How XPeng Built the World's Most Advanced AI-Native Smart EV Platform

On July 16, 2026, XPeng CEO He Xiaopeng sat in the passenger seat of the all-new XPeng L03 as it navigated Munich's narrow old-town streets, construction zones, and roundabouts—without a driver touching the wheel. The vehicle handled right-before-left priority rules at unmarked intersections, detected cyclists and e-scooter riders, and made early lane changes to avoid parked cars. This wasn't a controlled demo track. It was real Munich traffic, and the system worked. XPeng is not just another Chinese EV company—it's betting the entire company on becoming a "Physical AI" company, and the results are starting to show.

From EV Maker to Physical AI Company

XPeng was founded in 2014 and went public on the NYSE in 2020. For years, it was known as one of China's "Big Three" EV startups alongside NIO and Li Auto—competent, but not the market leader. BYD dominated volume, NIO captured the premium brand, and Li Auto found a niche with extended-range SUVs. XPeng was the "tech-focused" one, but that positioning didn't always translate into sales.

That changed in 2025. XPeng pivoted from being "an EV company with good tech" to "a Physical AI company that happens to make cars." The distinction matters. Under this new vision, XPeng's core competency isn't manufacturing—it's AI. The vehicles are just the first application. The same AI platform powers the company's humanoid robot (IRON), its flying vehicles, and its Robotaxi fleet.

"We are building a Physical AI company. We want to use technology to change how people move and how people live. Technology should be for everyone—not just for a few people." — He Xiaopeng, Chairman and CEO of XPeng, July 2026

VLA 2.0: The AI Brain That Powers Everything

At the heart of XPeng's transformation is the second-generation Vision-Language-Action (VLA) architecture, unveiled in November 2025. VLA 2.0 is what XPeng calls the "industry's first L4 autonomy-oriented foundation model for the physical world."

Traditional autonomous driving systems use a modular architecture: separate modules for perception (what the car sees), prediction (what other objects will do), planning (what the car should do), and control (executing the plan). Each module is developed independently and connected through interfaces. This works, but it's brittle—errors compound as they flow through the pipeline.

VLA 2.0 takes a fundamentally different approach. It's a fully end-to-end AI architecture that processes visual inputs directly into vehicle actions. There's no intermediate language translation, no hand-coded rules, and no separate modules to integrate. The model learns from real driving data—over 100 million data clips processed, equivalent to approximately 65,000 human driving years.

What makes VLA 2.0 genuinely novel is its cross-domain capability. The same AI platform is designed to work across passenger cars, Robotaxis, humanoid robots, and flying vehicles. This is the "Physical AI" thesis in action: if you build a model that understands the physical world well enough to drive a car, the same understanding can be applied to a robot navigating a factory floor or a flying vehicle navigating airspace.

Real-World Performance

Series deliveries of VLA 2.0-equipped vehicles began in China in March 2026. Within the first month, XPeng reported:

• 25.87% reduction in driver interventions per 100 kilometers
• 35.96% decrease in takeovers on narrow roads
• 96.97% user activation rate—meaning almost everyone who has the system uses it regularly
• On complex urban roads, the average intervention mileage improved by a factor of 13x compared to the previous generation

The 96.97% activation rate is particularly telling. Many advanced driver-assistance systems (ADAS) have high feature availability but low actual usage—drivers don't trust them enough to turn them on. A near-97% activation rate suggests that XPeng users genuinely find the system useful and trustworthy in daily driving.

The Turing AI Chip: XPeng's Silicon Bet

Building an AI-native car company requires AI-native hardware. XPeng developed its own AI chip—called Turing—specifically optimized for VLA 2.0 workloads. The Ultra version of the new L03 uses three Turing chips delivering a combined 2,250 TOPS (trillions of operations per second) of computing power.

This is a significant departure from the industry norm. Most automakers buy off-the-shelf chips from NVIDIA (Orin, Thor) or Qualcomm (Snapdragon Ride). XPeng's decision to design its own chip mirrors Tesla's approach with its FSD (Full Self-Driving) chip—vertical integration from silicon to software. The advantage is optimization: a chip designed specifically for VLA 2.0 can run the model more efficiently than a general-purpose AI accelerator.

However, it's worth noting that this approach also carries risks. Custom chip development is expensive and time-consuming. If VLA 2.0's architecture changes significantly, the chip may need to be redesigned. And for XPeng's lower-cost models, the company still relies on NVIDIA's Orin X chip, which provides a more cost-effective path to L2+ autonomy.

The XPeng Product Line: AI for Every Price Point

Going Global: The Munich Moment

XPeng's July 2026 Brand Day event in Munich was a watershed moment. It was the company's first vehicle debut to take place simultaneously across 65 markets. The L03, XPeng's most global model to date, was unveiled with a clear message: XPeng is not just exporting cars—it's bringing its full AI stack to global markets.

The Munich test drives were particularly significant. He Xiaopeng and Liu Xianming, head of XPeng's General Intelligence Center, demonstrated NGP (VLA 2.0) in real European traffic conditions. The system handled:

• Narrow old-town streets and complex intersections
• Construction zones with temporary signage
• Roundabouts with multiple entry and exit points
• Early lane changes to avoid parked vehicles
• Pedestrians, cyclists, and e-scooter riders—especially vulnerable road users
• European traffic signs and right-before-left priority rules

These aren't just technical achievements—they're regulatory prerequisites. For XPeng to sell autonomous driving capabilities in Europe, it needs to prove the system works under European traffic rules. The Munich tests demonstrated exactly that.

Google Maps Partnership

At the same event, XPeng announced a major partnership with Google Maps. XPeng becomes the first APAC automaker to ship a vehicle with Google Maps Auto SDK integration. This means XPeng can build its own navigation system using Google Maps data and services—critical for NGP's global operation. The system will also use Google Maps in-vehicle Map Data Services to support both NGP (VLA 2.0) and XPILOT Assist, the foundational assist system.

This partnership addresses one of the biggest challenges for Chinese automakers going global: navigation data. Autonomous driving systems need high-quality maps, and Google Maps is the default in most non-China markets. By integrating the SDK rather than relying on smartphone screen mirroring, XPeng creates a seamless, native experience.

The Regulatory Tailwind

XPeng's global autonomous driving plans are supported by two regulatory developments at the United Nations Economic Commission for Europe (UNECE/WP.29):

• UNR 171 Series 02 (DCAS): Creates the regulatory framework for urban NGP functions. Expected to become part of European law by the end of 2026.
• UNR ADS: Covers automated driving levels L3 to L5, expected to significantly accelerate approval of autonomous mobility solutions worldwide, especially L4 Robotaxis.

These regulations provide a clear legal pathway for XPeng to deploy its full NGP capabilities in Europe and other international markets. The company plans to begin the international rollout of NGP (VLA 2.0) in early 2027, subject to regulatory approvals. The introduction will happen gradually via over-the-air updates.

Beyond Cars: The Physical AI Ecosystem

XPeng's ambitions extend well beyond vehicles. The company is building a comprehensive Physical AI ecosystem:

Robotaxis

The first production Robotaxi based on NGP (VLA 2.0) is already undergoing internal road testing in China. The XPeng GX serves as the prototype platform, and XPeng plans to begin operational testing of its Robotaxi platform while exploring partnerships with global mobility service providers.

Humanoid Robot: IRON

XPeng plans to introduce its humanoid robot, IRON, to global markets from 2027. The robot runs on the same VLA 2.0 architecture that powers the cars—applying the same physical-world understanding to a different form factor. This is the cross-domain capability that XPeng sees as its ultimate competitive advantage.

Flying Vehicles

XPeng's flying car division continues development, though specific timelines were not updated at the Munich event. The underlying AI platform is designed to support aerial navigation as well as ground-based driving.

Challenges and Risks

XPeng's ambitious vision faces several significant challenges:

1. Competition

BYD, Tesla, NIO, Li Auto, Huawei-backed AITO, and Xiaomi EV are all investing heavily in autonomous driving. XPeng's technology lead is real, but competitors are closing the gap. BYD's scale advantage (millions of vehicles on the road generating training data) is a particular threat.

2. Regulatory Uncertainty

While UN regulations are moving in XPeng's favor, the timeline for L4 approval in major markets remains uncertain. A delay in regulatory approval would directly impact XPeng's revenue projections for its autonomous driving features.

3. Hardware Dependency

XPeng's custom Turing chip is a strategic asset, but the company still relies on NVIDIA's Orin X for most of its current lineup. If US-China chip restrictions tighten further, XPeng's access to NVIDIA chips could be affected—exactly the scenario that makes the Turing chip development so strategically important.

4. Profitability

XPeng has historically struggled with profitability. The heavy R&D investment required for custom chips, AI models, and global expansion puts pressure on margins. The company's ability to achieve sustainable profitability while maintaining its technology investment pace is not yet proven.

Conclusion: A High-Stakes Bet on Physical AI

XPeng is making one of the most ambitious bets in the automotive industry: that the future of mobility belongs to companies that control the full AI stack—from silicon to software to the physical vehicles themselves. VLA 2.0, the Turing chip, and the cross-domain Physical AI strategy represent a coherent vision that few competitors can match in scope.

The Munich event demonstrated that this vision is not just a PowerPoint presentation. The L03 driving through real European traffic, the Google Maps partnership, and the 65-market global launch plan all point to a company that is executing on its strategy, not just talking about it.

Whether XPeng can sustain this pace—and whether the market will reward a "Physical AI company" valuation over a traditional automaker valuation—remains to be seen. But one thing is clear: XPeng is no longer just another Chinese EV startup. It's attempting to redefine what a car company can be, and the global auto industry is watching.