On 12 November, Xixuan Chuangzhi, an edge-chip company founded by Shenzhen robot maker UBTech and GPU maker MetaX, said it had closed a RMB 500 million angel round at a post-money valuation of RMB 4 billion. The first chip it intends to build carries a single stated goal: to comprehensively surpass Nvidia Thor.
The slogan is loud. The product is still a drawing. Under the company’s own plan, first silicon arrives in 2028.
Why a robot maker goes into silicon
Look at what UBTech has been buying. Walker X ran two Intel i7-8665U processors and an Nvidia GT1030 graphics card. The Tiangong Xingzhe research platform uses Nvidia’s Orin AGX 64 as its development compute unit, rated at up to 275 TOPS, in single and dual configurations depending on version.
From Intel to Nvidia, the high-end AI compute in UBTech’s robots has come from outside. Those platforms were not defined around any one robot maker’s task set, which means model deployment, sensor interfaces, power budget and bill of materials all get negotiated inside constraints somebody else chose.
The physical constraint is real. Bigger models need more compute, and a humanoid has limited volume, battery and heat rejection. Onboard inference meets a power wall; cloud inference meets a latency wall. Xie Yipeng, who leads industrial solutions at Unitree, has made the same point: power limits how far embodied compute can scale, and automotive high-compute designs cannot simply be transplanted onto a robot.
There is a commercial constraint too. Depending on an outside vendor for the top compute layer ties a robot maker’s product cadence, cost and software adaptation to that vendor’s roadmap. Buying a chip answers whether you can run. It does not answer whether the chip fits.
Putting the chip team and the robot team in one company
Xixuan Chuangzhi’s answer is corporate rather than technical: put the people who understand silicon and the people who understand robots inside the same entity. The company was incorporated in May 2026 by UBTech, MetaX and Fenglong. MetaX, founded in 2020 and listed on the STAR Market in 2025, brings full-stack GPU design from architecture and instruction set through to software ecosystem. UBTech brings bodies, systems integration and application scenarios.
The first part is defined as a combined brain-and-cerebellum SoC for robots, built around a general-purpose GPU core, with a toolchain designed to sit close to CUDA so that developers can reuse existing models and applications. Whether migration is genuinely smooth has to wait for the platform.
The company insists this is not a general-purpose chip ported across. The claim is that real robot tasks drive the definition, with multimodal perception, cognitive decisions and motion coordination designed in from the start.
The team reflects that split. Chief executive Yang Xianbo was previously a senior vice president at H3C. Chief technology officer Li Zhaoshi holds a doctorate in integrated circuits from Tsinghua. Hardware vice president Wei Li has taken multiple general-purpose GPU chips through tape-out and volume production. MetaX founder Chen Weiliang chairs the board; UBTech founder Zhou Jian is vice chairman.
Zhou put the intent plainly: the aim is to carry requirements accumulated in robot development and deployment into chip design earlier, and to validate the product against real tasks.
The architectural choice is not unanimous across the industry. Zhongke Huiling has planned a dual-processor X100 plus P100 design for its next humanoid, splitting high-level decisions from real-time motion control. It has also demonstrated X100 handling both classes of work alone, which suggests the two approaches are not mutually exclusive. What settles the argument is real-time behaviour, power, cost and whole-machine validation, not the chip count.
The target is high, the calendar is long
Keep the goal and the facts apart. Surpassing Thor is a line in a research plan, not a result. The company intends to give industry partners a development and validation platform in 2027 for adaptation and testing, then ship the first SoC and a supporting embodied AI solution in 2028.
Today the team is still working through architecture design for its own controller. It has not completed a tape-out.
A company with no product and no tape-out raising at RMB 4 billion reflects what capital expects from edge compute, not what it has priced from delivered performance. The higher the expectation, the heavier the delivery in two years. Tape-out, volume production, software ecosystem and customer adaptation all sit in that window, and a slip in any one pushes the calendar.
Self-sufficiency is the industrial backdrop and it is a real one. It is also not a substitute for competitiveness. A robot chip stands or falls on four concrete things: energy efficiency and real-time behaviour under actual duty cycles, stability of motion control, total cost against the platforms engineers already use, and whether a CUDA-like ecosystem can persuade robot makers and model teams to move their stacks across.
Until then, the 2028 date is the only thing in the plan that can be checked.
Editor’s note: This is an adapted translation of the original Gasgoo report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.gasgoo.com/apps/e060bad9eca91db7f01403fc7120041f/robot/news/70477771/.