Investors in Unitree and its suppliers should watch capacity, reliability and service networks as closely as model progress, since its founder names them as the bottleneck after a breakthrough.
Unitree Robotics founder Wang Xingxing told the APEC CEO Summit in Shenzhen on November 17 that general-purpose robot models may soon cross a technical threshold, and that the harder problem may then be building enough machines. Output now runs to perhaps tens of thousands of robots a year, he said, without naming a company, while a real breakthrough could call for millions or even tens of millions, according to 21st Century Business Herald. Caixin Global reported the same warning about scaling output.
Wang spoke in a panel discussion alongside Baidu chairman Robin Li, Industrial and Commercial Bank of China chairman Liao Lin and the global chairman of PwC, according to China Ningbo Net, whose Yongpai app covered his remarks. He said humanoid robots are "already close to the critical point of a technical breakthrough," and that if robot AI models make a concentrated leap next year, manufacturing capacity, product reliability and after-sales service would all face a severe test.
The remarks matter because they come from the head of China’s most visible listed robot maker, three months after its Shanghai debut. Wang gave no production target, no date for any model release and no figure for Unitree’s own model spending at the summit, so his comments describe a direction of travel rather than a plan.
Wang names generalization as the barrier to robots in ordinary homes
Wang said the most difficult problem in the field is generalization, a robot model’s ability to handle settings it was never trained on. His goal, as reported by 21st Century Business Herald, is a single model that lets a robot go into any household and start working immediately, without being retrained every time the scene changes. Caixin’s summary of his remarks describes the same hurdle: robots that can work in homes right away.
He linked the current excitement about robots to progress in language models. Text-based AI has become more useful in people’s work and daily lives, he said, so people now want AI to make robots do real work. Asked to sum up 2026 in one phrase, Wang called it "the most exciting" year, and said that much of the technical progress is happening "below the surface," where breakthroughs are arriving faster each day.
Wang also flagged two costs that would grow with any leap in capability. He said the robot industry’s demand for computing power will far exceed what most people now imagine, according to China Ningbo Net, and 21st Century Business Herald reported that he named robot safety as a further challenge. Caixin reported that he described China as strong in manufacturing but weaker in models, computing power and talent, and that he called for cooperation.
Unitree’s capacity plans sit far below the volumes Wang described
Unitree listed on Shanghai’s STAR Market on August 19, 2026, at an issue price of RMB150.80 per share, giving it a market value of about RMB61 billion at the offer price, according to a Securities Times report carried by Tonghuashun. The same report said the company shipped more than 5,500 humanoid robots in 2025 and sold more than 30,000 quadruped robots across 2023 to 2025, and that nearly half of the listing proceeds would go to model research and development.
Unitree’s planned manufacturing base is designed for 75,000 humanoid robots and 115,000 quadruped robots a year once complete. Those figures fit Wang’s description of current output in the tens of thousands. A demand curve of millions of units, the scale he sketched in Shenzhen, would require a different order of factory capacity, component supply and service coverage.
Wang’s point about reliability and after-sales service follows from the same arithmetic. Robots working in millions of homes would need repair networks, spare parts and software support at a scale far beyond a company now shipping thousands of humanoids a year. His remark that compute demand will outrun expectations also points to rising costs for training and running the models those robots would depend on.
A model leap in 2027 is the scenario Wang asked the industry to plan for
Wang framed 2027 as the year a concentrated breakthrough in robot AI models could arrive, and treated production, reliability and service as the problems that would follow. Other Chinese companies are working on the model side of the same problem: Huawei Cloud recently open-sourced a robot world model trained on Ascend chips.
For Unitree, the near-term markers are concrete. The company’s progress on its planned manufacturing base, any change to its stated capacity targets, and the share of its listing proceeds that reaches model development will show whether it is preparing for the volumes its founder described. A Unitree model release that lets one robot perform household tasks without retraining would be the clearest sign that the threshold Wang described in Shenzhen has been crossed.