Unitree's next target is the robot brain, Wang Xingxing says
In Wuzhen on November 3, Unitree founder Wang Xingxing said the company's next focus is robot brains and that it will seek partners globally, extending a three-part plan set in October: volume manufacturing, small models tuned to its hardware, and very large compute clusters.
On the morning of November 3, before the World Internet Conference opened in Wuzhen, Unitree founder Wang Xingxing gave an interview. He said the company's next focus is "robot brains" — and that beyond its own breakthroughs, it will look for partners globally.
Three directions, set a month ago
The framing is not new. At its World Physical AI Conference and partner ecosystem event in October, Wang laid out three lines: scale up manufacturing for demand that could reach tens of millions of units; pursue in-house development and partnerships in parallel to land small physical AI models tuned to its own hardware; and build out very large compute clusters as the base for iteration.
Together the logic holds: bodies have to be buildable at volume, models have to be small enough to run on those bodies locally, and compute has to be sufficient to iterate. Miss one and the other two stall.
"Small models" is the phrase to watch
Unitree is not taking the general-purpose model route. It wants models tuned to its hardware — fewer parameters, faster inference, running inside the robot's chest rather than waiting on a cloud round trip.
The industry tends to split control into two layers: a cerebellum for balance and gait, a cerebrum for understanding and planning. What Unitree has proven over the past few years is mostly the cerebellum — it can run, jump and survive a fall. What it needs now is the cerebrum.
That is the opposite of what Meta and Google are doing. They are pushing general models plus open platforms so developers can drive different bodies with one model; Unitree puts the body first and makes the model serve the hardware. Each has a cost: the general route has stronger scale effects, the dedicated one has tighter control over latency and cost.
Building its own compute has a practical rationale: training and simulation both queue, rented compute is billed by the hour, and over the long run owning is cheaper. That is why "very large compute clusters" sits alongside the other two priorities.
For a company that has already sold a large number of quadrupeds and humanoids, the dedicated route carries a hidden advantage — it holds data from real deployments. Whether a model is small is an engineering problem; whether the data is sufficient is a matter of time, and that is harder to buy.
"Seeking partners globally" is a candid line
Hardware companies lack model capability; that is openly acknowledged across the industry. Building it takes time, and partnerships buy time. Putting "seeking cooperation" next to "technical breakthrough" is an admission that this cannot be done entirely in-house.
Who the partners might be was not said. Chip makers, model companies, overseas research groups — "globally" leaves plenty of room. For a newly listed company that has to keep telling a growth story, this kind of statement is also aimed at the market.
The numbers behind the confidence
Unitree reported revenue of RMB 1.817 billion for the first nine months, up 55.71%, and attributable net profit of RMB 387 million, up 270.54%. R&D spending was RMB 263 million, or 14.47% of revenue.
The trouble with being listed is that the homework comes due every quarter. Hardware revenue can be carried by orders; model progress can only be carried by milestones, and milestones usually arrive more slowly than orders.
Those figures explain two things: why it can run three tracks at once, and why it needs a new story on the model side. The hardware growth curve has a visible ceiling; models and compute are places where investment, and narrative, can continue.
When the brain will actually go into a robot, and what the robot will be able to do then, was not given a timeline in this interview.
The venue fits. The World Internet Conference is one of the few domestic gatherings that puts policy, capital and technology in the same hall, so talking about "robot brains" there reaches more than engineers.
This piece summarizes public reporting and did not independently verify vendor demos or experimental data. Not investment advice.