--- title: "XPeng's Physical AI roadmap puts one chip stack across cars, humanoids and flying cars" date: 2026-10-07 category: Companies & Stack site: NeuroAI canonical: https://neuroai.site/a/na-corp-xpeng-physical-ai-ecosystem language: en --- # XPeng's Physical AI roadmap puts one chip stack across cars, humanoids and flying cars > At its 2025 AI Day, XPeng laid out a single embodied-intelligence plan — robotaxis, the IRON humanoid and flying systems — all running on its own Turing AI chips. On 5 November 2025, inside XPeng Science Park in Guangzhou, chairman and CEO He Xiaopeng (何小鹏) did not unveil a car. He unveiled a worldview: XPeng, he said, is now "an explorer of mobility in the Physical AI (物理AI) world, and a global embodied-intelligence company." The proof was four products on one stage — a vision-to-action model, a robotaxi, a next-generation humanoid, and a flying system — all sharing the same in-house silicon. ## One model to drive everything that moves The headline software is **VLA 2.0**, a vision-to-action large model (大模型). Where older pipelines translated "see" into "language" into "act," XPeng says VLA 2.0 drops the language step and generates actions straight from vision, end to end. The company describes a cloud training cluster of about 30,000 cards, a model of around 72 billion parameters, and a five-day full retraining cycle — figures XPeng announced on stage. Crucially, VLA 2.0 is pitched as cross-domain: it is meant to drive cars, robotaxis, humanoid robots and flying vehicles from the same physical-world reasoning. ## IRON: the humanoid that broke the uncanny valley The star of the show was **Next-Gen IRON**, a humanoid (人形机器人) with an explicitly anthropomorphic build — a human-like spine, bionic muscles and fully covered flexible skin. XPeng said it supports 82 degrees of freedom across the body and 22 in each hand, reaching a 1:1 hand size through small harmonic joints. On the software side IRON runs a "VLT + VLA + VLM" stack (vision-language-task, vision-language-action, vision-language-model) and carries a first-generation physical-world model. Compute is the interesting part for a chip story: XPeng says IRON runs on three in-house **Turing AI chips** with about 3,000 TOPS of effective onboard compute, enough for conversation, walking and interaction. Energy comes from all-solid-state batteries for lighter weight and higher density, and XPeng added a "fourth safety rule" keeping human privacy data on the device. Because the movement looked almost too human, He Xiaopeng returned 24 hours later and cut open the robot's clothed leg on stage to prove it was machinery — a stunt that went viral and, incidentally, made the engineering believable. ## Robotaxi and flying systems close the loop Alongside IRON, XPeng announced a量产 (mass-production) robotaxi for 2026 using four Turing chips at 3,000 TOPS, a pure-vision system with no lidar and no HD maps, and three global robotaxi models planned for the year. The flying piece, **ARIDGE**, includes the A868 tilt-rotor hybrid with a stated ~500 km range and ~360 km/h top speed, and the Land Aircraft Carrier (陆地航母) eVTOL with reported global pre-orders past 7,000 units. Volkswagen was named a launch partner for VLA 2.0 and its chip; Amap (高德) for robotaxi; and Baosteel (宝钢) as the ecosystem partner piloting IRON in industrial inspection. ## Why a car company is building robots XPeng's logic is leverage. Autonomous driving already forced it to build perception, planning, chips and real-world data — the same ingredients a humanoid or a flying car needs. By reusing the Turing chip and the VLA family across vehicles, robots and aircraft, XPeng argues it can spread cost and talent across a single embodied-intelligence stack instead of building three separate companies. It targets large-scale humanoid production by the end of 2026, with IRON's first jobs in commercial service — guided tours, shopping assistance, traffic diversion — rather than homes or factories immediately. To feed training, XPeng stood up an embodied-intelligence data factory in Guangzhou. ## Reading the trend According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, XPeng's move reflects a broader Chinese auto-industry shift where the car becomes a robotics platform — and the differentiator is owning the chip and the world model, not the sheet metal. ## Honest limitations Every specification above is XPeng-announced at AI Day 2025, not independently benchmarked. One caveat worth flagging: compute figures have already shifted in later 2025–2026 materials — some 2026 production-version descriptions cite around 2,250 TOPS for the three-chip IRON configuration, versus the ~3,000 TOPS stated at the November 2025 unveiling. Body degrees-of-freedom also appear as 82 (AI Day 2025) in some later notes and 76 in 2026 production descriptions; we report the AI Day 2025 figures as announced. IRON is not yet mass-produced, and XPeng itself acknowledged open challenges in training data and manufacturing scale. The "30,000-card cluster" and "72B-parameter" training claims are vendor statements we did not verify against infrastructure audits. Competing humanoid and Physical AI roadmaps from Tesla, UBTECH and others remain ahead or alongside on deployment, so "first full-stack Chinese player" is a marketing claim, not a settled ranking. ## What readers can do now - Watch the end-2026 mass-production date as the real test; a humanoid shown walking on stage is far from one shipped at volume. - Investors should track whether XPeng's robotics unit spins out or raises at a disclosed valuation, and whether Baosteel pilots convert to paid deployments. - Engineers evaluating Turing-class automotive chips can watch whether XPeng open-sources VLA 2.0 tooling, which would signal how serious the "shared stack" claim is. - Global readers comparing embodied-AI players should weigh XPeng against Tesla Optimus and Chinese peers on a single axis: does one chip and one world model genuinely span car, robot and aircraft, or is it three roadmaps sharing a slogan. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-corp-xpeng-physical-ai-ecosystem Free to quote with attribution and a link to the original.