XPeng Abandons Multi-Billion ADAS for Unified VLA Foundation Model in Physical AI Pivot
Summary
Key Takeaways
He Xiaopeng revealed XPeng's decade-long ADAS development resulted in a 'Stitch Monster' combining software rules and AI algorithms, incapable of true autonomous driving. Last year, the company pivoted to a second-generation VLA (Vision-Language-Action) foundation model, abandoning the multi-billion RMB legacy system. The new model's upper limit reaches 100,000+ capability points vs the old 1,000, though its lower bound was initially worse. XPeng now positions as a physical AI company, investing in humanoid robots. Data costs exceed 1 billion RMB annually, with training runs processing dozens to hundreds of terabytes. Compute uses tens of thousands of NVIDIA H100 GPUs.
Why It Matters
XPeng's pivot is a defensive move against Tesla's end-to-end approach and Waymo's rule-based systems. By betting on a unified VLA model, XPeng aims to leapfrog competitors but hides the massive compute cost and initial instability. The new model may suffer from high inference latency in real-time driving scenarios, and the training data management cost is astronomical (over 1B RMB/year). The abandonment of a proven system risks short-term quality degradation.
PRO Decision
【Vendors】Tesla and Waymo should highlight XPeng's initial lower bound instability and showcase their own systems' reliability in edge cases. They can also exploit XPeng's resource diversion to strengthen their own autonomous driving deployments.
【Enterprises】Fleet operators should demand independent benchmarks of XPeng's VLA model in challenging scenarios like unknown parking lots and bad weather. Audit potential cost pass-through for data management.
【Investors】Recognize this as a high-risk, high-reward bet. Monitor training cost trends and the timeline for lower bound convergence. The pivot to physical AI may dilute automotive focus.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)