The closest business to cash in China’s embodied-AI chain is not the robot maker. It is the integrator that takes a general-purpose body, fits it to a specific factory, connects it to the customer’s systems and keeps it running. That layer is now being squeezed from above by the people who sell the bodies.
The vendors moved first
In May, Unitree opened its UniStore application platform, letting developers upload motion models and customers download and deploy them. In October, Leju Robot published its LeTools skills toolchain, packaging basic capabilities as reusable skills that developers can call, combine and extend. In August, Songyan Power joined Acceleration Evolution and Luming Robot to invest in Oakline, a company building a cross-body operating system and application layer.
Overseas the same shift is visible. Swiss quadruped maker ANYbotics renamed its Data Navigator platform to Shift in October, letting customers schedule fleet missions, review inspection results and push findings straight into maintenance workflows.
Each of those moves gives integrators better tools. Each also removes a line item they used to charge for. A function that had to be custom-built in 2025 is a product feature in 2026, and the customer can see that.
The ROI question the vendors cannot escape
Vendors will not simply absorb the whole integration layer, because doing so is expensive and often unprofitable. Zhao Jiangpeng, co-founder of Feikuo Technology, draws the line at strategic relevance: a vendor can justify an in-house team where the application is tightly coupled to its own hardware, but education, culture and tourism, and general industry each require different customer relationships and service networks.
Agility Robotics co-founder Jonathan Hurst wrote in October that supporting early deployments consumed enormous time and effort, potentially offsetting the profit from those deployments. Transport, customer support and fleet software all become problems the moment a robot is switched on.
That is exactly the work integrators do, and exactly why some of them survive. AgiBot’s Deng Taihua laid out the sequence in April: reach tens of thousands of units in industrial and commercial deployment first, then work with ecosystem partners to go from one to many. Its July overseas expansion folded local logistics, technical support and customer service partners in Australia and New Zealand into its RaaS network.
Three businesses wearing one label
The companies called integrators are already diverging into three different businesses.
The software route is Oakline’s: one operating system, applications for motion generation, grasping and navigation, and portability across body brands so a customer who switches hardware keeps the application layer.
The industry route is Xianheng International’s. Its Qidian Sensing unit upgraded acoustic and infrared products, built an integrated payload for DEEP Robotics quadrupeds and sells through DEEP’s existing channels, with inspection packages for substations, cable tunnels and solar plants already in small-volume delivery. Here embodied AI extends an existing instrumentation business rather than replacing it.
The service route is TAU’s in the United States. It runs modified Unitree G1 units with human teleoperation for home cleaning in San Francisco, designed its own cameras, grippers and onboard compute, and trains models on the video and commands generated by its own jobs.
Zhao expects the label to split further into technical services, industry solutions, software platforms and operations. The test he sets is simple and brutal: if ten projects require ten separate engineering teams, revenue growth and cost growth arrive together, and there is no business model, only a queue of orders.
For anyone underwriting Chinese embodied AI, this is the layer where revenue is real today and margins are least protected. The vendors above are not competitors yet. They are the suppliers deciding, quarter by quarter, how much of the integration business to take back.
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-8321206-8110-30709645.html.