG notes it has substantial internal demand inside Tesla's own factories but has not shipped at volumes that register against Chinese output ^[13]. Captive demand is not a supply chain. For an operator writing a 2030 line plan, the live question is whether a machine whose actuators, spares and firmware trace back to two firms in Shanghai and Hangzhou ^[4][5] is acceptable, and what the answer costs if it is not.