A robot waits at the end of a line for a tablet to arrive on a conveyor. It lifts the tablet, loads it into a test rig, waits, unloads it, and sorts it by result. Then again, and again, for six days, while cameras stream everything.
That test, run in June at a Nanchang factory, is still the most detailed public measurement of a Chinese humanoid doing paid work.
Key takeaways
- AgiBot (智元机器人) says it shipped 9,700 robots in H1 2026 on revenue near 2 billion yuan (≈ US$282M), passing 20,000 units in September.
- On 29 October it proposed Worken — active robots × autonomous hours × output efficiency — to measure work done, not machines sold.
- Its best-documented deployment, at Longcheer's Nanchang plant, ran eight robots for six days: 17,625 tablets, about 64,800 operations, 99.99% success.
- Economics turn on utilisation: a 300,000-yuan (≈ US$42,000) machine costs 30 yuan an hour at 2,000 hours a year, about 62 at 960.
- Unitree, listed in Shanghai in August, grew nine-month revenue 56% while third-quarter profit fell 17% as research spending rose 250%.
A new unit, announced on a stage
On 29 October in Shanghai, AgiBot president and CTO Peng Zhihui (彭志辉) introduced a word: Worken, from "work" and "token". It multiplies active robots by autonomous working hours by an output efficiency ratio, measured against a skilled worker on the same task. "Ten thousand units off the line is a victory for manufacturing," he said. "Ten thousand units on the job is a victory for physical AI."
Worken is a company invention, not a standard. It addresses a problem the sector created: firms valued on demos, where a robot that walks on stage is impressive and commercially worthless.
It also concedes something. If you need a new unit to describe how much work a robot does, the old unit was measuring the wrong thing.
The one line with public numbers
Eight AgiBot G2 robots joined a tablet inspection line at Longcheer's Nanchang plant. Over six days they ran 64 hours of live-streamed operation, handled 17,625 tablets and completed roughly 64,800 operations, with a reported success rate of 99.99%.
The task is modest: pick a tablet off a conveyor, load it, wait, unload, sort by result. A fixed arm could do it. What the humanoid adds is swapping between SKUs without rebuilding the line.
AgiBot also disclosed failures: communication faults between robots and test equipment, and robots colliding. Both were fixed. But the gap between 99.99% of operations and 100% of shifts is where factory managers live. One stopped line costs more than a month of robot wages.
What an hour actually costs
The arithmetic that decides this follows; the numbers are illustrative and exclude subsidies.
Take a humanoid at 300,000 yuan (≈ US$42,000). Depreciation over five years with no residual value is 60,000 yuan a year. Run it 2,000 hours — eight hours a day, 250 days — and depreciation alone is 30 yuan an hour.
Now run it 960 hours, which is what ten working days a month looks like. Depreciation doubles to about 62 yuan an hour. Add maintenance, power, integration and supervision — at Chimelong the robot-to-staff ratio started near 1:1, is now about 3:1, and is targeted at 10:1 — and the hourly cost roughly triples.
A humanoid's cost per hour is a function of how often it is switched on, and almost nobody publishes that. A report this year from CAICT, the telecoms research institute, found that many deployed humanoids run fewer than ten days a month — a figure that fits the shipment data: Counterpoint counted about 22,000 units shipped globally in the first half of 2026, more than 60% to stages, showrooms and laboratories.
The robots exist. The shifts do not.
Growth without profit
The listed numbers say the same from the other side. Unitree (宇树科技), which listed in August, reported nine-month revenue of 1.82 billion yuan (≈ US$256M), up 55.7%, and net profit of 387 million yuan (≈ US$54M), up 270%. Then look at the third quarter alone: revenue up 70% to 665 million yuan, profit down 17% to 113 million, research spending up 250% to 127 million, or 19.1% of revenue.
That is a hardware company racing to stay relevant in software: revenue scales with units, research faster, profit later, if at all.
Where the deployments work
The most convincing deployments are not factories. They are places where a robot's job is to be seen: Chimelong theme park in Hengqin, where AgiBot deployed 300-plus robots within a month and now runs close to 500.
That is a real business — tickets, shows, crowds — but a different one from manufacturing. In a park, a robot that works 70% of the time is an attraction. On a production line it is a liability.
According to Li An, chief scientist at BrainNet (脑机网), the shift from counting units to counting hours is the first time the sector has priced its own weakness: the question is no longer whether a robot can do a task once, but whether a customer will pay for the second thousand repetitions.
Honest limitations
Everything about AgiBot here comes from its own announcements and briefings at its 29 October event. It is private; none of its shipment, revenue or deployment figures have been audited.
The Longcheer test was a demonstration, not a contract. It ran six days, not six months, and 99.99% refers to individual operations — an easier bar than line availability.
The cost-per-hour arithmetic is mine, built on a hypothetical price. Real quotes vary, and local subsidies change payback entirely.
The "fewer than ten days" figure comes from a CAICT report whose sample is not public, so treat it as an indicator.
What readers can do now
- If a vendor quotes a price per unit, ask for cost per operating hour at your duty cycle. The same machine is cheap or unaffordable depending on how many shifts you can fill.
- If you are piloting a humanoid, measure line availability, not task success. A robot that completes 99.99% of motions but stops the line twice a week has negative value.
- If you are underwriting the sector, discount unit counts and ask for hours.