AINEWS
Hardware is largely solved. The bottleneck is real world training data, and the race to capture it is just beginning.
On August 17, a black-and-white humanoid named Superman sprinted 12.66 meters per second across a Beijing arena, briefly faster than Usain Bolt's recorded peak, then drove itself headfirst into a barrier. The clip has circulated since as a one-second summary of where Chinese robotics sits in August 2026: the hardware is startling, the usefulness is not.
Unitree, the Shanghai-based maker of that robot, debuted on the city's stock exchange a few days later. Its shares leapt roughly 460% on the first day of trading. The same week, The Economist ran a column arguing that Chinese firms will struggle to make money from humanoids even after winning the hardware race. The two stories sit on top of each other for a reason. The bottleneck has moved.
Chinese makers are on track to ship about 50,000 humanoid units in 2026, more than triple last year's figure, according to Morgan Stanley research reported by CNBC. Actuators, reducers, joints, the supply chain that turned Shenzhen into the world's hardware bench, are largely mastered. So the next constraint is software: the foundation models that turn a walking, two-armed machine into something that can pick up a glass, judge its weight, and pour without breaking it. Today's humanoid foundation models run at a few billion parameters. Replicating human-body versatility will take hundreds of billions, the scale of a frontier large language model. The moat is no longer the chassis. It is the data that trains the brain.
That is where the hardware winners run into a problem they cannot outsource. Unlike text-based AI, robot foundation models cannot learn from scraped web pages. Training data has to come from physical interaction, what researchers call "real machine" data, captured by tele-operating a robot through tasks or by letting a fleet run autonomously and recording the failures. Or it comes from "egocentric" data: humans wearing haptic gloves and head-mounted cameras, performing the task themselves so the model can learn from first-person video and grip force. Either approach is slow, expensive, and largely bespoke. Insurance Journal reported in July that the industry is sending robots out into the world specifically to learn how to be human in it.
China is positioning to win this next phase. The same supply-chain density that lets a Unitree humanoid ship in volume also lets a single operator run hundreds of machines through a warehouse, a restaurant kitchen, or a hotel corridor and collect the resulting trajectories at a cost no Western rival can match. Egocentric capture, the haptic-glove pipeline, is cheaper inside a country with mature glove, camera, and motion-capture manufacturing. The race is moving from who can build the robot to who can record what the robot needs to learn.
The United States has the model side covered. The data side is harder. Tesla's Optimus program is the obvious counterweight, and CNBC has framed the Unitree IPO as the opening of a US-China humanoid contest. American data-collection pipelines are thinner, and US privacy law plus labor cost make a thousand-robot tele-operation fleet difficult to stand up. The next two years will look less like a sprint and more like a procurement problem: who can put gloves and cameras on enough humans, and robots in enough rooms, to feed the next generation of models.
Watch the data partnerships, not the share price. Unitree's 460% debut is the market pricing a hardware story it already knows. The story that pays is the one nobody has filed yet: a multi-year real-machine and egocentric capture deal between a humanoid maker and a Chinese logistics, hospitality, or elder-care operator, with a unit price attached. That filing, when it comes, will be the first real evidence of who owns the moat the column says is still up for grabs.