XPeng Skips L3 for L4 Autonomy in 1-3 Years

🔑 Enhanced Key Takeaways

• •XPeng's second-gen VLA utilizes 30,000 GPU hours for training, four self-developed Turing AI chips providing 3,000 TOPS compute power, and a 72-billion parameter model trained on 100 million video clips equivalent to 65,000 years of human driving experience.[1][3][4]
• •Three Robotaxi models planned for 2026 launch, equipped with VLA 2.0, no reliance on LiDAR or HD maps, dual redundancy in six systems, and capability for mass production without factory modifications.[3][4]
• •XPeng merged its autonomous driving and smart cockpit centers into a General AI Center led by Liu Xianming, focusing on foundation models, infrastructure, and cross-domain applications for vehicles and robots.[7]

🛠️ Technical Deep Dive

• •Four Turing AI chips per vehicle deliver 3,000 TOPS computing power, enabling pure vision-based L4 autonomy without LiDAR or high-definition maps.[3][4]
• •VLA 2.0 foundation model has 72 billion parameters, trained with over 2 billion yuan in costs using nearly 100 million global video clips covering extreme long-tail scenarios.[1]
• •System supports advanced capabilities like Narrow Road NGP, improving takeover mileage by 13x on complex roads, gesture recognition, in-place starting, pull-over parking, and speed bump handling.[1][4]
• •Deployed in Ultra/Ultra SE models and upcoming Robotaxi/Robo vehicles with dual redundancy across six systems for safety.[3][4]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline