When Unitree Robotics put its H1 humanoid robot on sale in August 2023, the sticker price was around RMB 650,000. The industry average that year hovered near RMB 590,000. By 2026, the company’s G1 base model was going for RMB 85,000, and its new R1 dual-arm machine started at just RMB 26,900. That’s a drop of more than 95% in under three years. Overseas, the story is the same: average humanoid robot prices have slid from roughly $85,000 to $25,000.
Across the industry, the average selling price (ASP) is down about 70% in two years. Unitree’s average selling price has fallen 71%. Some estimates put the three-year decline at 80%.
Here’s the part that doesn’t add up: gross margins at the leading manufacturers have gone up, not down.
The G1 base model has an estimated BOM cost of about RMB 41,600. Its gross margin is still 40.7%. The EDU version clears 60%. That kind of divergence between price and margin is usually a sign of trouble — or a sign that something fundamental has changed in how these machines get built.
The cost curve, not the price tag
The price cuts aren’t a race to the bottom. They’re the visible edge of a cost structure that has been collapsing from the inside.
Start with components. Reducers, servo systems, and controllers account for more than 70% of a humanoid’s total cost. As of May 2026, the localization rate for core components reached 75% to 90%. Chinese-made components run 30% to 50% cheaper than Japanese or German equivalents. Take RV reducers: domestic production went from 18.7% of the market to 72%. The price for an equivalent specification fell from RMB 1,280 to RMB 410.
Then there’s China’s new-energy vehicle (NEV) supply chain. Motors, reducers, and sensors used in humanoid robots overlap heavily with those in NEVs. Barclays has described humanoid robots as “miniature cars” — the bearings, gears, and actuators inside a robot joint are close cousins of what you’d find in a car. China’s NEV industry spent a decade driving down the cost of those parts. Humanoid makers are collecting the dividends.
Scale is doing its part. Chinese humanoid production went from about 20,000 units in 2025 to more than 40,000 in the first half of 2026. The full-year estimate is north of 100,000.
And geography matters enormously. An Optimus built on the Chinese supply chain carries a BOM of about $46,000. Build it entirely outside China and that number jumps to $131,000 — nearly three times as much.
So when prices fall 70% and margins rise, it’s not a paradox. It’s what happens when the cost curve drops faster than the price tag. Some analysts have read the combination as a sign that volume production is beginning to ramp up.
The profit has moved upstream
The price war has redrawn the industry’s profit map. The phrase you hear in Chinese industry circles is “integrators bleed, component makers eat.”
Most integrators are bleeding. Of the more than 140 humanoid robot makers in China, fewer than five have reached profitable scale. UBTECH booked RMB 821 million in humanoid revenue in 2025 and a net loss of RMB 790 million for the year.
Unitree is the exception — the first Chinese legged-robot company to turn a profit at scale. But its margins don’t come from assembling robots. They come from making more than 90% of its core components in-house. Pure assembly work carries gross margins below 18%. Of the RMB 4.2 billion Unitree raised in its IPO, RMB 2 billion is going to component production lines, not robot assembly. And even Unitree isn’t immune: in 2026, its revenue grew 68% while net profit was cut in half. Its profitability still leans heavily on research labs buying machines, not factories.
Component makers, meanwhile, are enjoying a very different year. Harmonic reducers fetch 40% to 55% gross margins. Planetary roller screws run 50% to 70%. Six-axis force/torque sensors command similarly high margins. Across the board, component suppliers are holding steady at 40% to 65% gross margins — almost untouched by the price war downstream.
Leaderdrive, a Chinese harmonic reducer manufacturer, grew revenue 47% in 2025 and has production booked through 2027. Investors have taken notice. About 70% of humanoid robot financing in 2026 went to component companies, not integrators.
One caution: some observers see those fat component margins as a signal in themselves. High margins attract capacity. Long-term, there’s still 30% to 40% of cost reduction left in the component stack. But the upside is likely to concentrate among the top players.
The demand side is where the math gets hard
Cheaper hardware gets you past the entry gate. Whether the business model actually works depends on whether the economics of replacing human labor hold up.
The optimistic case: a humanoid robot can work 20 hours a day. Factor in depreciation and maintenance, and the effective cost runs about $3 to $5 per hour. Human labor in a tier-one Chinese city costs $8 to $12 per hour. By that math, the crossover point has already arrived. Forty percent of surveyed companies say they’d adopt at scale at a price between RMB 100,000 and RMB 200,000. The industry consensus is that full-size models need to fall below RMB 50,000 for broad labor replacement to make sense.
The cautious case: production, delivery, paid deployment, and repeat orders are four progressively narrower filters. In the first half of 2026, China produced 40,000 humanoid robots and delivered 23,000. Of those 23,000, fewer than 5% were deployed in factories. Global shipments rose 272% year over year to 19,100 units, but third-party tests of general-purpose tasks still show low success rates on physical robots. Cross-task generalization is weak. The “general brain” — a single system capable of handling a wide range of tasks — hasn’t arrived.
The industry is calling 2026 the year the conversation shifts from viral demos to billable robot hours. The real inflection point isn’t whether a robot can do a backflip. It’s whether it can hold down a night shift, do dull work reliably, deliver repeatable results, and have clear accountability when something goes wrong. Overseas analysts say the same thing: China’s prices have come down dramatically, but commercialization still hinges on operational stability.
Which is why, despite an 80% drop in ASP over three years, industry-wide ROI is still negative. When costs will actually fall below human labor across the board remains an open question.
The model didn’t break. It moved.
Here’s the clearest way to read what’s happening: the old business model — sell a robot, make a margin on the hardware — is basically dead. Price wars plus a three-to-six-month technology iteration cycle have crushed pure integrators to 10% to 20% gross margins and persistent losses. Over the next three to five years, nine out of ten small and mid-sized integrators may disappear.
The “sell components” model is working. Upstream core parts account for 45% to 70% of a robot’s BOM cost. Orders come in ahead of production, customers are diversified, and the parts are standardized and reusable. It’s the most certain profit pool in this production cycle.
The “sell use cases and reliability” model is just beginning. The leading manufacturers are moving from trade-show prototypes to actual deployment. UBTECH’s Walker S passed a three-month validation at a NIO factory and gained qualification for production deployment. Unitree’s G1 is now deployed at a BYD factory. State Grid plans to buy around 8,500 specialized robots this year. Goldman Sachs expects global humanoid shipments to hit 76,000 units by 2027 and 502,000 by 2032. Omdia’s long-range forecast is 2.6 million by 2035.
Wang Xingxing, Unitree’s founder, has said the real inflection point for general-purpose robots entering homes is still a decade away. The biggest bottleneck, he argues, is the generalization capability of embodied AI.
So the 70% price collapse didn’t kill the humanoid robot business. It killed the pure-assembly business model. The model that replaces it looks different depending on your time horizon: short-term, it’s a supply chain play. Medium-term, it’s about delivery and repeat orders. Long-term, it’s about whether anyone can build that general brain.
