Xiaomi posted Q2 results on the evening of August 18. Revenue of RMB 108.9B, adjusted net profit of RMB 6.2B, R&D spending of RMB 9.2B, up 18.9% year over year. That is what the market talked about that night.
The paragraph worth stopping on sits further down the release, under robotics. Xiaomi wrote that its next-generation humanoid has widened its footprint inside the company’s own car plant: dual-side operation at the self-tapping nut loading station now runs at 98.0%, and two harder scenarios were added, center console side-panel sequencing and returnable crate folding, both at 90%.
A listed company put humanoid workstation names and per-station success rates into a quarterly disclosure it is legally accountable for. That is the part to file away, not the 98.0.
Everyone else reports the aggregate
For a year now, the humanoid numbers reaching the outside world have all been aggregates.
In July we covered Chinese humanoids reaching 85% of human output on logistics floors. That is a cross-site average. On August 10, AgiBot passed Unitree and more than 70% of first-half shipments went to industrial and commercial buyers. That is a shipment mix. On August 11, three automakers each showed a humanoid, and all three were doing the same job: pulling parts out of bins and loading a trolley. Reporters worked out the task by watching. No manufacturer named it.
Xiaomi named it. Three of them. The station has a name, the task has a name, and the rate is reported one station at a time. Self-tapping nut loading is one motion. Console side-panel sequencing is another. Crate folding is a third. 98.0%, 90%, 90%.
An aggregate can be lifted by choosing which sites to count. A station-level rate cannot. The station has a fixed position on the line, and if the cycle does not finish, the next cycle waits.
The gap between 98.0 and 90 is not eight points
Read those three numbers as headcount and they say completely different things.
98.0% means two interventions per 100 nut-loading cycles. 90% means one per 10. The first is a ratio where one worker can cover several stations. The second is a ratio where somebody has to stand there.
So the staffing shape on that line today is a robot working and a person watching, not a robot replacing a person. The threshold where headcount actually moves is not 90 and not 98. It is somewhere in the 99s, where a station fails less than once a shift and the person on it has a reason to go do something else.
Xiaomi did not blur this. On the same earnings call, president Lu Weibing put the robot’s position plainly: humanoids are a long way from large-scale commercialization, and Xiaomi has to land its existing businesses first, smart factories and smart homes. He also said the robot will not ship as a standalone product. It plugs into Xiaomi’s own “human, car and home” stack.
That second sentence is the load-bearing one. A robot that is not a product does not need a buyer, and does not need a sales cycle to clear.
Builds it, runs it, no customer in between
Xiaomi’s structure here is not the industry’s structure.
Almost every humanoid company has to sell a robot to a factory, wait for the factory to choose a station, wait for validation, then wait for an expansion decision. Procurement, validation and rollout each move in quarters. Xiaomi skips all three. It builds the robot, owns the plant, assigns the station, and can change any of it the same day.
Hyundai and Boston Dynamics run the same structure. LG is moving toward it: on August 13, LG signed a humanoid agreement with NVIDIA, but the machine with an actual date on it was CLOiD, a wheeled unit going onto a Tennessee washing-machine line before year end. When a company with its own factories builds robots, the first customer is always itself.
The structure sets the iteration speed and it sets the disclosure cadence. Xiaomi’s auto business booked RMB 23.9B this quarter on 104,199 deliveries, up 28.2%. This is a line running volume, not a demonstration cell built for visitors. Every nut the robot loads there is loading against takt time.
The World Robot Conference opened in Beijing E-Town on August 19, and Xiaomi’s 1.7-meter humanoid makes its first public appearance there. The show has more than 300 exhibitors this year, up 36%, over 2,000 exhibits, more than 150 global debuts. There will be a great deal of impressive movement on the stands. The information already shipped the night before, in a filing.
Which jobs these three stations are
Self-tapping nut loading, console side-panel sequencing, returnable crate folding. Inside an automotive plant those map onto two job families: part-loading operators on the final assembly line, and in-plant logistics handlers who cycle empty crates.
Both families are repetitive, takt-bound, and low on judgment. They are also the exact layer conventional automation never took. A fixed industrial arm needs the part presented in a known position and a known orientation; parts in a bin are loose, side panels face a different way every time, and an empty crate folds with a different feel on every cycle. The bet on humanoids lives right there — walk up to the station a person used to stand at, and change nothing about the line, the tooling or the crate.
That makes the signals to watch specific, and there are two.
The first is whether the station count keeps climbing. Xiaomi went from one station to three this quarter on the same model and the same body. Six or eight next quarter would mean transfer cost has dropped far enough to add stations weekly instead of quarterly.
The second is whether 98.0 becomes 99-point-something. When that number crosses, the first real headcount adjustment happens, and it almost certainly will not arrive as a layoff. The usual form is that the station stops being backfilled: someone leaves, and the seat stays empty.
In the quarter the number appeared in a filing, nobody left the plant because of it. But the disclosure moved from “we are working on robots” to “this robot hit this rate at this station.” Once the language moves that way, next quarter cannot move back.
