China accounts for 97 per cent of global humanoid robot shipments. It also has a structural problem that shipments cannot solve: once a general-purpose brain becomes good enough to drive any body that meets a standard interface, the body stops being an asset and starts behaving like a peripheral.
The economics that started the argument
SenseTime’s Astra model is genuinely capable at the planning layer. It scores 100 per cent on LIBERO-Long++ and 81.9 per cent on RoboDojo, which means it can map tasks to actions across different bodies. The company’s own testing also produced the more awkward number: USD 19 to USD 55 per task, and 131 to 263 minutes per task, with force-feedback breakpoints appearing in real hardware. Cloud-hosted general models cannot close a millisecond control loop at those costs.
So every serious body maker is now building its own execution layer. The question is what they hold once they have it.
Three cards, one moat
Chinese robot makers hold three apparent advantages: manufacturing supply chain and cost structure, the largest set of real deployment scenarios in the world, and that 97 per cent shipment share.
Two of the three are already eroding. Whole-machine prices have fallen nearly 70 per cent in two years, with volume models now in the tens of thousands of dollars. Cost leadership is no longer a moat when your competitors buy the same components from the same cluster. The shipment share sits on top of a technology gap: a CCID report notes that core algorithms and high-end chips remain constrained, and public estimates put domestic self-sufficiency in core algorithms below 40 per cent. China ships the volume. It does not yet hold the algorithms.
That leaves physical interaction data, generated daily in real sites, as the only asset a general model cannot scrape from the internet. But data only becomes a moat if it flows back into your own model. A site that does not produce data is just a location.
Who is buying a body and who is growing one
The three best-funded players have taken three different routes, and all three keep the data inside.
Unitree is turning the body into a flywheel. It has deployed GPT-6 on its robots for task testing, but Wang Xingxing has stressed that real-machine data matters more, and the company plans a general embodied foundation model within three years, built on a large real-world dataset with automated annotation.
AgiBot has gone the other way and pushed the money into software. Three quarters of its research headcount and more than three quarters of its research budget go to the brain and cerebellum programmes, and it describes itself as an embodied foundation model company rather than a hardware vendor.
Galaxy General trained its cerebellum on 2 billion frames of motion-capture data. AstraBrain-WBC 0.5 reports 92.58 per cent zero-shot motion tracking with 0.39 millisecond inference, a jump from 2 million frames to 2 billion in one generation.
The common thread is that none of them is merely using a general model. Each is borrowing it while keeping data, execution layer and iteration speed in house. In their hands the body is a data inlet, not a physical interface.
The middle of the market is the exposed part
Below that tier the picture is different. Most Chinese body makers buy their embodied software stack. Public industry counts suggest that among robot companies with funding in the hundreds of millions of renminbi, only a minority have reached thousand-unit shipments. The rest are at prototype stage, assembling purchased parts around purchased intelligence.
A body assembled from purchased parts and purchased software is, by definition, replaceable. That is the structural weakness behind the shipment number, and it is why so much of the current investment is going into control layers rather than into new mechanical designs.
The Chinese robotics industry has spent three years proving it can build bodies at scale. The next three years will decide whether those bodies own anything the brain cannot replace.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://robot.ofweek.com/2026-11/ART-8321206-8420-30709634.html.