XPeng starts an 80%-automated line to build IRON humanoids

XPeng commissioned a dedicated humanoid robot manufacturing facility on September 8, 2026, and says more than 80% of the line’s core processes are automated. The company describes it as the world’s first automated production line for advanced general-purpose humanoid robots — a claim XPeng makes about itself, and one no third party has verified. Mass production of its IRON humanoid is targeted by the end of 2026, with official launch and customer deliveries in China and overseas markets set for 2027.

The robot is not the news here. The line is. Every humanoid programme with a credible model has been throttled by the same thing — the inability to build the machine repeatably, at cost, in numbers above the low hundreds. XPeng is the first of the vehicle makers to answer that with automotive process discipline rather than another demo reel.

Manufacturing was always the bottleneck

Chinese makers shipped the overwhelming majority of the roughly 19,100 humanoids delivered worldwide in the first half of 2026, and they did it with conventional assembly — skilled operators, jigs, and a great deal of hand-fitting. That works to a few thousand units. It does not work at the volumes the sector’s valuations imply. XPeng arrives holding the asset almost no humanoid startup has: vehicle plants, a tier-one supplier base, and process engineers who already run high-mix lines at scale.

That advantage is not automatic, and the counter-argument is specific. Car bodies are large, rigid and geometrically forgiving. Humanoid assembly is dominated by the opposite — small compliant subassemblies, tendon routing, fine cabling, and hands carrying 21 degrees of freedom apiece. Those are precisely the steps that resist automation, and “over 80% of core processes” leaves the residual 20% undefined. If the remainder is hand assembly, the hand is still the rate limiter.

What IRON is, on XPeng’s numbers

XPeng specifies 76 degrees of freedom across the body, 21 in each hand, and three of its in-house Turing AI chips delivering up to 2,250 TOPS, with its physical-AI foundation model running on the robot rather than in the cloud. On-robot inference at that budget is the more meaningful figure: it removes a latency and connectivity dependency that has quietly constrained factory deployments elsewhere. Chairman and CEO He Xiaopeng framed the build as unmapped territory — “The robot production lines were created from scratch with no precedent to follow.”

The financing is already in place. XPeng carved the robotics business out as Dogotix and raised about $900m at roughly $6.3bn post-money, announced August 24, 2026, led by IDG Capital with Gaorong Ventures, Alibaba and Tencent among the investors. That is capital sized for tooling and inventory, not for research.

How to read the claim

Treat “world’s first” as marketing and the 80% as unaudited. What is checkable is the sequence: XPeng said in July it would push monthly IRON output past 1,000 units before a 2027 rollout, and the September line is consistent with that plan rather than a substitute for it. The number that settles the question is not automation percentage. It is units shipped to third-party customers in 2027, and the price they pay. Watch the bill of materials as well: three Turing chips and 76 actuators are a cost structure, and an automated line only pays for itself if the per-unit price falls fast enough to open a market beyond XPeng’s own showrooms.