In late August we covered Unitree turning over 85% of its float on listing day and running a full round trip in five sessions. The conclusion there was that the week moved the shareholder register and nothing else. The price went out and came back. Where the robots go to work did not change by a word.
On September 9 the story stopped being the market talking to itself. The regulator spoke.
The Information reported, and Reuters followed, that the China Securities Regulatory Commission has informally told some banks and companies that humanoid robotics listing candidates should demonstrate recurring revenue, progress toward narrowing losses, or significant technological innovation. The Wall Street Journal reported a similar signal separately. Reuters could not independently verify the report, and Chinese regulators have not commented.
Three conditions, and only one of them is hard
They are not equivalent. Narrowing losses is financial discipline and technological innovation is a narrative; both can be assembled for a prospectus. Recurring revenue is a physical test.
It does not ask whether the robot sold. It asks whether the customer keeps paying after delivery: rent, billed hours, a service contract, a software subscription, any of them, on the condition that the machine stays on site doing something worth paying for.
That is precisely the column the humanoid industry does not have yet.
Open Unitree’s books
Unitree is the healthiest set of financials in the cohort, which makes it the most useful sample rather than the least.
Per The Robot Report, Unitree generated 1.70B yuan in 2025 revenue, up from 392.77M yuan in 2024. Humanoid robots contributed 868M yuan, or 51.78% of the total, on shipments of more than 5,500 units. For the first half of 2026 the company guided to 1.052B–1.128B yuan, growth of roughly 36% to 45%.
This is a company that sells things and earns money. It is a different species from most of its peers.
The two ratios the WSJ pulled out are the actual news of the day:
- Under 10% of 2025 revenue came from industrial applications
- More than 40% came from overseas, leaving roughly 60% domestic
Add one more from the Financial Times: China has established more than 90 humanoid training centres, many co-funded by local governments and robot makers. These centres buy humanoids and use them to generate training data, often through teleoperation.
Put the three together and the composition of that 1.70B yuan resolves: research and education, culture and entertainment, commercial display, overseas buyers, and a set of state-co-funded training centres purchasing robots to manufacture data. Industrial production lines are the sub-10% slice.
Right now humanoids create jobs. They do not yet take them.
This is the finding we keep hitting and the trade press keeps skipping: at this stage humanoid robots are net job-positive.
The training centres are the proof. A robot arrives and a teleoperation station arrives with it. Someone in a headset with hand controllers demonstrates the motion, over and over. That role did not exist in 2023. It exists now, and it multiplies as the centre count climbs. Data labeling, scene construction, mechanical maintenance, deployment engineering: same chain, same direction.
In late August we counted $230M of Chinese state procurement of humanoids in six months. That money lands where this picture predicts: training and demonstration, not takt-timed production.
The U.S. sample says the same thing from the other side. Earlier this month we covered Agility’s S-4, which put 2025 net sales at $1.78M against a $140.2M operating loss, with Digit deployed at 9 customer sites, more than 65,000 accumulated operating hours, and over $300M in multi-year contracted orders. Nine sites is the order of magnitude for one of the most serious warehouse-humanoid companies on the planet.
Both readings agree. The robots got built, the capacity got financed, the valuations got printed. The column marked people who no longer come to work because of this is still close to empty.
Why this regulatory line is worth watching
At Wednesday’s close Unitree traded at 513.93 yuan, worth roughly $30B, or about 125 times 2025 revenue and more than 350 times adjusted 2025 earnings. That is down about 39% from the debut close and about 53% below the first-day high, and still more than three times the offer price.
The market is taking back multiple, not conviction. What changes conviction is the first of the CSRC’s three conditions.
Requiring recurring revenue hands everyone a shared indicator. The day a humanoid maker can publish a real recurring-revenue table — monthly rent per unit, billed operating hours, service renewal rates — that company’s robots are holding down a post at a customer for the long run. Holding the post is what displacement is. Until then, every “tens of thousands of units into factories” projection measures capacity, not jobs.
For people working in or near this industry, the line has three practical uses:
- Next 12 months: robotics is still adding jobs on net. Teleoperation, data collection, deployment and field maintenance are the openings with real volume behind them. These roles are transitional by construction. Their existence is evidence the models are not good enough yet.
- The signal to watch: vendor revenue mix. When one of them moves industrial-application share from single digits to double digits and can say whether it is rent or hours, that is when the displacement clock starts.
- The metrics to ignore: shipments, capacity plans, letters of intent, market cap. All four have gone up by more than an order of magnitude in a year while the number of posts held has not moved.
Sources
- China curbs humanoid IPOs after Unitree’s volatile debut, The Information reports (Reuters, September 9, 2026)
- Chinese Regulators Seek Higher Bar for Humanoid Robotics Listings (The Wall Street Journal)
- Unitree shares down 53% from IPO debut (The Robot Report, September 9, 2026)
- China’s humanoid robot training centres (Financial Times)
