China's Humanoid Robot Boom Hits a Data Wall

At the World Humanoid Robot Games in Beijing in August 2026, a bipedal machine sprinted 100 meters in 8.64 seconds. A few blocks from the venue, a different humanoid pulled espresso shots in a fixed-layout coffee kiosk. The two images traveled the same way around the world: as proof that the humanoid era had arrived. The less visible fact is that both robots only worked because their environment was pre-mapped and their task was repetitive. The next leap past that ceiling now depends on data the world has barely produced.

That is the contention of a recent analysis by Recode China AI, a newsletter on the Chinese AI industry. The field's bottleneck, the analysis argues, has shifted from actuators and chips to the volume of real-world physical-interaction data available to train a robot's "brain," the vision-language-action (VLA) model that turns camera input into motor commands. The frame borrows from the LLM playbook. Today's text models were trained on trillions of tokens scraped from the open web. The equivalent for robots is hours of human teleoperation, ego-centric video of people doing chores, and other forms of motion data. By the most widely cited estimate, the world has produced only about 500,000 hours of high-quality, real-world physical-interaction data, roughly 1/20,000th of the data used to train basic LLMs. The figure circulates across Chinese media; the Recode China AI author was unable to verify its original primary source, and it should be read as an order-of-magnitude claim rather than a benchmark.

YouTube alone uploads around 720,000 hours of video a day. The largest public robot dataset, Open X-Embodiment, holds between 1,150 and 2,000 hours of actual robot operation. The contrast is not a hundredfold. It is six orders of magnitude.

If the bottleneck is data, the natural question is which supply strategy is most likely to work. The clearest scaling-law evidence published so far comes from Robbyant, an embodied-AI lab at Ant Group. Robbyant reports that pre-training a VLA model on 3,000 hours of robot data delivered large downstream gains, that scaling to 20,000 hours produced further gains with no saturation, and that a second-generation model is now being trained on 60,000 hours. The numbers are company-reported and come from a privileged incubator, so they sit at the optimistic end of the curve. They also suggest that the relationship between data and robot capability is the kind of curve that rewards whoever can keep pouring data in.

A different supply strategy is already visible on the ground in China. Rest of World reported that by December 2025, more than 40 state-funded humanoid training centers had been announced across the country and roughly 24 were already operating. The Beijing Shijingshan site run with Leju Robot covers more than 10,000 square meters across 16 scenarios. A People's Daily profile of the Shijingshan Phase II facility lists concrete training parameters: 110 staff, 7 to 8 days to teach a single scenario, 1,250 repetitions to teach a humanoid to put a pan on a stove, a reported task-success rate above 95%, and several million data entries generated per year. The numbers are the facility's own, not independent measurement, but they make the operational cost of the data-factory model legible. The bottleneck is not the robots. It is the rooms, the staff, and the calendar.

A third approach bets that the data already exists in human video. Figure AI, the U.S. humanoid maker valued at $39 billion, released "Index" in 2025: a crowdsourced dataset of 16 million first-person videos. Peking University's DAGroup released "HumanNet" in May 2026: a one-million-hour human-centric video corpus mixing first- and third-person footage. The bet is that a sufficiently large pile of human manipulation video, combined with the right pre-training recipe, can substitute for hours of robot operation. Carnegie Mellon's VideoManip project is an example of researchers trying to extract robot-action signal from arbitrary human manipulation video without specialized rigs. If either of the egocentric-video bets works, the data wall is shorter than it looks. If it does not, the field falls back to the slow, room-based, staff-heavy model.

The capital backing all three strategies is large enough that a bubble warning is now part of the public record. IT Juzi, a Chinese venture-data provider, tallied more than 63.4 billion yuan (about $9.4 billion) invested into Chinese robot-brain and embodied-AI startups between January 2023 and August 2026. China's National Development and Reform Commission issued a rare bubble warning in November 2025. More than 150 humanoid companies now operate in China. Rest of World quoted Marco Wang of Interact Analysis and Carnegie Endowment analyst Pavlo Zvenyhorodskyi flagging overcapacity risk. UC Berkeley's Ken Goldberg told the outlet that the cost-effectiveness of the data-factory model is unknown and that even hundreds of workers will not close the gap quickly.

The robots in Chinese coffee shops, pharmacies, and traffic-management roles are running in fixed layouts with repetitive, pre-defined tasks, the data-rich narrow cases rather than the general-purpose humanoids that headline reels imply. Goldman Sachs projects the global humanoid market could reach $38 billion by 2035, with shipments on the order of 250,000 units a year by 2030. UBTech sold 566 million yuan (about $80 million) of humanoids to three data-collection centers, and China Mobile placed 124 million yuan (about $17.6 million) of orders with Unitree and AgiBot in July. The orders are real. The general-purpose humanoids they imply are not here yet.

The race to escape the demo is now a race to produce data at a scale the world has not yet produced. Whether that scale is reached by state-funded training centers, by crowdsourced egocentric video, or by the next 10x jump in Robbyant's curve will determine which humanoid companies, in which country, are still standing by 2030. The first scaling-law curves are in. The bubble warning is on the record. The next data point to watch is whether the egocentric-video bets, Figure's Index and Peking's HumanNet, produce a humanoid that can leave a pre-mapped room.