How XPeng Is Reusing Its Turing AI Chips for IRON Humanoid Robots

XPeng is extending its self-developed Turing AI silicon from intelligent EVs into humanoid robotics by treating vehicles and robots as different physical platforms for a common AI stack. Its latest claim is that the IRON humanoid carries three Turing chips with up to 2,250 TOPS of effective compute, enough to run XPeng’s Physical AI foundation model on the robot rather than relying on remote operation. 18

That approach matters because it combines a chip program, local AI inference, and a shared model-and-software strategy. But it should be read as a product architecture and company roadmap—not yet as independently validated evidence that a humanoid can operate reliably for long periods in complex environments.

One Physical AI stack, multiple machines

XPeng describes its Physical AI system as a full-stack effort spanning chips, large models, and intelligent hardware for AI cars, robotaxis, humanoids, and flying vehicles. 30 Its VLA 2.0 model is designed to generate actions directly from visual signals through what the company calls a “Vision-Implicit Token-Action” path. 10

The company’s basic proposition is straightforward: a car, a robotaxi, and a humanoid have different bodies and tasks, but each needs to perceive the physical world, make decisions, and produce actions. Reusing a common edge-compute platform and portions of the AI software stack could reduce the amount of technology XPeng must develop independently for every category.

What XPeng says is inside IRON

XPeng says the IRON humanoid uses three in-house Turing AI chips for up to 2,250 TOPS of effective computing power. According to the company, that capacity enables local deployment of its Physical AI foundation model, which it says can support complex tasks without remote operation while reducing inference latency and improving data security. 18

Reporting on IRON similarly describes the robot as using the Physical AI foundation model shared with XPeng’s intelligent-driving cars and flying-car efforts. 19 The important qualifier is that the 2,250-TOPS figure and claims about local autonomy are XPeng’s own specifications and demonstrations; they are not a public, independent benchmark of sustained robot capability.

The vehicle connection: Turing in Ultra models

The same compute figure is central to XPeng’s advanced vehicle program. XPeng has described its Next P7 and IRON as both featuring 2,250 TOPS of computing power. 30 In newer Ultra vehicles, three Turing chips support VLA 2.0, while the company has positioned a third chip as enabling a vehicle-wide voice-driven “super agent.” 13

VLA 2.0 is not simply a conversational assistant. XPeng characterizes it as a physical-world model for understanding, prediction, and action generation, with visual signals translated directly into action commands. 10 Applied to a car, that means driving-related perception and decisions; applied to a humanoid, the intended result is perception and action in a robot body.

Why a shared chip platform could be useful

Reuse across hardware and software

A common silicon family can give XPeng a foundation for reusing model optimization, compiler work, runtimes, data pipelines, simulation practices, and over-the-air deployment methods. XPeng has already had to optimize its stack to put VLA-style models into production vehicles, and it presents VLA 2.0 as part of a cross-device Physical AI system. 1030

This does not mean that an automotive model can simply be copied into a humanoid. It means the underlying deployment experience—running large, latency-sensitive AI models at the edge—may be reusable.

Local inference, lower latency, and less data transfer

For a robot, processing sensor input on-device can avoid a network round trip before it acts. XPeng specifically cites lower latency and improved data security as reasons for running its Physical AI model locally on IRON. 18

Local processing can also limit the need to send camera, audio, and environmental data to remote servers. The practical benefit will depend on what data XPeng retains or transmits in real deployments, which has not been publicly detailed.

Manufacturing and supply-chain leverage

XPeng is an EV manufacturer with experience developing and integrating electronics at scale. Extending its in-house Turing program into robots could let it spread chip-development, integration, and validation investments across more than one product category. XPeng has explicitly framed its broader system as spanning self-developed chips, models, and hardware. 30

Whether that becomes a material cost advantage will depend on production volumes, chip yields, robot component costs, and the amount of specialized hardware IRON requires—none of which are publicly disclosed in enough detail to quantify.

Shared silicon does not make cars and humanoids the same problem

Autonomous driving and humanoid robotics both require perception and action, but their control demands differ sharply. A vehicle must understand scenes, predict movement, plan routes, and control driving. A bipedal humanoid also has to maintain balance, manage contact forces, coordinate many joints, manipulate objects, and recover safely from disturbances.

That means Turing chips are unlikely to be the robot’s entire control system. High-rate deterministic motion-control loops, actuator control, sensing, and safety systems may need specialized components alongside the AI compute. XPeng has not publicly released a detailed diagram showing how IRON partitions work across its three chips, conventional controllers, and any dedicated safety hardware.

The unanswered questions behind 2,250 TOPS

The published number alone does not reveal real-world robot performance. Key missing details include:

• Model partitioning: XPeng has not publicly explained what runs on each of IRON’s three chips, how they communicate, or which functions remain outside the foundation model.
• Power and thermals: 2,250 TOPS is an effective-compute claim, not a disclosed continuous power budget. XPeng has not released sustained wattage, thermal behavior, battery-life impact, or performance after throttling for IRON.
• Latency under motion: The company says local inference improves latency, but has not published end-to-end latency measurements during walking, manipulation, or recovery events. 18
• Independent validation: Public evidence largely consists of XPeng announcements and demonstrations. There are no cited public standardized tests comparing IRON’s long-duration, contact-rich, safety-critical performance with competing humanoid compute platforms.

What would show the strategy is working

XPeng has established a coherent thesis: use in-house Turing chips and a shared Physical AI model foundation to extend its capabilities beyond EVs. IRON’s reported three-chip, 2,250-TOPS configuration is the clearest expression of that thesis. 1819

The next evidence to watch is more operational: a public chip-and-controller architecture, measured power and latency under realistic robot workloads, endurance results, clear disclosure of what tasks run fully on-device, and independent safety testing. Those data points—not the TOPS rating alone—will determine whether XPeng’s common silicon strategy becomes a durable advantage in humanoid robotics.