The humanoid robot industry has a numbers problem. According to a report circulating through Chinese industry circles, at least 80% of the "orders" you read about in headlines are not orders at all — if you use a strict definition. Behind the banner numbers is a quieter story: the people actually paying real money are universities, local governments, and — coming this year — one state-owned utility giant. That mix is the clearest signal yet that the industry has crossed from narrative into demand validation.
The word "order" has been broken
In 2025, Chinese media disclosed 48 large humanoid robot deals (¥10 million or more each), totaling over ¥5.7 billion. Impressive. But "order" in Chinese industry reporting routinely collapses seven or eight distinct stages into a single word: intent to cooperate, framework agreement, formal procurement, customer purchase order, production completion, shipment, acceptance, and revenue recognition. All of it gets called 订单 — "order."
The standard used by the HUXIU report is deliberately strict: a real commercial order means a customer buys out of production need, the robot is delivered and accepted, and it works stably without continuous engineer intervention. By that bar, at least 80% of the so-called orders are fake — and the authors say even that is the conservative reading.
Consider WAIC 2026 in July: over four days, exhibitors announced roughly ¥20.36 billion in intended humanoid-and-AI procurement. Most items disclosed neither amounts nor delivery dates. These are intention agreements, not purchases, and whether any convert into delivered systems remains to be tracked.
The actual buyers: classrooms and fiscal budgets
Data from the Humanoid Robot Scenario Application Alliance shows that in the first half of 2026, China saw 218+ publicly disclosed humanoid/embodied-AI winning bids worth over ¥1.72 billion. The buyer breakdown: education and research institutions 55.5%, government and public platforms 20.6%, industrial and technology enterprises 21.1%, others 2.8%.
July's procurement analysis by MIR echoes this: of 46 projects, universities and vocational colleges dominated, taking more than 90% of winning bid value. Unitree's own revenue mix confirms it — in the first nine months of 2025, 73.6% of humanoid revenue came from research and education, 17.4% from commercial consumers, and only 9.01% from industry. And of that industry slice, more than half was display-and-guide robots for corporate lobbies. Actual deployments in smart manufacturing or inspection: ¥15.7 million, or 2.64% of total revenue.
In other words, the money is fiscal money: universities buying teaching instruments, local governments standing up robotics training centers. Industry insiders point to three categories of inflated orders. First, local-government investment-attraction projects — procurement dressed up as industrial subsidy to pull robot companies into a region. Second, upstream-downstream "mutual protection" deals: component suppliers promise trial-plus-purchase to keep their biggest customers, and robot makers return the favor with whole-machine orders to prop up shipment metrics. Third, the overseas-sample game: an "intent order" is often just an overseas integrator asking for evaluation samples, and the promised thousand-unit mass-production contract quietly becomes a permanent "testing period." Supply-chain financing professionals who cross-checked logistics and invoicing records found that a large share of contracts generated zero engineering service fees within six months — the machines were never delivered, not even as demos. The assembly capacity behind the shipment numbers is concentrated in China's robot-manufacturing county towns — we analyzed how those clusters took shape.
The factory test: a payback ceiling, not a hype floor
The real dividing line is "entering the factory." Line testing counts as entering the factory. Data collection counts as entering the factory. Replacing an actual workstation also counts as entering the factory — and only that last one justifies real money. The hardware math explains why so few deployments clear the bar. Traditional industrial arms carry joint lifetimes beyond 10,000 hours; humanoid joints are rated in the hundreds. Humanoids run 2-4 hours per charge while factories run 7×24. UBTech's average unit price is about ¥760,000, and industry consensus puts the mass-substitution threshold below ¥150,000. Unitree founder Wang Xingxing conceded publicly that robots in factories are still less efficient than real workers, with success rates dropping sharply when the environment shifts even slightly. Even past the economics, mass production itself is a bottleneck: we broke down the delivery bottleneck — line ramp-up, yield and lead times — after the WRC trade show.
Yet one order stands apart: State Grid's April announcement of a ¥6.8 billion, 8,500-unit procurement. It is the first humanoid-scale order in China justified by payback math rather than subsidy logic. The breakdown is unusually specific — 500 live-line humanoid operation robots at ~¥5 million each (high-voltage line connection and grounding-ring installation); 5,000 quadruped inspection robots at ~¥300,000 each for substations and mountain transmission lines; and 3,000 dual-arm inspection robots at ~¥600,000 each for indoor switchgear rooms. Each device is expected to save ¥500,000-800,000 per year in labor, implying a two-to-three-year payback. The order is being released in three batches this year, and so far only a small distribution-line live-work framework (about ¥70 million) and a ¥128 million quadruped pilot have actually completed formal tender procedures. The headline makers — Unitree, Zhiyuan, UBTech, Fourier — are listed as "core suppliers" at the planning level; concrete allocations await subsequent tender announcements.
What this actually signals
Read together, the numbers describe a sector crossing a threshold. The humanoid industry still tells a 21st-century industrial-revolution story, but today it mostly sells scientific instruments and stage props. The State Grid deal is the inflection point in the other direction: a buyer specifying unit prices, deployment counts, and a two-to-three-year payback. That is how you can tell demand validation has begun — someone added a payback column. The generalist-robot narrative is heating up in parallel — whether the field is having its GPT-3 moment is a question we analyzed earlier.
For operators and investors, the practical checklist is simple: count accepted deliveries and engineering service revenue, not intent letters. Track three numbers — joint lifetime hours, unit price heading toward ¥150,000, and industrial deployments paid for by production budgets rather than fiscal budgets. The companies that bridge that gap will own the segment when the truly large industrial orders land; everyone else is still selling instruments.
