Remember when XPENG’s IRON humanoid robot walked itself off the company’s production line in Guangzhou in September 2026? It was the kind of moment built for social feeds: a mass-produced, general-purpose robot strolling out of a factory on its own. People shared it, argued about it, and moved on. One month later, XPENG followed with something far less cinematic and, from where I sit, a lot more important. It put its own AI chips in a car, and not just in the expensive version.
The car is the L03, a coupe-SUV that debuted in Munich and launched simultaneously across 65 markets. I review toolkits for a living, so I’m going to treat this the way I treat any new stack. I’ll look at what’s in the box, what the specs promise, and where the claims stop and the testing has to begin.
What actually shipped
The L03 runs on XPENG’s proprietary Turing AI chips, which deliver 2,250 TOPS of compute. That compute handles two jobs: intelligent driving and the smart cockpit. On top of the silicon sits the company’s second-generation VLA system, a vision-language-action model built to interpret road environments. Tying it all together is Sepa, XPENG’s smart electric platform architecture. Sepa integrates the core hardware, electrical systems, infotainment, and full autonomous driving software into one platform, with the stated goals of lowering production costs and speeding up software upgrades.
The headline detail is that this is the first Chinese EV to feature in-house autonomous driving silicon in every configuration. Every trim gets the same brain.
Why “every configuration” beats the TOPS number
Big compute figures make good headlines, and 2,250 TOPS is a big figure. But TOPS is a spec sheet number, not a performance result. It tells you how much raw math the chip can do. It tells you nothing about how well the software turns that math into safe lane changes in the rain. I’ve reviewed plenty of tools with impressive benchmarks that fell apart on real workloads.
The “every configuration” decision is the part that changes things. When a carmaker tiers its driving hardware, the base model buyer often ends up with a different, weaker system. That splits the fleet. The software team has to support multiple hardware targets, and the cheaper cars can get left behind on updates. Putting one chip platform in every L03 means one target for developers and one baseline for every owner. As a toolkit reviewer, that’s the kind of boring architectural choice I trust more than a flashy benchmark.
The VLA model is what I’d want to test
A vision-language-action model is a solid concept on paper. The system sees the road, builds some internal understanding of what’s happening, and decides what to do. Calling it second-generation suggests XPENG has iterated on the approach rather than shipping a first draft.
But I can’t tell you how well it works, and neither can anyone reading the launch materials alone. The verified details describe what the model is meant to do, not how it performs across edge cases, weather, or the wildly different road conditions you’d find across 65 markets. That last point deserves attention. A simultaneous launch in that many markets is ambitious, and driving behavior varies a lot from one country to the next. The real review of this system will come from owners, not press releases.
Sepa and the upgrade question
Sepa is the piece most likely to matter over the life of the car. Integrating hardware, electrical systems, infotainment, and driving software into a single platform is how you make over-the-air updates faster and cheaper to ship. In toolkit terms, it’s the difference between a well-maintained monorepo and a pile of vendor plugins that each break on their own schedule.
There’s a flip side, and XPENG’s own customers are already raising it. One commenter on the company’s launch post, identifying as a 2026 P7+ or X9 customer, said they hope XPENG will offer a paid retrofit. That’s a fair ask. When a company makes its new silicon the standard, earlier buyers start wondering whether their hardware just became the legacy branch.
What works, what doesn’t (so far)
- Works: one in-house chip platform across every trim, which simplifies support and updates.
- Works: an integrated architecture in Sepa aimed at cheaper production and faster software upgrades.
- Works: a single compute pool serving both driving and cockpit features.
- Unproven: real-world performance of the second-generation VLA system.
- Unproven: how one driving model adapts across 65 markets at once.
- Open question: what XPENG offers owners of earlier models with different hardware.
My take
IRON walking off the line got the attention. The L03 deserves more of it. Vertical integration, from chip to model to platform, is the same playbook that has worked for companies that control their whole stack. XPENG has assembled the parts. What I can’t yet score is the experience, because that depends on how the VLA system behaves on real roads. On specs and architecture, the L03 looks like a well-designed toolkit. I’ll hand out the final grade once the field reports arrive.
🕒 Published: