Unitree filed its listing notice on August 18 and lists on Shanghai’s STAR Market under 688836 the next morning. Post-offer valuation: about 61B yuan, roughly $8.5B. Issue multiple: 219.23x. The retail allocation ratio came in under 0.02%, meaning fewer than two in every ten thousand orders were filled.
We covered the pricing side on August 6, the week the FCC barred new Chinese humanoids from the US market. What is new in this week’s filing is a production figure Unitree had not published before: as of July 2026, roughly 18,000 bipedal humanoid robots have come off its line. The company attached a qualifier, counting only bipedal humanoids, excluding wheeled-base humanoids and other quasi-humanoid form factors.
That qualifier tells you Unitree understands what the number will be used for. Of everything disclosed this week, 18,000 is the only figure that converts directly into a labor question.
The ten-year ledger
Revenue ran 159M yuan in 2023, 392M in 2024, and 1.699B in 2025. Net income crossed over in the same window, from a 11.1M yuan loss to 94.5M and then 278M. Stripping out non-recurring items, the 2025 line reads 591M yuan. Gross margin on the core business climbed from 44.18% in 2022 to 60.13% in 2025.
Against that 591M, a 61B yuan offer valuation is a static multiple a little north of 100x. The 219.23x figure uses the diluted basis. The gap between the two is the market’s forecast for this year’s earnings.
The mix shift matters more than the totals. Humanoids brought in roughly 877M yuan in 2025, passing quadrupeds for the first time and taking more than half of core revenue on shipments of about 5,500 units. Quadrupeds did not stall either; the 2023-2025 stretch put more than 30,000 of them into the field.
So the 18,000 cumulative bipeds break down roughly as follows: last year supplied about a third, the rest came from the post-G1 ramp in 2024 and the first half of 2026. A company registered in 2016 with 100,000 yuan of capital now ships bipedal robots at five-figure annual volume. That is a manufacturing result, and it is real.
The growth deceleration inside the same filing is a story we already took apart when the prospectus landed on August 1. No need to relitigate it here.
What 219x is actually buying
Nobody pays 219 times earnings for a manufacturer that booked under 1.7B yuan of revenue. The market is not pricing Unitree as a manufacturer.
Guoxin Securities put the assumption in plain language in its latest note: the multiple only works if Unitree grows far faster than its sector for several consecutive years, and the price is not buying 2026 results but discounting a trillion-yuan future market back to today. CCB International is more conservative, running 32x a 2026 price-to-sales target for an implied 109B yuan. Some sell-side voices are talking about 200B.
Those three numbers are not three valuations. They are three different industrial assumptions. 61B buys a company that already makes money selling robots. 109B requires Unitree to hold high growth for several more years and prove humanoids keep contributing profit. 200B requires the market to believe Unitree becomes the platform the rest of the industry has to route through, which is a claim about lock-in, not unit sales.
The strategic placement list points the same direction. When it was published on August 6, DeepSeek, Tencent entities, PetroChina’s Kunlun Capital, China Southern Power Grid’s finance arm and China Telecom’s Tianyi Capital each took 136M to 141M yuan. A power grid and a telco are not buying robots. They are buying an option on robots entering their own maintenance and inspection work.
Read in that light, disclosing 18,000 units the night before the bell is not incidental. When a valuation rests on discounted future deployment, the most persuasive evidence available is not the income statement. It is physical units already delivered.
Where do 18,000 robots go to work
Here is where the number stops flattering and starts asking.
The prospectus lists the deployment contexts for Unitree’s humanoids: scientific research and education, culture and entertainment, commercial display, and some industrial applications. The first three carry the bulk. A university lab buys a G1 as a research platform. An entertainment venue or a brand activation buys one as a performer or a storefront. All three generate revenue. None of them generate substitution. The G1 doing martial arts on the Spring Festival Gala stage displaced no one.
Gao Feng, a professor at Shanghai Jiao Tong University’s School of Mechanical Engineering, drew the line precisely in comments to The Paper. Legged robots, he said, are still crossing from “can walk, can jump” to “can actually be used.” Running, jumping and backflips demonstrate mechanical systems and motion control. Real machine intelligence shows up as perceiving, judging and completing a task autonomously in an unstructured environment. If the robot cannot finish the job without teleoperation or continuous human intervention, it is not there yet.
Rewrite that as an employment test and it gets blunt. A humanoid that needs a person standing beside it with a controller has not replaced labor. It has created a job: operator.
The cohort comparison is already on the record. On August 10 we wrote about AgiBot passing Unitree on cumulative output, and the point of that piece was not the ranking. It was the mix: roughly 70% of AgiBot’s deliveries go into industrial settings. Two Chinese makers at the same order of magnitude, one shipping mostly into factories, one shipping mostly into labs and exhibition halls. The multiple assigned to Unitree this week is the multiple you assign to the first kind of company.
Look one layer further down the line and the ceiling is visible. On August 11 we catalogued the humanoids BMW, GM and a third automaker have in test. The task they are all actually doing is the same one: kitting parts. That is the full width of industrial work humanoids can currently deliver reliably, across every vendor.
So 18,000 splits in two. As a manufacturing record it holds: ten years to take a bipedal robot from lab equipment to a mass-produced product, with the motor, joint, actuator and motion-control chain proven out. As a labor signal it has not arrived. How many of those 18,000 units are doing work someone was previously paid to do is not disclosed, and the ordering of the deployment contexts in the prospectus suggests the share is small.
For specific occupations that implies two clocks running at different speeds. The first is already running: robot assembly, motion-control engineering, component supply chain, teleoperation and on-site commissioning are jobs the 18,000 figure created, and Unitree’s own sharply rising R&D expense is the evidence. The second has not started: line operators, warehouse pickers, and repetitive service roles in hotels and retail are the categories the valuation model is discounting, and reaching them requires exactly what Gao described, autonomous task completion in unstructured environments, not athletic ability.
What separates the first clock from the second is not production capacity. Unitree has cleared capacity. The remaining proof is whether the robots it has already built can finish a shift without a human holding the controller.
61B yuan is the starting line and 18,000 units are the chips. The exam has one question on it: how many of those 18,000 show up next year somewhere that keeps a timesheet.
Sources
