Robot Compute Chips

What the part does

A humanoid robot needs two kinds of computing. One handles real-time motion control, vision and path planning; the other runs the AI models that perceive, reason and decide. Rockchip describes this split as a "cerebellum" and a "brain": its main chips take the first role, and coprocessors such as the RK182X take the second (2026 interim report, p. 12).

The chips matter to the rest of the chain because they decide how much of a model can run on the robot itself. UBTech says it has built its on-device inference stack with hardware and software co-design to fit its models to the chip (2026 interim results, p. 9).

Where China stands

The filings show three approaches:

• General-purpose chip designers. Rockchip (瑞芯微, 603893.SH) says several robots using the RK182X were released and entered volume production in the first half of 2026, and that its chips are used in robot dogs and in industrial, service, companion, sports and agricultural robots. First-half revenue was ¥2.88bn, up 40.6%, and it reports no robot revenue (pp. 6, 10, 12).
• Automotive chip makers reaching into robots. Horizon Robotics (地平线, 9660.HK) says its associate D-Robotics has empowered over 100 robot categories, serves over 400 customers and covers more than half of China's embodied-intelligence companies. D-Robotics was deconsolidated on 31 March 2026, so Horizon's own results do not show its size (2026 interim results, pp. 1, 4).
• Robot and car makers building their own. XPeng says its IRON humanoid runs on three in-house Turing AI chips with up to 2,250 TOPS of computing power (XPeng announcement). UBTech says its Cruzr Y1 wheeled humanoid is built on a base of domestic computing chips, without naming them (2026 interim results, p. 7).

What the filings do not show

None of these filings gives robot chip revenue, unit shipments for robots or the share of a robot's cost that goes to chips. The foreign supplier most often discussed for robot computing is Nvidia, and UBTech's report says its inference link was rebuilt on TensorRT, an Nvidia software library (2026 interim results, p. 9). Neither it nor Unitree's prospectus says which chips or how many they buy from abroad. Unitree does report that about 20% of its raw material purchases are imported materials bought through domestic agents, which it flags as exposed to export control changes (prospectus, p. 24).

Our view: this is the part of the chain where the filings say least. The suppliers describe reach (customers, categories) rather than revenue, and the two public humanoid makers disclose no chip suppliers, so the size and the import dependence of robot computing in China cannot be read from these reports.

What to watch

• Robot revenue disclosure. A first reported robot revenue line, from Rockchip or Horizon, would turn the claims about reach into a measurable market.
• Named chips in robot makers' filings. UBTech's and Unitree's reports are the likely place where chip suppliers would first appear.
• In-house chips. XPeng's Turing chips show carmakers moving into robot computing, which would leave less for outside suppliers.
• Supply. Rockchip says it is stocking ahead because memory and other upstream supply remain tight (p. 12).