XPeng carves out its robotics unit and raises $900M at $6.3B

XPeng announced on August 24, 2026 that its robotics business has raised more than $900 million at a post-money valuation of more than $6.3 billion. IDG Capital led the round, Gaorong Ventures participated, and Tencent and Alibaba came in as strategic investors. XPeng calls it the largest single-round private financing in the history of China’s embodied AI industry, and on the disclosed numbers that claim holds.

The composition is more interesting than the headline. Of the $900 million-plus, roughly $600 million comes from external investors, about $200 million from an XPeng subsidiary and about $100 million from company leadership. Around a third of the round is money moving inside the house. That does not make the valuation fake — internal participation is common in carve-outs and often demanded by outside investors — but the amount of genuinely third-party validation here is $600 million, not $900 million.

The carve-out is the real announcement

Alongside the raise, XPeng is spinning its robotics assets, intellectual property and staff into a standalone subsidiary over the next 18 months, retaining roughly 82% ownership. Chief executive He Xiaopeng took direct control of the robotics business in June 2026.

At 82%, the unit stays consolidated on the group’s financial statements while acquiring a separate market valuation and a separate cap table. That is the standard pre-listing shape. It gives the robotics business its own currency for hiring and acquisitions, gives outside investors a defined instrument, and gives XPeng the option of a separate listing later without a further restructuring. Every Chinese automaker with a robot programme is now watching whether this structure gets rewarded; Unitree’s post-IPO reception, covered separately today, is the counter-argument.

What IRON is, and what it has not done yet

IRON carries 76 degrees of freedom across the body and 21 in each hand, wrapped in what XPeng describes as a fully enclosed flexible lattice structure. Compute is three of XPeng’s in-house Turing AI chips for a combined 2,250 TOPS, which the company says lets it run its physical-AI foundation model on the robot itself rather than off-board — and therefore, it claims, perform complex tasks without remote operation.

That last claim is the one to hold onto. Teleoperation behind the curtain is the standing suspicion in humanoid demonstrations, and on-device autonomy is exactly the assertion that cannot be verified from a stage. IRON has not been sold to a customer. The proof arrives when a unit is operating somewhere its operator does not control the lighting.

Why the manufacturing argument carries weight

XPeng’s pitch is that automotive supply chains and automotive quality standards transfer to humanoids. Unlike most companies making that argument, it is currently building hardware at volume: 103,295 vehicle deliveries in the second quarter of 2026, up 64.8% quarter-on-quarter. A firm shipping a hundred thousand complex assemblies a quarter has a credible claim on process control that a robotics startup does not.

The transfer is not automatic. A car is built once and driven by its owner; a humanoid is built once and then has to keep working through contact, wear and unstructured environments its maker never saw. Automotive volume solves the manufacturing half of the problem and says nothing about the second half. The 1-million-units-by-2030 figure should be read as a factory-planning assumption, not a demand forecast — no customer has ordered at that scale, from anyone.

Still, the near-term milestone is unusually checkable. XPeng says IRON goes into production by the end of 2026 and starts work inside its own stores and campuses. That is four months away, in venues anyone can walk into.