XPeng is following the right strategy (vision only, custom made silicon, data collection from the fleet), but I highly doubt they will make it in 2026, at least not without HDMaps. Please remember that Tesla needs 50,000 H100 GPUs to train FSD, China does not have access to that. What makes you think they can do it without this? China is developing its own AI Cluster technology with Huawei, but until it is ready at scale for XPeng to use it, plus the time needed to actually train, debug, and real world test the models (remember, if Tesla scales Robotaxi next year, it will have taken 18 months between FSD13 and Robotaxis at scale!), 2026 looks completely unreasonable. I think 2028 is the best-case scenario.
I have always said the second OEM to crack FSD would be Chinese and probably XPeng, but 2026 sounds to me like a marketing scam (unless Lidar/HDMaps are used, but XPeng publicly said they would follow Tesla's vision-only strategy).