On September 8, in a ceremony in Guangzhou, XPeng chairman He Xiaopeng clipped a staff badge onto a humanoid robot. The gesture was theater, but the setting was not. The robot, Iron, had just walked off what XPeng calls the world’s first automated production line for advanced humanoid robots, a line the company says runs with more than 80 percent of its core processes automated. Twelve days later, Dongfeng Motor told Yicai that its own humanoid will enter factories in October for sorting and quality inspection work, with small-batch trial production by year-end. Chery’s robotics unit Aimoga, incubated only in January 2025, has already shipped more than 3,000 robots, two-thirds of them overseas, and is eyeing an IPO.
Strip away the badge ceremony and the pattern is unmistakable: the humanoid industry’s hardest problem is no longer intelligence demos or venture funding. It is manufacturing, and the companies solving it are not robotics startups. They are Chinese automakers.
The Step That Actually Mattered
XPeng’s September 8 commissioning deserves the attention it got, and skepticism about the details matters too. What is established: the Guangzhou line exists, it was purpose-built for Iron, and the company describes it as automotive-grade, meaning the quality systems, supplier qualification processes, and traceability standards used for its electric cars have been pointed at a humanoid product. Iron walked off the line under its own power. What is not established: output volumes, cycle times, yields, or a delivery total. As humanoid.guide noted, none of those manufacturing metrics have been disclosed, and the company has not clarified how autonomous the walk-off itself was.
The honest reading is that one robot walking off one line is a comma, not a conclusion. But it is the first comma in a sentence the rest of the industry has not started writing. Tesla, the company most associated with humanoid manufacturing ambition, is still converting a car line at Fremont for Optimus assembly. Musk predicted roughly 10,000 Optimus units in 2026, admitted in January that none were doing useful work, and has described near-term output as “quite slow.” XPeng did not convert a line. It built a new one, from nothing, for a product category that has no precedent, which is precisely how He Xiaopeng framed it: “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.”
Why Automakers, Why Now
The deeper story is structural. Mass-producing a humanoid robot requires five things at once: capital at automotive scale, a supplier base for actuators, harmonic drives, batteries, and precision castings, in-house or privileged access to AI silicon, a first customer willing to absorb imperfect units, and a distribution channel. Chinese automakers hold all five. Almost nobody else does.
Consider XPeng’s stack. Iron runs on three of the company’s in-house Turing AI chips delivering up to 2,250 TOPS, enough to run its Physical AI foundation model on the robot without teleoperation, at least per company claims. That silicon was developed for cars: the same Turing platform powers XPeng’s VLA 2.0 driver assistance. The company’s robotics arm raised over $900 million at a $6.3 billion valuation in August, the largest single private round in China’s embodied AI sector, led by IDG Capital with Tencent and Alibaba participating. XPeng’s stated trajectory is more than 1,000 robots per month, scaling toward a million a year by 2030. And the first deployments will not go to external customers at all: Iron units go into XPeng’s own stores and campuses before any commercial launch, with sales and deliveries in China and abroad scheduled for 2027.
Notice the loop. The company builds the robot on an automated line, deploys the robot inside its own retail and manufacturing footprint, harvests operational data from those deployments, and iterates. The first customer, the test bed, and the data engine are all the same corporate parent. This is the Tesla playbook executed one step ahead: Tesla pilots Optimus in its own factories, but it is still waiting on its own line. XPeng has the line and the pilots running concurrently.
Dongfeng’s version of the loop is even more explicitly internal. Zhang Zhenlin, the automaker’s chief engineer of intelligent technology, said the humanoid will handle sorting and quality inspection inside Dongfeng factories starting in October, with trial production at year-end and a target of bringing the robot’s working capability to parity with human workers by the end of 2027. His stated rationale is technology reuse: embodied intelligence development feeds and feeds off automotive large models. Dongfeng’s robot dog, notably, will commercialize first, guiding customers at dealerships, which is to say the company’s robot business begins at its own sales counters.
Chery’s Aimoga shows what the loop looks like further along the curve. Incubated in January 2025, it has delivered more than 3,000 robots including 2,000 overseas, a sharp acceleration from 300 humanoids and 1,000 robot dogs shipped by the end of 2025, operates in more than 60 countries, has deployed 110 police humanoids, and targets 10,000 deliveries next year. Since April it has sold its Mornine M1 humanoid directly to consumers on JD.com at 285,800 yuan, about $42,000. The Mornine is not Iron: roughly 1.68 meters tall, around 40 degrees of freedom, a service and reception robot rather than a general laborer. But it is a shipping consumer product with a price tag, a sales channel, and a delivery record, three things most Western humanoid companies cannot claim.
The Margin Question Nobody Can Answer Yet
The most audacious XPeng claim is not the 2,250 TOPS or the 80 percent automation rate. It is the assertion that Iron’s per-unit gross margin will run significantly higher than the company’s EVs. Electric vehicle margins sit in the single digits across most of the Chinese market, which would put a “significantly higher” robot margin somewhere in the double digits. If true at scale, that inverts the industry’s unit economics overnight: the robot stops being a subsidized bet and becomes the most profitable product an automaker sells.
If true. No humanoid company has demonstrated a positive gross margin on a general-purpose machine at any volume, and XPeng has not shipped a robot to a paying customer. The claim rests on assumptions about actuator costs, automated assembly reducing labor content, and silicon economics that only hold if the company hits volume. One thousand robots a month is the number to watch, and December is the deadline XPeng itself set for mass production. Every humanoid timeline in this industry has slipped. XPeng’s has too. The difference is that this time the line exists, which converts a promise into a falsifiable milestone: either Iron units come off the Guangzhou line at rate before January, or the automaker turn stalls at the comma.
There is also a genuine technical unknown inside the automation claim. An 80 percent automated line is good for repeatability, but humanoid robots are not cars: they have 76 degrees of freedom in the body and 21 in each hand, wrapped in a flexible lattice skin, with tolerances and failure modes that automotive processes were never designed to hold. Nobody, anywhere, has automated the assembly of dexterous hands at rate. That last 20 percent of manual processes is where the real cost lives, and it is the part least amenable to the automotive playbook.
Two Models of Scaling, One Direction
Set the Chinese automaker wave against the American picture and the industrial logic of this moment comes into focus. The United States is scaling humanoids through AI-first companies: Figure chasing capability leaps with Helix, Agility taking Digit public through a SPAC on the strength of a logistics order book, Tesla vertically integrating Optimus from silicon to factory. The frontier of capability arguably lives in the American model. The frontier of production now lives in the Chinese one.
The regulatory backdrop makes the divergence sharper. The FCC’s Covered List rule, finalized in July, blocks new equipment approvals for foreign humanoids in the US, and Chinese automakers read the writing: XPeng plans overseas deliveries in 2027 but its near-term deployments are domestic, Aimoga’s overseas volume skews toward markets that are not the United States, and Dongfeng’s entire first chapter happens inside its own fence line. The result, as the supply chain splits, is a division of labor neither side designed: American firms must prove humanoids can be built outside China’s component ecosystem, while Chinese automakers must prove that scale without American customers still produces a real industry.
The automakers’ wager is that it does, because the demand side does not require consumers at all. Factories, dealerships, police forces, and export markets across Southeast Asia, the Middle East, and Latin America absorbed Aimoga’s first 3,000 units. XPeng’s first thousand Irons have a guaranteed home in company stores. Dongfeng’s first batch has a guaranteed home in Dongfeng plants. None of that depends on a Western consumer ever buying one.
What December Tells Us
Three checkpoints now sit between the industry and its next reckoning, all in the same quarter. XPeng must hit end-of-2026 mass production on the Guangzhou line, and the metric that matters is not the ceremony but the monthly rate. Dongfeng must convert an October factory pilot into a year-end trial batch. Aimoga must hold its trajectory toward 10,000 deliveries next year across its 60-country footprint.
If even two of the three land, the humanoid industry’s center of manufacturing gravity finishes its move this year, and the 2027 conversation changes from “can anyone build these” to “who can match Chinese automaker cost curves.” Tesla’s Optimus V3, Figure’s Helix deployments, and Agility’s public-market scrutiny will all then be measured against a benchmark set in Guangzhou: a robot with a staff badge, walking off a line that mostly built itself.
The badge was theater. The line was not.