AgiBot senior vice-president Wang Chuang said the company is in preliminary cooperation with more than one domestic compute firm, and that domestically made chips are going into cerebellum control, joint microcontrollers and communication chips in its robots. He spoke at the World Artificial Intelligence Conference in Shanghai; the remarks were reported by Chinese financial outlet Yicai on July 21 and picked up in English the following day. Separately, Xiong Kun, chief technology officer of AgiBot dexterous-hand subsidiary Lingjiedian, said the MCU in the hand's palm uses a general-purpose part chosen on performance and price, with some customization work running in parallel.
UBTech is on a longer path. Its joint venture with chipmaker MetaX, signed June 11, is targeting tape-out in the second half of 2027 and mass production in 2028.
None of this says Chinese humanoids are leaving NVIDIA.
The substitution is happening at the cheap layer first
Modern humanoids split compute across at least two tiers. A high-level "brain" runs perception and vision-language-action policies and needs GPU-class silicon. Below it, a "cerebellum" and per-joint controllers run balance, gait and torque loops on microcontrollers — deterministic, low-latency, unglamorous parts, bought by the dozen per robot.
It is the second tier that is localizing. MCUs and communication chips are the highest-volume, lowest-margin, most substitutable components in the bill of materials, and automotive-grade Chinese MCUs already exist in volume. Swapping them cuts unit cost and removes an export-control exposure without requiring a domestic answer to a Thor-class accelerator.
Localization is starting where it is easy and cheap — joints and comms — not where it is hard and expensive. — EW analysis
The contrast is visible in AgiBot's own product line. Its A3 Ultra flagship, announced on July 18, runs its high-level embodied-AI stack on NVIDIA's Thor. Unitree, Galbot, EngineAI and UBTech have all been Jetson AGX Thor adopters since the platform launched in 2025. The same firms courting domestic chip partners for joint control are shipping American silicon in the head.
Key Facts
- AgiBot SVP Wang Chuang says the company is in preliminary cooperation with more than one domestic compute firm, per Yicai (July 21, 2026)
- Domestic chips are going into cerebellum control, joint MCUs and communication chips
- UBTech's joint venture with MetaX, signed June 11, targets tape-out in H2 2027 and mass production in 2028
- AgiBot's A3 Ultra, announced July 18, still uses NVIDIA Thor for high-level compute
Why it matters
Component localization is the quietest variable in humanoid unit economics. A full-size humanoid carries dozens of joint controllers; shaving cost and lead time across that layer moves gross margin more reliably than any single headline part. It also changes what an export restriction can reach: control the accelerator and you constrain the fleet's intelligence, not its ability to walk.
What this announcement does not yet establish is a domestic path for the top tier. Until a Chinese accelerator ships inside a flagship humanoid at Thor-class performance, "no need for NVIDIA" describes the joints, not the brain — and the brain is where the capability gap lives.
Frequently Asked
Which Chinese robot makers are moving to domestic chips?
AgiBot senior vice-president Wang Chuang said at WAIC that the company is in preliminary cooperation with more than one domestic compute firm, with Chinese chips going into cerebellum control, joint microcontrollers and communication chips, as reported by Yicai on July 21, 2026. UBTech is on a longer timeline: its joint venture with chipmaker MetaX, signed June 11, targets tape-out in the second half of 2027 and mass production in 2028.
Does this mean Chinese humanoids no longer use NVIDIA?
No. The localization reported so far is at the microcontroller and communications layer. AgiBot's A3 Ultra flagship, announced July 18, 2026, runs its high-level embodied-AI processing on NVIDIA Thor, and Unitree, Galbot, EngineAI and UBTech have all adopted NVIDIA's Jetson AGX Thor platform since its 2025 launch.
Why start with MCUs instead of the main AI chip?
MCUs and communication chips are the highest-volume, lowest-margin and most substitutable parts in a humanoid's bill of materials, and automotive-grade Chinese MCUs are already produced at scale. Replacing them lowers unit cost and reduces export-control exposure without requiring a domestic equivalent to a GPU-class accelerator.