# Has XPeng Beaten Tesla to Humanoid Production at Scale?

[Xpeng Robotics](https://humanoidintel.ai/companies/xpeng-robotics) has commissioned a dedicated humanoid production facility in Guangzhou, where its IRON robot autonomously walked off the assembly line following completion of manufacturing — a milestone that lands as [Tesla (Optimus Division)](https://humanoidintel.ai/companies/tesla-optimus) pushes commercial Optimus sales to the second half of 2027, slipping past an earlier target of late 2026. The IRON platform carries 76 [degrees of freedom](https://humanoidintel.ai/glossary/degrees-of-freedom) across its full body — including 21 DOF per hand — and runs three proprietary Turing AI chips delivering up to 2,250 TOPS of effective onboard compute. More than 80 percent of the production line's core processes are automated, according to XPeng. Mass production is targeted for the end of 2026, with initial commercial-scenario deployments beginning in XPeng's own stores and campuses, followed by a formal market launch and deliveries in China and overseas markets in 2027.

The convergence of two datapoints — a Chinese automaker achieving production-line deployment while the most-watched Western competitor slips its schedule by roughly a year — marks a concrete shift in who is setting the manufacturing pace in humanoid robotics.

---

## IRON's Hardware Stack: What the Specs Actually Mean

The 76-DOF total body count, with 21 DOF in each hand alone, positions IRON at the higher end of the [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) spectrum among robots currently in production. For context, high-DOF hands are necessary for the kind of in-hand re-orientation that real-world general-purpose tasks demand — but they also introduce significant reliability and calibration complexity at manufacturing scale. XPeng has not disclosed actuator type (whether harmonic drive, linear, or otherwise), which matters enormously for backdrivability, force control, and long-term maintenance burden.

The onboard compute story is notable. Running 2,250 TOPS of inference directly on the robot — rather than offloading to cloud or edge servers — is a deliberate architectural choice targeting latency reduction and operational data sovereignty. For a [Physical AI](https://humanoidintel.ai/glossary/physical-ai) strategy built around self-reinforcement from real-world experience, keeping inference local means the robot can continue functioning without network dependency and accumulate proprietary operational data that doesn't leave the platform. That data flywheel, if it compounds across a deployed fleet, becomes a durable competitive asset.

XPeng describes IRON as built around a proprietary fully enclosed flexible lattice structure, designed to balance mobility, aesthetics, and safety. The company frames IRON as a general-purpose platform intended to improve through self-reinforcement using real-world experience — language that gestures toward on-robot learning loops, though the source material does not specify the model architecture or training methodology in detail.

---

## The Automotive Manufacturing Angle Is the Real Thesis

XPeng's most credible claim isn't about AI — it's about manufacturing. The company is explicitly applying automotive production discipline to humanoid robotics: high process automation rates, dedicated facilities, and the quality-control infrastructure that vehicle manufacturing demands. This is the same thesis underpinning BYD's and other Chinese automakers' moves into robotics: the supply chain relationships, stamping and assembly expertise, and quality-at-volume experience built over decades of car production transfer meaningfully to humanoid manufacturing.

That said, skepticism is warranted. Automotive manufacturing expertise handles metal forming, painting, and powertrain assembly well — but humanoid robots introduce cable routing, sensor calibration, actuator backlash tuning, and software commissioning steps that have no direct automotive analogue. An 80%-automated production line is impressive, but the remaining 20% of manual processes in a system this mechanically complex can dominate cycle time and yield. XPeng has not disclosed production rate, yield figures, or unit cost targets.

---

## Tesla Optimus: What the Schedule Slip Signals

According to the source, Tesla now expects Optimus production to begin in the coming months, with commercial sales targeted for the second half of 2027. Earlier plans had called for production this summer and initial sales by late 2026. The next-generation model's design and scale are described as dependent on field experience — suggesting Tesla is taking a data-driven, iterative approach rather than locking specifications ahead of deployment.

A roughly one-year slip on commercial sales is significant but not surprising. Scaling a novel electromechanical system from prototype to reliable production units involves failure modes that only emerge at volume. The more telling signal is Tesla's posture: waiting for field experience before finalizing next-gen design suggests the company recognizes it doesn't yet have sufficient real-world data to commit to a production architecture with confidence. That's intellectually honest engineering, but it cedes the manufacturing timeline narrative to competitors who are willing to move faster and accept higher early-unit variance.

---

## Industry Trajectory: China's Production Lead Is Real

XPeng joins a cohort of Chinese humanoid developers — including [AGIBot](https://humanoidintel.ai/companies/agibot), [Unitree Robotics](https://humanoidintel.ai/companies/unitree-robotics), and [UBTECH Robotics](https://humanoidintel.ai/companies/ubtech) — that have moved into or toward dedicated production facilities faster than most Western peers. The pattern reflects both state-level policy support for humanoid robotics in China and the structural advantage of being embedded in one of the world's densest manufacturing ecosystems.

For Western competitors and their investors, the relevant question is whether early production volume translates into the data advantage that matters for [Physical AI](https://humanoidintel.ai/glossary/physical-ai) development. A robot fleet operating in real commercial environments — even in limited scenarios like XPeng's own stores and campuses — generates the kind of unscripted, long-horizon interaction data that simulation cannot replicate. If IRON accumulates that data while Optimus remains pre-commercial, the gap becomes harder to close on the AI side, not just the hardware side.

XPeng CEO He Xiaopeng framed the ambition directly: "We aim to build a new type of robot with full generalization capabilities that can truly become part of everyday life, create a better life for people, and ultimately become a companion in their lives." The aspiration is standard for the category. The differentiated question is whether a production line in Guangzhou and a campus deployment fleet can generate the diversity of real-world experience needed to approach [zero-shot generalization](https://humanoidintel.ai/glossary/zero-shot-generalization) across unstructured environments — something no humanoid has demonstrated convincingly at scale.

---

## Key Takeaways

- **XPeng has commissioned a dedicated IRON humanoid production facility in Guangzhou**, with more than 80% of core production processes automated, per the company.
- **IRON carries 76 DOF total body, 21 DOF per hand**, and runs up to 2,250 TOPS onboard via three Turing AI chips — all figures sourced directly from XPeng's announcement.
- **Tesla Optimus commercial sales have slipped to H2 2027**, roughly a year later than prior guidance of late 2026, per the source.
- **XPeng plans mass production by end of 2026**, with initial commercial-scenario rollouts in its own stores and campuses, followed by China and overseas market launch in 2027.
- **The automotive-to-robotics manufacturing transfer thesis is credible but unproven at yield and volume** — XPeng has not disclosed production rate, unit cost, or yield data.
- **China's humanoid production momentum is structural**, not episodic — driven by supply chain depth, policy support, and automotive manufacturing infrastructure.

---

## Frequently Asked Questions

**What are the specs of XPeng's IRON humanoid robot?**
According to XPeng, IRON has 76 degrees of freedom across its full body and 21 DOF in each hand. It runs three Turing AI chips providing up to 2,250 TOPS of effective onboard computing power. The robot is built around what XPeng describes as a proprietary fully enclosed flexible lattice structure.

**When will XPeng IRON be available commercially?**
XPeng says mass production is targeted for the end of 2026, with initial deployments in its own stores and campuses. A formal market launch and deliveries in China and overseas markets are planned for 2027.

**Why is Tesla Optimus delayed?**
The source reports that Tesla's Optimus production timeline has shifted, with commercial sales now targeted for the second half of 2027, compared to earlier guidance of late 2026. Tesla has indicated that its next-generation model's design and scale will depend on field experience, suggesting an iterative approach tied to real-world data.

**How does XPeng's automotive background help in humanoid robotics manufacturing?**
XPeng argues its automotive manufacturing expertise — including high process automation, supply chain relationships, and quality-at-volume infrastructure — transfers to humanoid production. The company reports more than 80% automation on IRON's production line. However, humanoid-specific steps like actuator calibration, cable routing, and software commissioning have no direct automotive equivalent, so the transfer is partial rather than complete.

**What is XPeng's Physical AI strategy for IRON?**
XPeng describes IRON as a [Physical AI](https://humanoidintel.ai/glossary/physical-ai) platform designed to operate safely and adaptively across environments, improving through self-reinforcement from real-world experience. Onboard inference via the Turing chips is intended to reduce latency and keep operational data on the robot, supporting a data flywheel as the fleet scales.