Chinese humanoid robot maker Unitree Robotics, one of the most-watched names in the country’s booming embodied AI sector, has poured cold water on suggestions that a “GPT moment” for robots is imminent. Speaking at Nikkei’s Global Digital Summit 2026 in Tokyo, Irving Chen, Unitree’s director for Asia-Pacific and other regions, said the industry remains 2 to 5 years away from the kind of transformative leap that ChatGPT delivered to generative AI in late 2022.
“The LLM is powerful in the virtual world,” Chen said during the summit, arguing that today’s large language models remain poorly connected to robots and machines that must actually operate in the physical world. According to a report by Nikkei Asia, Chen’s diagnosis is that the hardware side of the industry has moved faster than the software brains needed to make humanoid machines commercially useful, a candid admission from one of the field’s most celebrated companies.
The comments land at a moment of extraordinary hype. Unitree’s viral videos of humanoid units performing synchronized dances, martial arts routines, and acrobatic flips have made the Hangzhou company a global symbol of Chinese robotics prowess. But Chen argued that those spectacles primarily demonstrate that the hardware is reliable and mass-producible, not that the machines are intelligent enough to perform genuinely productive labor. The gap between choreographed demonstrations and useful autonomous work, he suggested, is still wide.
The physical-world problem
Chen’s central technical point is that language models were trained on the near-infinite corpus of text and images available on the internet, but robots must reason about a far messier set of variables: friction, gravity, unexpected obstacles, lighting conditions, human bystanders and countless edge cases that cannot be scraped from the web. Wider adoption, he said, will require robots to reliably carry out longer, more complex tasks in changing environments without being pre-programmed for each scenario.
That gap is reflected sharply in commercial demand. Chen cited industry data showing that in the first nine months of last year, roughly 74% of humanoid robot revenue came from research and education customers, with only 9% originating from industrial buyers. “Right now, most robotics projects are in the pilot stage,” he said. Unitree’s own roadmap assumes proof-of-concept trials will dominate through next year, with meaningful commercial deployment arriving around 2030, followed by broader general labor assistance thereafter.
Two ingredients are needed to close the gap, Chen said: more capable robot “brains”, software models that can interpret surroundings in real time and translate high-level instructions into precise physical actions, and large-scale, low-cost manufacturing. Neither is trivial. On the software side, world models and vision-language-action systems remain at the research stage. On the hardware side, unit economics still favor niche buyers over mass deployment.
An IPO built for the long game
Even as Chen tempered expectations for the technology, Unitree is preparing to raise a very large sum of money to bet on exactly that long game. The company, founded in 2016 by Wang Xingxing, received regulatory approval earlier this month for a Shanghai IPO that aims to raise 4.2 billion yuan (roughly $620 million).
The allocation of proceeds is telling. According to Unitree’s prospectus, 48% of the funds will go directly into AI and embodied intelligence research, a clear acknowledgment that solving the “brain” problem Chen described is the company’s binding constraint. Another 26% will fund new hardware development, and 15% will finance new factories with capacity for roughly 20,000 robots per year. Unitree shipped 6,500 units last year, so the manufacturing expansion implies a roughly threefold jump in capacity even before demand materializes at scale.
Unitree’s ultimate ambition, Chen said, is “one model that can fit most scenarios” across homes, offices, and industrial sites, a general-purpose robotic foundation model that would represent, in effect, the robotics industry’s answer to GPT. Reaching that milestone is what Chen believes is still two to five years away.
A crowded, cash-hungry field
Unitree’s caution is notable because the surrounding sector has been operating on a very different narrative. Chinese humanoid robot models now exceed 400, more than half of the global total, as EastFrontier reported this week in a survey of the field’s explosive growth. Competitors are racing to capital markets at an unprecedented pace: AgiBot has tapped CITIC, CICC, and Morgan Stanley for a Hong Kong listing, while world-model startup Psibot recently closed a round at a $1.48 billion valuation, and GigaAI has filed for a Hong Kong IPO of its own.
Investors have largely accepted the premise that embodied AI is the next major platform after LLMs. The valuations reflect that faith. But Chen’s remarks suggest that operators within leading companies see a longer, more grinding road than public markets currently price in. His two-to-five-year window would push a genuine commercial inflection point to somewhere between 2028 and 2031, beyond the horizon of most venture funds and well past the IPO lockup periods now being negotiated.
That mismatch could become a source of tension. Moonshot AI, another Chinese darling, has already demonstrated how quickly hype can outrun operational reality: the company was forced to pause new subscriptions to its Kimi K3 model earlier this month due to a compute shortage. Robotics companies face an even harder version of that problem, they must scale not only servers but also physical supply chains, service networks and safety validation.
A measured pitch from a leader
Unitree’s willingness to publicly downplay the timeline is, in some ways, a competitive statement. By anchoring expectations to a multi-year horizon while simultaneously raising $620 million and expanding manufacturing capacity, the company is positioning itself as the mature, patient player in a field increasingly populated by startups pitching near-term breakthroughs.
For Chinese policymakers, who have identified embodied intelligence as a strategic priority in the coming 15th Five-Year Plan, Chen’s comments are a useful reality check. Beijing’s industrial planners have been directing significant state capital into humanoid robotics, and local governments across the country are competing to attract robot manufacturers. If the commercial payoff is still five years away, the state’s patience and its willingness to keep funding losses will be tested.
For now, Unitree’s message is that the choreography is easy and the intelligence is hard. The GPT moment for robots is coming, Chen suggested. It just isn’t coming this year or next.