Zhang Feng read the number out from the World Robot Conference main stage on the afternoon of August 20. China shipped more than 40,000 humanoid robots in the first half of 2026, he said, releasing the 2026 Humanoid Robot Industry Development Report, and China’s share of global shipments has climbed to 97%.
We published a different number ten days ago.
On August 10, Smart Analytics Global put global first-half humanoid shipments at 19,100 units, up 272%, with AgiBot passing Unitree 8,400 to 5,900. China’s share in that dataset was also above 97%.
The two share figures match exactly. The two absolute figures are more than double apart.
The gap sits inside one verb
The question is not who did the arithmetic wrong. It is what the word “shipments” is covering.
Several Chinese institutions have published a consistent figure through July and August: first-half humanoid production in China above 40,000 units, with a full-year run at 100,000. Securities Times used the production framing. A July 20 breakdown split it further, putting first-half mass production above 30,000 units and actual shipments around 23,000.
So 40,000 is the number that carries the label “production” everywhere else. The report filed it under shipments.
That is not a semantics complaint. Built, shipped, and working on a line are three separate gates, and each one drops units. Production minus shipments is inventory. Shipments minus deployment is machines parked at the customer waiting on integration, commissioning, and sign-off. Nothing about a job changes at any of those stages except the last one.
The report’s own sentence is more careful than its headline
Zhang’s framing was hedged. Humanoids are “accelerating from small-batch trials toward routine deployment and large-scale application,” he said, and industry adoption has entered the deep water of full rollout.
Accelerating toward is not the same as arrived at. The distance between them is roughly the factor of two.
The rest of the document is serious work. It runs five dimensions — industry overview, development trends, technology map, supply chain, application scenarios — and compares China, the United States, Germany, South Korea, and Japan across seven deployment categories. It argues that technical innovation has entered a “cluster emergence” phase and that the supply chain now spans chips through finished units domestically. None of that conflicts with what we have seen on factory floors this quarter. Only the headline number does.
What a factor of two looks like
Two anchors put it in scale.
On August 18, Unitree disclosed roughly 18,000 cumulative bipedal humanoids delivered in its listing filing, the first time it had published the figure. That is the company’s ten-year total.
The report says China did 40,000 in six months.
If 40,000 units genuinely reached customers, one half-year cleared more than twice the category leader’s decade. That slope is not impossible, but it needs a demand side to match. We took apart how that demand gets assembled on July 5: China’s push to place 10,000 humanoids into real working roles runs through an administrative channel, not through individual corporate capex. The SASAC-led central SOE robotics consortium being pushed at this same venue this week extends it.
An administrative channel can concentrate orders. It cannot manufacture deliveries. Both things hold at once: the capacity behind 40,000 units of production is real, and the customers behind 19,100 units of shipment are real.
The number that actually touches work is smaller than both
For anyone reading this to gauge their own exposure, neither 40,000 nor 19,100 is the right denominator.
The SAG dataset carried a second line that mattered more. Industrial and commercial deployment went from roughly half of shipments to more than 70% inside twelve months. Multiply it back and roughly 13,000 humanoids entered industrial or commercial settings worldwide in the first half.
Then discount that. Xinhua toured China’s factory humanoids on July 24 and found 97% success separating one fabric type and a robot that still cannot finish a shirt without a person taking over. We counted the automotive side on August 11: nearly every OEM humanoid deployment is doing the same task, kitting parts. A machine that enters a plant usually absorbs one motion at one station, not a whole job.
The chain reads: 40,000 built, 19,100 shipped, 13,000 into industrial and commercial settings, most of them holding a single step at a single station. Every link loses volume.
Why this particular ambiguity keeps recurring
The Chinese term at issue, chuhuoliang, gets used for both “units that left the factory” and “units that left the warehouse to a customer.” English coverage collapses both into “shipments.” Nobody is lying. The word simply does not force the distinction, and no Chinese humanoid maker is currently required to disclose installed base separately from output.
That regulatory gap is closing from one direction only. Unitree had to publish a cumulative delivery figure because a listing filing demanded it. Xiaomi named per-station success rates inside an earnings release it is legally accountable for on August 18. The numbers that survive scrutiny in this sector are arriving through disclosure obligations, not through industry reports. That is worth remembering when a conference-stage figure runs double a filed one.
The report also builds out the softer half of the picture, arguing that ecosystem construction now runs on four coordinated tracks: standards development, talent cultivation, international cooperation, and innovation infrastructure. Talent cultivation sitting in that list is not decoration. It is the same bottleneck showing up from the industry’s side of the table, and it points at the one part of this market where headcount is expanding rather than contracting.
The forward test
The report projects full-year Chinese output at 100,000 units. That gives a checkable prediction for January.
If the year closes near 100,000 units of production and third-party shipment trackers land near 45,000 to 50,000, the two-to-one ratio held, meaning roughly half of everything built in 2026 never reached a customer inside the year. If shipments close much closer to production, the gap was a first-half timing artifact and the deployment curve is genuinely steeper than we have been reading it.
Those two outcomes point at different labor timelines, and they are distinguishable with public data in about five months. Until then, treat 40,000 as a manufacturing statistic.
How to use the report
It is still worth reading. The reading protocol is what changes.
For direction, use it. The five-country comparison, seven scenario categories, and full supply-chain map are the most complete public treatment available in Chinese, from an organization with first-party access to manufacturer data.
For absolute figures, cross-check. From today, any coverage citing “40,000 units in the first half” needs the follow-up question attached: production or shipments? The two words get used interchangeably in Chinese financial coverage, and by the end of that chain readers believe 40,000 robots are already standing in factories.
The test for an individual worker has not moved. It is not how many robots got built. It is whether the station you work at appears on some systems integrator’s delivery schedule. What separates those two questions is exactly the factor of two published today.
Sources
- Report: humanoid robot industry adoption enters the deep water of full rollout — Chinanews, 2026-08-20
- 2026 World Robot Conference maps the embodied-intelligence industry — Chinanews, 2026-08-20
- Production pushes toward the 100,000-unit mark — Securities Times
- First-half mass production above 30,000 units, shipments around 23,000 — Sina Finance, 2026-07-20
