AINEWS
Hardware is largely solved; the binding constraint is now training data, and China is the frontrunner for the kind only humans and robots can capture.
On August 17 in Beijing, a humanoid robot called Superman hit 12.66 meters per second on the track, faster than Usain Bolt's peak sprint speed, before veering off and slamming into a wall. Two days later, Unitree's shares leapt 460% on their Shanghai debut, and the Beijing Humanoid Robot Games opened in the capital.
Underneath both events sits the same shift. The hardware race in Chinese humanoids is effectively over; the next sprint record is incremental. The binding constraint on the category has moved from motors, actuators, and capital to the training data needed to make those machines do useful work, and that is the race China is now positioned to win.
Chinese makers are projected to ship about 50,000 humanoid units in 2026, more than triple the 2025 volume per the livemint/Economist-style analysis. TrendForce, in an April 9 release, pegged Chinese 2026 output up roughly 94% year over year, with Unitree and AgiBot together holding close to 80% of the market. In a separate August 19 note, the research firm framed 2026 as the year commercial validation begins and sized the Chinese market at CNY 15 billion (roughly $2.1 billion, USD approximation at then-prevailing rates).
Those numbers describe a category that knows how to build robots. They do not yet describe a category that knows how to make money from them. The bottleneck now lives in software, specifically in the foundation models that have to convert a torrent of motion into something like physical common sense. Today's humanoid foundation models run on a few billion parameters. To replicate human-body functioning at anything like chatbot-LLM fluency, the field needs to scale toward hundreds of billions of parameters, with the training corpus physically captured rather than scraped.
Two data streams matter most. "Real machine" data is the record of a robot actually moving under teleoperation or autonomous action. "Egocentric" data is the record of a human moving while wearing sensor headsets, haptic gloves, and body cameras. Neither exists at scale on the open web. You cannot download a year of someone loading a dishwasher; you have to pay a worker in a capture facility to do it on camera, then annotate the result frame by frame so a model can learn what a stable grip on a wet mug looks like.
What that looks like in practice is a warehouse room of human teleoperators guiding a robot through hours of household and industrial chores while dozens of cameras record joint angles, contact forces, and the visual scene. Interact Analysis, cited in the livemint piece, counts 53 dedicated facilities in China built to capture this kind of human-movement data, most built in the past two years, with another 34 under construction or planned. The Hubei Humanoid Innovation Centre in Wuhan is one anchor tenant. China is now the frontrunner for real-machine data specifically because the country has been willing to build the physical capture infrastructure at a pace no Western robotics programme has matched.
The lead matters because the next stage of the humanoid race runs through the foundation models themselves. Whoever trains the next generation of humanoid foundation models on a meaningfully larger embodied corpus will set the default interface layer between humans and physical machines, the way today's leading text LLMs sit between workers and most knowledge work. China's 53-facility lead does not guarantee that outcome, but it does mean the country is no longer dependent on Western research labs to define what physical AI should look like.
That changes which firms and which geographies deserve attention. Unitree's debut pop is a capital story, not an earnings story; the share-price move separates IPO enthusiasm from the harder question of recurring revenue, which is what a humanoid.guide aggregation of Ubtech and Unitree filings tracks through Chinese disclosure. The next leg of the humanoid race will be decided by whose captured data first lets a robot pour coffee without crashing into the counter.