A humanoid in a shop is an installation.
XPeng Robotics raised over 900 million dollars on 24 August and said the first IRON humanoids go into its own stores and campuses. That is not a hiring decision, it is a site fit-out, and the questions that decide whether it works are the ones we ask on every install.
On 24 August 2026, XPeng announced that its robotics business had raised over US$900 million at a post-money valuation of over US$6.3 billion, led by IDG Capital with Tencent and Alibaba in as strategic investors. It is the largest single private round in China's embodied AI sector so far. The robot it funds, IRON, carries 76 degrees of freedom in the body and 21 in each hand, runs on three in-house Turing chips rated at a combined 2,250 TOPS, and is meant to enter mass production by the end of this year (XPENG, 2026).
The line that actually tells you something is further down the release. Before any customer takes delivery in 2027, IRON goes into XPeng's own stores and campuses. The first commercial humanoid deployment from a company with real manufacturing behind it is a retail floor.
A shop is a venue, and a venue has a brief.
It is tempting to read that as a soft launch, somewhere gentle to fail. It is more interesting than that. A car showroom is a controlled site with a known floor plan, known lighting, known dwell times and staff who can be trained, which makes it the easiest place on earth to run a machine that has to work in front of the public. It is also, whether XPeng frames it this way or not, an installation.
We build in those rooms. A humanoid on a showroom floor raises the same list a kinetic sculpture or an interactive fixture raises, and none of the items are about intelligence.
- 01Footprint and exclusion zone. What is the machine allowed to occupy, and what does the public get to stand in.
- 02Power and charging. Where the robot goes to sleep, who plugs it in, and what happens on a Sunday when the person who knows is off.
- 03Uptime that a shop manager can live with. Not the demo, the eleventh hour of the ninth day.
- 04The stop. One physical control, reachable, obvious, and tested by someone who was not there when it was fitted.
- 05The failure state. What it looks like when it stops. Frozen mid-gesture is a bad look, parked and quiet is fine.
- 06Who owns it on site. Every install we have ever done has one named person, or it degrades within a month.
That list is not a criticism of humanoids. It is the same list we would write for a projection rig or an automated door, and it is the part that decides whether the thing is still running at Christmas.
Compare the version that already works.
Agility Robotics passed 100,000 totes at a GXO fulfilment site in Flowery Branch in November 2025. Digit picks totes on and off an autonomous mobile robot, puts them on a conveyor, stacks them at another floor location (Agility Robotics, 2025). One verb, one site, one measurable count, on a service contract. It is deliberately unglamorous and it is the only humanoid deployment most people can name a number for.
Scale is worth keeping in view. Global humanoid shipments in 2025 came to 13,318 units on Omdia's count, up nearly 480 per cent, with Chinese manufacturers taking around 87 per cent and AgiBot leading on 5,168 units (Omdia, via SCMP, 2026). A 480 per cent rise off a base that small is a pilot industry growing fast, not a labour market changing. Anyone selling you a workforce story this year is a few years early.
What we would tell a client asking about this now.
Two things, neither of them exciting. First, if a humanoid is on your roadmap for a store, a lobby or a stand, brief it as a fixture with a job description, not as a person. Name the single task it performs, the hours it performs it, the state it holds when it is not performing it, and the human who resets it. A brief that says "greet customers" will fail. A brief that says "escort a visitor from the door to the demo bay, then return to the dock" can be built, measured and argued about.
Second, buy the boring layer first. The value in these rollouts sits in the supervision, the telemetry and the recovery path, and that layer is worth building whether the machine on the floor has legs, wheels or nothing at all. We have put that layer under screens, doors, sensors and lighting rigs for years. A humanoid changes the actuator, not the architecture.
IRON is a serious piece of hardware backed by serious money, and putting it in a shop before selling it is the right call. Just read the move for what it is. Somebody at XPeng is about to learn a lot about floor surfaces, charging cupboards and who has the key, which is what everyone building physical things learns eventually, usually a fortnight after opening.