Xpeng Opens Its Self-Driving Brain to the World, with Volkswagen as Its First Partner
Xpeng Opens Its Self-Driving Brain to the World, with Volkswagen as Its First Partner
China’s rising EV star aims to become the “Mobileye of China,” kicking off a new round in the global race for smarter cars.
GUANGZHOU — Chinese electric carmaker Xpeng just threw down the gauntlet in the global tech arena. On Wednesday, during its 2025 AI Day, CEO He Xiaopeng unveiled the company’s next-generation Vision-Language-Action self-driving system—and immediately opened it to automakers worldwide. The first to sign on? None other than Volkswagen.
That single move could rewrite how carmakers think about artificial intelligence. It’s the first time a Chinese manufacturer has licensed a full autonomous-driving system to a major Western brand. For Xpeng, it’s a shift from selling cars to selling brains—the software and silicon that make them think.
“Volkswagen is our first customer for the second-generation VLA, and they’ve officially certified our Turing AI chip,” He announced proudly. “This marks the biggest leap in intelligent driving since end-to-end neural networks arrived.”
Computing Power Becomes the New Horsepower
At the heart of Xpeng’s new system sits a 72-billion-parameter foundation model trained on nearly 100 million driving videos—enough to cover the equivalent of 65,000 years of human driving scenarios. Training happens on Alibaba Cloud’s massive 30,000-GPU supercomputer, completing a full cycle every five days. Few automakers can even dream of that pace.
But big data means little without real-world performance. Xpeng’s custom Turing chip delivers 2,250 trillion operations per second across three processors, tackling the long-standing problem of running massive AI models in real time inside a moving car.
The payoff shows up where drivers need it most: tricky “long-tail” moments. Think of those narrow residential lanes, chaotic intersections, or confused pedestrians that leave even human drivers on edge. Xpeng’s new “Narrow Road NGP” reportedly improves performance thirteen-fold in these situations, dramatically reducing how often a human has to take over. If outside tests confirm that, it’s a genuine milestone.
A Divide Over How Cars Should Think
Xpeng’s VLA approach, where vision and language work together to guide action, has split opinion across China’s fast-moving automotive AI world. Startups like Yuanrong Qihang and giants like Li Auto have embraced language-guided systems. Others, such as Huawei and Horizon Robotics, think that’s the wrong road to take.
Huawei’s engineers favor what they call “World-Model plus Action”—a direct line from visual input to vehicle control, without the language layer in between. Horizon, whose highway system often gets compared to Tesla’s Full Self-Driving, rejects VLA entirely. They argue that natural language just slows things down and muddies precision.
As one industry analyst put it, “Language is like an extra sensor—useful, powerful, but not magic.” Commands like “slow down a bit” may sound human, but they’re too vague for a computer deciding whether to tap the brakes or swerve.
Xpeng, however, isn’t choosing sides—it’s blending both. The company uses language for high-level reasoning and learning from rare edge cases while keeping traditional control systems for safety and precision. In practice, that means the car can interpret a delivery driver waving it through a cramped European alley but still relies on tried-and-true code to manage the exact steering and throttle.
Betting on a Platform, Not Just Cars
Volkswagen’s participation gives Xpeng’s grand pivot real credibility. The German automaker already brought in about RMB 1.72 billion for Xpeng through earlier tech partnerships in 2025. Adding VLA software and Turing chips deepens the relationship—creating multiple income streams, from licensing fees to per-vehicle software charges and cloud services.
The bigger picture is timing. Between 2026 and 2028, Western automakers will need ready-to-deploy AI systems to stay competitive in China, even as Chinese brands focus on their own vehicles. Volkswagen’s parallel work with Horizon Robotics and its Cariad software arm sets a ticking clock. Xpeng must prove indispensable before VW’s in-house AI matures within the next few years.
That’s where Xpeng’s narrow-road expertise could shine. Europe’s tight, medieval street grids and unpredictable pedestrians mirror the complex Chinese cities where Xpeng’s data was born. If the thirteen-fold improvement holds true overseas, automakers like Renault or Stellantis might find a ready-made fix for their lagging self-driving programs.
Still, there’s a caveat. Xpeng says its platform is “open to global commercial partners,” but that likely means source-available under strict terms, not true open source. So while others can integrate it, they won’t control it. The real proof will come in 2026 when Volkswagen’s China models hit the market. If those cars list “Powered by Xpeng VLA,” and another major automaker signs on, the bet will look brilliant. If not, that massive 30,000-GPU cluster could become a costly burden.
The Stakes Ahead
Optimists believe Xpeng could earn RMB 2–3 billion annually in high-margin tech revenue by 2027, reinventing itself as the “Mobileye of China.” Skeptics counter that Volkswagen might outgrow its dependence just as political tensions make Western firms cautious about Chinese AI. In that scenario, Xpeng’s grand model would serve only its own limited lineup—a heavy load for one brand to carry.
What’s undeniable is that Xpeng has crossed a line few dared to approach. It’s no longer just a carmaker; it’s now a platform company in one of the toughest, most lucrative games on Earth. Whether others follow Volkswagen’s lead will decide if China has built its first true global automotive software powerhouse—or just its boldest experiment yet.
NOT INVESTMENT ADVICE