VW First Customer for XPeng Gen2 VLA

🔑 Enhanced Key Takeaways

• •XPeng's VLA 2.0 was trained on nearly 100 million video clips, equivalent to 65,000 years of extreme driving scenarios, enabling annotation-free training from real driving videos.[1][5][6]
• •VLA 2.0 runs on three self-developed Turing chips providing 2,250 TOPS peak compute, with mass production of the chip starting in Q2 2026.[4][5][7]
• •Volkswagen also nominated XPeng's Turing AI Chip alongside adopting VLA 2.0, marking validation from a major Western carmaker for Chinese AD tech.[3][4]
• •XPeng demonstrated XNGP to global UN/WP.29 delegates in Shanghai, showcasing real-world urban and highway performance to regulators from EU, US, Japan, and others.[2][3]

🛠️ Technical Deep Dive

• •VLA 2.0 eliminates language translation by directly mapping visual inputs to continuous control actions (steering, acceleration), improving precision in complex scenarios and enabling self-supervision without annotations.[1]
• •Powered by XNet (perception NN), XPlanner (planning/control large model), and XBrain (AI LLM), building on XNGP+ 'pre-end-to-end' architecture.[1]
• •Onboard compute via three Turing chips at 2,250 TOPS; trained on 30,000-GPU Alibaba Cloud cluster with 72-billion-parameter model, iterating full-stack every five days.[5]
• •Integrated world simulator evaluates decision quality by generating and scoring alternative driving strategies from VLA data.[1]
• •New capabilities include hand gesture recognition, proactive traffic light response, and 'Narrow Road NGP' boosting takeover mileage 13x on complex roads.[6]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline