On 24 August 2026, XPENG said its robotics business signed share-purchase agreements for more than US$900 million at a post-money valuation above US$6.3 billion. The company called it the largest single-round private raise in China’s embodied-AI industry so far. IDG Capital led. Gaorong Ventures joined. Tencent and Alibaba came in as strategic investors. That is a capital event, not a new datasheet. The same release still asks a factory buyer to trust DOF counts and chip TOPS while joint torque, payload, and continuous duty stay off the page.
He Xiaopeng framed IRON as a general-purpose humanoid built on XPENG’s in-house stack — chips, controllers, motion modules, dexterous hands — and said the round will fund R&D, Physical AI model training, data generation, mass-production facilities, and global commercial expansion. Mass production is targeted by the end of 2026. First deployments are XPENG’s own stores and campuses. Customer deliveries in China and overseas are aimed at 2027. XPENG keeps controlling ownership and consolidates the robotics unit. Those are the dates that matter next to the valuation.
76 DOF and 21 per hand are claims. Torque is not.
The funding release puts IRON at 76 degrees of freedom across the body and 21 in each hand, wrapped in what XPENG calls a proprietary fully enclosed flexible lattice structure. Three Turing AI chips are rated at up to 2,250 TOPS of effective on-board compute so the Physical AI foundation model can run on the robot without remote operation. Those numbers are specific. They are still incomplete for anyone sizing a cell.
A DOF count tells you how many axes the kinematics team booked. It does not tell you peak or continuous torque at the hip, knee, or wrist, reduction ratio, thermal limit after an hour of pick cycles, or how much mass the arm can hold at full reach. Payload is not in the Aug 24 release. Height and mass are not in it either. Secondary catalogs float 173 cm / 70 kg and a 20 kg payload for earlier IRON listings; those are not the numbers XPENG put in this raise. I will not treat a third-party card as the 2026 production sheet. If you are buying motion hardware, ask for the torque-speed curve and the continuous rating, not the DOF headline.
Hands are where the marketing and the work diverge fastest. Twenty-one DOF per hand sounds dense. Without published fingertip force, tactile resolution, and a duty cycle for repeated grasp-release, it is a finger count. Dexterous hands fail on heat, cable routing, and seal life long before they fail on a conference stage. The raise funds iteration. It does not publish those failure modes.
An earlier XPENG AI Day next-gen IRON brief talked about 82 DOF and 22 per hand with a humanoid spine and solid-state battery language. The Aug 24 raise says 76 and 21. That is a live naming and numbering problem. Until XPENG publishes one locked table for the mass-production unit, treat every DOF figure as release-specific, not as a stable SKU. Solid-state pack claims belong in the same bucket until there is a watt-hour number and a thermal envelope next to them.
Store and campus first is the honest path
Starting inside XPENG retail and campus spaces is the right order if the goal is data and controlled floors. A showroom has known lighting, known furniture, and staff who already know the brand. It is not a Tier-1 body shop with changing carts, oil, and people who will walk through the work envelope. The release does not claim factory takt time, mean time between assists, or a published safety case for shared floors. “Automotive-grade quality standards” is a manufacturing promise. It is not a published joint continuous-duty number.
The Guangzhou mass-production base story from earlier in 2026 — on the order of 110,000 square metres in Tianhe — is the industrial bet behind the raise. Capacity without a published actuator bill of materials is still a building and a schedule. Watch whether the first store robots run a full shift without a pendant in someone’s hand, and whether XPENG ever posts wrist and knee torque the way Unitree posts a knee number on H1.
Compute at 2,250 TOPS matters if the model actually runs locally at that budget under load. Edge inference for locomotion plus manipulation plus speech is a power and thermal problem as much as a FLOPS problem. The press release does not publish pack capacity, average draw, or how long IRON walks and manipulates before a swap. Without that, TOPS is a chip slide, not a shift length.
What this raise does not buy the reader
Nine hundred million dollars buys time, talent, and a factory fit-out. It does not invent a public payload figure, a thermal map, or a continuous-torque curve. It does not settle the 76-versus-82 DOF mismatch.
Compare the raise to what is still missing on a purchase order. No peak and continuous joint torque table. No rated payload at reach. No published mean-time-between-assist for a store shift. No open safety case for co-located work. Investors can underwrite a roadmap on DOF and TOPS. A plant engineer cannot. The useful test after this round is simple: does XPENG ship a locked production datasheet with the first store fleet, or does the marketing stay one slide ahead of the joint numbers?
FIRGELLI does not make XPENG’s rotary joint modules, and this is not a claim that we do. Linear actuators belong on the fixtures around a cell — the lift, the locate, the hatch — not as a substitute for an unpublished hip torque. If you are comparing IRON to Atlas, Optimus, or Walker-class industrial bipeds, hold XPENG to the same bar: publish height, mass, payload, joint torque, and duty cycle for the unit that will actually leave the Guangzhou line. The raise is real. The thin sheet next to it is also real.