Gasgoo Munich-In the second half of the smart-driving race, "technological leadership" is often synonymous with burning cash. For automakers, turning that tech into revenue is the ultimate dream.
On February 26, XPENG Chairman He Xiaopeng dropped a hint on social media: Volkswagen is set to be the launch customer for XPENG’s second-generation VLA (Vision-Language-Action) model. The move marks more than just the start of mass production and external supply for the new system. It sends a clear signal to the industry: XPENG’s smart-driving ecosystem has evolved from a money pit into a revenue generator.
VLA 2.0 Is Coming
XPENG’s roadmap shows a steady march toward the essence of autonomous driving. It started with rule-based systems, moved to mass-produced end-to-end models, and has now arrived at a new paradigm: the second-generation VLA.
XPENG’s latest iteration in smart driving is the second-generation VLA. Official disclosures suggest the system will go into mass production in the first quarter of 2026. The first models to feature it will include the 2026 XPENG P7+, G7, and X9.
Architecturally, the biggest shift in the second-generation VLA is a full end-to-end connection. Traditional systems typically rely on a three-step logic: visual input, language translation, and action output. Visual signals must first be translated into machine code before driving commands can be generated.
Image Credit: He Xiaopeng's Weibo
Conventional VLA models follow a "Vision (V) – Language (L) – Action (A)" logic, converting visual data into language instructions before triggering action. While this boosts comprehension, the intermediate translation step drains computing power and risks information loss.
XPENG’s second-generation VLA skips that translation step entirely, moving directly from visual signals to action commands. After extensive testing, He Xiaopeng remarked that the system possesses the intuition of a seasoned driver — a qualitative leap born of architectural disruption.
By enabling end-to-end output from vision to action (V-A), XPENG eliminates the redundancy of language translation. The result is a more agile system that reacts much like a human driver’s instinct.
Powering the system is XPENG’s in-house Turing AI chip. A single chip delivers 2,250 TOPS of effective computing power, and with multiple chips onboard, model inference efficiency jumps 12 times. The model’s parameter scale is 10 times that of mainstream solutions, trained on nearly 100 million driving video clips that cover a vast range of long-tail scenarios globally.
XPENG estimates this data equates to 65,000 years of human driving experience.
Functionally, the second-generation VLA brings key upgrades. "Small Road NGP," for instance, enables assisted driving on narrow country lanes or old city alleyways, boosting the average distance between driver interventions by 13 times.
XPENG defines the system as the "first mass-produced physical world large model." It acts as both an action generator and a model for understanding and simulating the physical world. Crucially, the architecture isn't limited to cars; it can be adapted for humanoid robots, flying cars, and other physical terminals in the future.
In short, the rollout of VLA 2.0 is more than just a smart-driving upgrade — it is a cornerstone of XPENG's strategy in "embodied intelligence."
VLA Starts to Pay Off
Technical breakthroughs matter, of course. But what's really turning heads in the industry is that XPENG has started turning this technology into profit.
The market used to worry that XPENG was spending too much on R&D. Now that the second-generation VLA has entered Volkswagen's supply chain, the company's profit logic is shifting. Beyond selling vehicles, technology licensing is becoming a key revenue pillar.
Volkswagen's role as the launch customer means XPENG's smart-driving tech has cracked the procurement system of a global top-tier automaker for the first time. This isn't a simple badge-engineering job or a parts supply deal; it involves deep-level technology licensing and joint development.
For a giant like Volkswagen, developing software in-house is slow and expensive. Sourcing XPENG’s mature, architecturally advanced VLA model offers a fast track to intelligent transformation. For XPENG, this is more than just a big order. It proves its smart-driving technology can be exported as a "standard product," marking a leap from competing on vehicles alone to building a technology ecosystem.
Image Credit: XPENG
Compared with selling cars, technology licensing boasts high margins and low marginal costs. By open-sourcing tools and interfaces for its VLA architecture, XPENG is building an open smart-driving ecosystem.
As more third-party automakers, suppliers, and even robotics firms (like XPENG Robotics) adopt its base model, XPENG stands to gain long-term revenue from software subscriptions, technical support, and data loops. This virtuous cycle — the bigger the ecosystem, the stronger the model; the stronger the model, the higher the revenue — is reshaping XPENG’s financial structure.
Admittedly, this revenue stream still relies heavily on Volkswagen, posing a risk of customer concentration. But XPENG is looking to replicate its technology export model. In 2025, the company initiated talks with several overseas automakers about smart-driving tech, and licensing the VLA architecture to more partners is a distinct possibility.
Meanwhile, XPENG is accelerating the commercialization of new ventures like Robotaxis, humanoid robots, and flying cars. All rely on the Turing chip and VLA architecture, promising new growth engines for revenue in the coming years.
