The interesting quantity here is denominated in hours, not dollars. Researchers put the requirement for robust, generalisable physical capability at roughly 100 million hours of real-world interaction data, against a present shortfall measured in the hundreds of thousands of hours [12], and the June directive asks for 10,000 units in commercial use by the end of the year [16]. Divide one by the other and each machine carries 10,000 hours, or about 417 days of round-the-clock operation, before the fleet has retired the deficit once [17]. That makes the ninety-plus training centres, most of them state-backed and explicitly built to generate the missing data [13], the actual product being funded: they exist to manufacture hours.
Then the price check. Annualise the first half's $230 million and you get $460 million of government buying [1], divide that by Morgan Stanley's revised forecast of 50,000 shipments [9], and you land at roughly $9,200 a unit [28]. Either the state is directly buying a minority of the units its policy is credited with driving, or the machines are cheap, and the source material does not say which. My reading is the former, or rather the more useful version of the former: the procurement line is the signal that mobilises utility-scale capex and provincial land, and the purchase orders are the smaller half of what it moves. China made between 12,800 and 20,000 humanoids in 2025, some 90 to 95 percent of world output [7], so one country's procurement policy is setting the global unit count.
The demand underneath is not imaginary. An ageing population has produced a labour shortage for which humanoids are the government's stated long-term answer [26], and Made in China 2025 frames the effort as a race with the United States in AI and advanced robotics [23]. None of that makes the hardware competent this year, since current models struggle with basic dexterity and degrade outside the environments they were trained for [11].
This is probably wrong in at least one direction, so here are the directions. Say the data mandate works, generalisable dexterity arrives out of those centres, and $230 million turns out to have been the cheapest capability purchase on the table. Or the hours accumulate but stay narrow, filling the mandate's 100 deployment scenarios [16] with 100,000 units [8] that are each good at exactly one task, which is a machine-tool industry rather than a robotics one and still worth owning. Or procurement flattens while roughly 150 manufacturers burn through it [20], and the consolidation the reporting treats as a matter of when rather than if arrives abruptly [25].
My view is the second, and the thing that would kill it is specific: second-half orders placed by buyers who are neither ministries, provinces, nor state-owned enterprises, at prices set by output rather than by policy targets. Until those show up, the sector's demand curve has a signature on it.
