A humanoid robot walked off an assembly line in Guangzhou on Tuesday under its own power, and in doing so moved China's most-watched embodied AI project out of the demo-stage era and into the far less forgiving world of manufacturing.
XPeng, the Chinese electric-vehicle maker better known for its G-series SUVs than its silicon, said on September 8 that it had commissioned what it calls the world's first automated production line for advanced humanoid robots. Its IRON robot completed final assembly and then walked off the line on its own legs — a piece of choreography that doubles as the company's thesis statement. The line, XPeng says, runs with more than 80% of its core processes automated. It sits inside a full-chain facility of roughly 110,000 square meters in Guangzhou's Tianhe district.
The specifications are the part that should make competitors uncomfortable. IRON stands 178 centimeters tall, weighs about 70 kilograms, and carries 76 degrees of freedom across its body plus 21 more in each hand. Its skin is a fully enclosed flexible lattice structure that XPeng treats as both cosmetic surface and safety system. And its brain is three of XPeng's own Turing AI chips, which the company says deliver up to 2,250 TOPS of effective compute — enough, it claims, to run its Physical AI foundation model entirely on the robot, with no teleoperation and no round trip to a data center.
A line built with no blueprint
He Xiaopeng, XPeng's chairman and CEO, framed the commissioning as a manufacturing milestone rather than a product launch.
The robot production lines were created from scratch with no precedent to follow, He said at the ceremony. Today's step is small, but XPENG is building the production lines for an entirely new product category.
He then clipped a staff badge onto IRON — theater, but pointed theater. XPeng's first robots are not going to customers. They are going to the company's own stores and campuses, where they will be watched, measured and, presumably, embarrassed in public before anyone is asked to pay for one.
The timeline XPeng reaffirmed is aggressive but bounded. Mass production begins by the end of 2026. An official launch, with sales and deliveries in China and overseas markets, follows in 2027. The company declined to disclose the line's specific capacity or IRON's price, though it has previously signalled an ambition to build more than 1,000 robots a month on the way to a million units a year by 2030. XPeng also expects gross profit per robot to run meaningfully higher than on its NEVs, citing high technical barriers and thin supply of genuinely capable humanoids — a claim it has notably declined to put a number on.
The money behind the push arrived three weeks earlier. On August 24, XPeng disclosed that its robotics unit, Dogotix, had signed share purchase agreements worth roughly $900 million in financing commitments, led by IDG Capital with Gaorong Ventures participating and both Alibaba and Tencent joining as strategic investors. The implied post-money valuation is about $6.3 billion if the equity incentive pool is fully used, and it stands as the largest single private round in China's embodied AI sector to date. XPeng retains control and will keep consolidating the unit's results.
One more engineering detail is worth flagging: reporting earlier this month indicated IRON has shifted to high-energy solid-liquid hybrid, or semi-solid, batteries — a quiet concession that all-solid-state cells are not arriving on the schedule the industry once promised.
Why in-house silicon is the real story
Strip away the walking and the badge, and the competitive claim here is about chips.
The humanoid field has split into two economic models. Chinese companies now control roughly 90% of unit volume; Unitree alone sold about 5,500 humanoids last year, making it the world's largest seller by count, and China's Ministry of Industry and Information Technology expects national output to exceed 100,000 units in 2026. Bank of America's 2026 analysis puts a Western-built pilot humanoid at roughly $90,000 to $100,000 against a Chinese bill of materials nearer $35,000. Figure, the U.S. valuation leader at around $39 billion, runs paid factory pilots. Tesla, having missed its 5,000-unit Optimus target for 2025 and with Elon Musk conceding in January that none were doing useful work yet, is still converting a car line at Fremont.
That is the gap XPeng is trying to exploit, and Turing is the lever. Buying inference silicon means paying someone else's margin on every unit and inheriting their roadmap — an especially fragile position for a Chinese firm facing export controls on high-end accelerators. Designing it in-house lets XPeng amortize one chip program across cars, robotaxis and robots, and it is the only credible path to a bill of materials where compute is not the dominant line item at scale. The policy tailwind helps too. China's lead comes from a combination of policy support, public investment, mature supply chain, and advancements made in AI software and hardware, Omdia analyst Lian Jye Su told Rest of World.
Running the foundation model onboard also changes the product. Latency drops, operational data stays local, and — critically for enterprise buyers — the robot keeps working when the network does not.
What to watch
One robot walking off a line is not mass production, and every humanoid timeline in this industry has slipped, XPeng's included. The tests between now and December are unglamorous: does the Guangzhou line hold yield as it ramps; does XPeng publish a capacity figure and a price; and do the IRON units stationed in its stores do anything a human would otherwise have been paid to do. Answer those three and the 2027 delivery date becomes a schedule rather than a slogan.
“The robot production lines were created from scratch with no precedent to follow. Today's step is small, but XPeng is building the production lines for an entirely new product category.”— He Xiaopeng, Chairman and CEO, XPeng