The first people to spend a working shift near an IRON will be XPeng store staff and campus employees, because that is where the company says its initial commercial-scenario rollouts begin [8]. Selling to outside buyers, in China or overseas, is a 2027 plan [9]. He Xiaopeng says the aim is a robot with full generalisation that can become a companion in people's lives [14]; the first posting is a shop floor the vendor owns.
That ordering is more useful than the walk-off itself. A line whose core processes are more than 80 percent automated [3] is a claim about build repeatability, not about what the machine does across an eight-hour shift. What the walk-off does establish is narrower and worth having: the unit that powered up and balanced came off a fixture sequence rather than out of the hands of the team that designed it [2]. That is a real difference from a stage prototype. It is also still one unit.
What an operator needs before hosting a fleet is hours of unattended task time per unit per week and a failure rate on the specific task. What the announcement supplies is 76 degrees of freedom across the body with 21 in each hand [4], plus three Turing chips rated at up to 2,250 TOPS in total [5], which works out to roughly 750 TOPS per chip [6]. No unit volume, cycle time, price or uptime figure appears in it [16]. XPeng runs its foundation model on the robot instead of teleoperating it [5], which the company says cuts latency and keeps operational data on the machine [7].
Set that against Tesla, where Optimus production is now expected in the coming months and commercial sales are targeted for the second half of 2027, where earlier plans had production this summer and first sales by late 2026 [10][11]. The sales date has moved by about a year [12]. Line the two calendars up and XPeng's lead shrinks to sequencing inside a single year, because its external delivery and Tesla's commercial sales both sit in 2027 [13]. On what each company has published, no outside buyer takes delivery of either machine before then [18].
So the test for anyone offered a humanoid pilot this year runs on two axes, and the first is who absorbs the cost when the robot fails the task. In a vendor-owned store, the vendor does, which is the sensible place to start, but it is thin evidence of anything beyond that. The second axis is whether the task is scored on completion or on presence. A humanoid greeting visitors in a showroom is scored on presence rather than on completing a task. Anything sitting in the vendor-pays, presence-scored quadrant is a demo with a production line behind it, which beats a hand-built prototype and falls short of a product.
The figures that move a programme out of that quadrant are units shipped to a party that can send them back, and hours of scored task time per unit. XPeng has published neither [16].
