Unitree Says Up to 20,000 G1 Humanoids Have Shipped and Puts Mass Production, Small Models and Compute Clusters at the Center

Unitree used its own stage to answer the question every robotics watcher is asking: how close is humanoid manufacturing to consumer-electronics scale? At the 2026 World Physical AI Conference and Unitree Partner Ecosystem Conference in Hangzhou on Wednesday, founder, chairman and CTO Wang Xingxing said the company has produced and shipped roughly 10,000 to 20,000 G1 humanoids since the model debuted in May 2024 — the first concrete cumulative production figure the company has offered, per coverage from People's Finance and multiple Chinese outlets.

Wang framed the number as a waypoint, not a destination. The company's first stated priority is mass production capability sized for a market of tens of millions and eventually hundreds of millions of units — a scale no robot factory on earth approaches today. The second is a dual-track model strategy: Unitree will keep developing its own physical AI small models tuned to its hardware while running a parallel program of partner-built models, an acknowledgment that no single lab will own embodied intelligence. The third is building out super-large compute clusters to feed the iteration loop, which Wang described as the foundation under everything else.

The most technically revealing slide was about process rather than product. Wang said Unitree is assembling a robot self-evolution system — a closed loop where AI runs its own R&D, simulation training, real-world testing and feedback iteration — designed to replace traditional manual debugging as the main engine of algorithm improvement. The claim positions Unitree's next capability jumps as products of compute and automation rather than headcount, echoing the argument its American competitors make about AI-driven research.

Wang also put a timeline on the field's inflection point, saying there are already visible signs that physical AI is about to break through, and that the breakthrough could arrive within the next one to three years. As a demonstration, the conference showed a G1 humanoid performing a physical sketch drawing under the control of OpenAI's GPT-6 model without task-specific pretraining — a live example of the general-model-controls-robot pattern that has been gaining traction across the industry this year.

The production figure lands in a market that is suddenly talking in billions. The same day, Musk told Tesla's earnings call that he believes the entire humanoid track will become the world's largest industry, with more humanoids than humans on earth eventually, and that Optimus could be delivered to companies other than Tesla once reliability and lifespan problems are resolved by the end of next year. Unitree's own H2 full-size humanoid, unveiled just this week, packs 31 degrees of freedom and a three-hour battery — specs the company says were aimed at the same mass-production logic.

The sober read is that 20,000 units is still a rounding error against the hundred-million-unit ambition, and Wang himself noted the company's output remains far below what mass adoption would require. But the three-track roadmap — factories, small models, compute — reads less like a demo-day pitch and more like a supply-chain plan. In 2026, the humanoid race is increasingly an industrial-policy question about who can ship, and Unitree has just published its answer in units.