Is Musk Conceding Ground? And Does XPENG’s IRON Signal China’s Lead in the Humanoid Robot Race? – CleanTechnica

---

Support CleanTechnica’s work through a Substack subscription or on Stripe.

---

GUANGZHOU, CHINA — In a rare public acknowledgment of competition, Elon Musk seemed to have commented on Chinese EV maker XPENG’s newly unveiled humanoid robot with what seems to be an admission: “Not bad … Tesla and China companies will dominate the market. Other companies in the West are weak.”

The statement was sent as a DM following XPENG’s dramatic unveiling of its next-generation IRON robot last week. The robot’s catwalk marks an inflection point in the global race for humanoid robotics—one in which the competitive axis has shifted decisively away from Silicon Valley versus the world, and toward a two-horse race between American and Chinese tech giants.

XPENG CEO He Xiaopeng posted swiftly on X: “Let this stand as the final proof: the robot that mastered the catwalk is built by a Chinese startup.” The exchange, part congratulatory and part combative, underscores an emerging reality: the humanoid robot market, long dominated by research prototypes and cautious timelines, is now a high-stakes industrial contest with mass production as the finish line.

Again, thank you to “Insidar,” my contact attending XPENG’s 2025 AI Day, for sharing his notes and recordings. CleanTechnica wanted very much to be in Guangzhou so we could intensely poke at Tesla to rush its robot, but visa complications obstructed that.

Built like us?

IRON’s architecture is organized around a “bone–muscle–skin” framework that mirrors human anatomy rather than simply approximating it. The robot features a humanoid spine capable of flexion, allowing it to bend, balance dynamically, and even recover from falls—a critical capability for real-world deployment. Bionic muscles provide actuation that mimics organic motion profiles, and the entire body is enveloped in flexible synthetic skin that XPENG says can be customized for different body shapes and proportions.

The devil, as always, is in the joints.

IRON supports 82 degrees of freedom across its entire body, a figure that puts it among the most articulated humanoid platforms announced to date. The hands are the engineering crown jewel: 22 degrees of freedom per hand, achieved using what XPENG claims is the industry’s smallest bevel gear joint. This enables 1:1 human hand sizing—critical for operating in environments designed for human ergonomics—and allows for delicate manipulation tasks such as grasping fragile objects or operating touch interfaces.

Powering this mechanical ballet is another industry first: IRON is the world’s first humanoid robot to run on an all-solid-state battery. The choice reflects XPENG’s automotive DNA—solid-state technology offers higher energy density and enhanced safety over lithium-ion, particularly relevant for robots intended to operate in enclosed public spaces or, eventually, homes. It’s also a not-so-subtle signal that XPENG is leveraging its EV supply chain advantage, a manufacturing edge Tesla also enjoys but has yet to deploy in Optimus.

The controversy that proved the point

XPENG’s reveal didn’t just showcase technology—it was also theatrical.

When the IRON robot walked onto stage at the company’s 2025 AI Technology Day with movements more fluid and lifelike than any other seen before, observers suspected a human operator was hidden inside (the same was said of Tesla robots). However, XPENG’s engineers took a chainsaw to skepticism—literally. In a moment that instantly went viral, they cut open the robot’s leg on stage, exposing metal joints, actuators, and wiring. No hidden human. Just hardware sophisticated enough to fool the eye.

The stunt was calculated, but the underlying engineering is what caught Musk’s attention. Where Tesla’s Optimus has emphasized functionality and iterative development, XPENG has pursued what it calls “extreme anthropomorphism”—a design philosophy that aims not just to replicate human capability, but to overcome the “uncanny valley” through biomimicry so precise it becomes socially indistinguishable from the real thing.

Where cars meet robots

If the hardware is impressive, the software is where XPENG is making its most aggressive bet. IRON runs on the company’s first-generation Physical-World Large Model, a multi-model cognitive stack adapted directly from XPENG’s autonomous driving program. This isn’t retrofitted AI—it’s a deliberate convergence strategy where “Physical AI” becomes the umbrella term for both self-driving cars and ambulatory robots.

The system integrates three AI subsystems into what XPENG calls the VLT + VLA + VLM architecture. The VLA (Vision-Operation-Action) model, inherited from the automotive division, is the action generative model responsible for translating sensory input into motor commands. The VLM (Vision-Language-Model) handles natural language dialogue and social interaction. But the novel component is VLT (Vision-Language-Task), a new large model XPENG developed specifically for robotic reasoning. VLT functions as the “core engine” for autonomous decision-making, allowing IRON to decompose complex instructions, plan multi-step tasks, and adapt to dynamic environments without constant human supervision.

Processing this multi-model stack requires serious compute. IRON is equipped with three of XPENG’s in-house Turing AI chips, delivering a combined 3,000 TOPS (Tera Operations Per Second) of onboard processing power. For context, that’s roughly equivalent to the neural compute available in a high-end autonomous vehicle—except now it’s managing 82 articulated joints in real-time, processing visual and linguistic input simultaneously, and maintaining balance on two legs.

XPENG has also introduced what it calls the “Fourth Law” of robotics for IRON: privacy data does not leave the robot. In an era of escalating data security concerns, particularly around AI systems deployed in homes and public spaces, this local-processing mandate is both a technical constraint and a marketing promise. It also hints at the regulatory headwinds any mass-market humanoid will face.

The mass production gauntlet

Musk’s comment about Tesla and Chinese companies dominating wasn’t idle speculation—it was an acknowledgment of a hard truth. Mass-producing humanoid robots at automotive scale is an industrial problem, not a research one, and it requires the kind of vertically integrated manufacturing infrastructure that only EV makers and a handful of industrial giants possess. XPENG is betting that its automotive supply chain, already optimized for high-volume production of complex mechatronic systems, gives it a time-to-market advantage that pure robotics startups cannot match.

The company has established its first embodied intelligence data factory in Guangzhou, purpose-built to generate the training data that humanoid systems desperately need. One of the field’s persistent bottlenecks is the lack of large-scale, task-diverse datasets for training robots in real-world environments. By creating a dedicated facility to capture interaction data, XPENG is addressing this constraint head-on, much as Tesla has done with its real-world driving data advantage in autonomous vehicles.

XPENG is targeting mass production by the end of 2026, with production preparation beginning in April 2026. If achieved, this timeline would position IRON among the first commercially mass-produced humanoid robots globally—a status that carries significant first-mover advantages in establishing supply chains, developer ecosystems, and customer expectations.

Your next barista

Unlike the grand visions of domestic assistant robots that have animated the field for decades, XPENG is taking a measured, phased approach to commercialization. The first deployment wave will target controlled commercial environments: IRON will serve as guides, shopping assistants, and receptionists in XPENG’s own retail showrooms. This strategy allows the company to accumulate task-specific operational data while maintaining tight control over the interaction environment and minimizing liability exposure.

Industrial partnerships are also underway. Chinese steel giant Baosteel has signed on as an ecosystem partner, with IRON units slated for deployment in inspection and material sorting roles within Baosteel facilities. These industrial applications serve dual purposes: they provide revenue and operational validation while stress-testing the platform in demanding, unstructured environments that will inform future development.

To accelerate ecosystem development, XPENG is releasing an SDK (Software Development Kit) for IRON, inviting third-party developers globally to build applications and extend the robot’s capabilities. This open platform strategy mirrors successful models in mobile and automotive software, and represents a bet that developer creativity will unlock use cases XPENG hasn’t anticipated. It also sets up IRON as a platform rather than a product—a distinction with significant long-term strategic implications.

The real race is Tesla vs. XPENG

Musk’s acknowledgment that “other companies in the West are weak” was less about XPENG specifically and more about the consolidation he sees coming. The humanoid robot market, still nascent, is already showing signs of a duopoly structure: companies with automotive-scale manufacturing, proprietary AI stacks, and deep pockets on one side; everyone else on the other.

Tesla’s Optimus has focused on functional iteration and cost reduction, with Musk repeatedly emphasizing the goal of a robot priced below a car. XPENG’s IRON has prioritized anthropomorphism and cognitive sophistication, betting that social acceptance and task flexibility will justify premium positioning. Both strategies are rational; both require years of execution to validate.

What’s clear is that the competitive dynamic has shifted. The humanoid robot market is no longer a speculative research domain. It’s an industrial race with visible finish lines, and the competitors with automotive DNA—Tesla, XPENG, and perhaps a handful of others—have pulled decisively ahead. Musk’s comment wasn’t a concession of defeat. It was recognition that the race is now a two-country sprint, and the outcome is far from certain.

He Xiaopeng’s response—”the robot that mastered the catwalk is built by a Chinese startup”—was more than a retort. It was a mission statement. The question now is whether fluid motion and elegant design will matter as much in the market as they do on stage.

Based on confirmed “speculations” by the end of 2026, we’ll have an answer.