XPeng shares fell 5% on Friday, on track for their worst day since late April, after Xianming Liu, head of the company's General Intelligence Center, revealed the scale of its autonomy spending. Speaking to Electrek, Liu said his team trains AI models for self-driving cars at a cost of roughly RMB 300 million a month — about $500 million a year on model training alone. "I do spend a lot of money. As a car company, you can never imagine such a big R&D investment because you can never make the money back," Liu said, adding that CEO He Xiaopeng has committed to it because XPeng is determined to be a "Physical AI company."

XPEV stock has already fallen 22% this year on weak deliveries and slowing demand, and retail sentiment on Stocktwits slid from 'extremely bullish' to 'bearish' over the past week. The reveal also sharpened a philosophical split with Tesla: Liu said cameras are the best primary solution, but unlike Elon Musk he does not believe cameras alone will ever be sufficient. "Even you can reach 99.9999%, you still have the chance to make mistakes," Liu said, explaining why radar and lidar stay on XPeng cars as a redundant safety layer. "You're talking about lives."

Stocktwits reports that Liu made the remarks the day after sharing the CVPR 2026 stage in Denver with Tesla's Ashok Elluswamy and executives from Nvidia and Waymo — and claimed XPeng's latest vision-language-action models have already reached parity with, or even surpassed, Tesla's FSD V14 in early testing this year. XPeng follows nearly the same camera-first philosophy as Tesla, but keeps ultrasonic sensors and radars in reserve strictly for redundant active-safety functions like automatic emergency braking, creating what Liu called an "orthogonal system" independent of the main AI driver. Tesla's FSD system itself remains under naming scrutiny in Europe, where regulators have said the 'Full Self-Driving' label misled buyers into expecting unsupervised automation.

Liu argued that China's notoriously chaotic roads give XPeng an advantage in data diversity and corner-case collection compared with Tesla, which primarily trains FSD in the US, according to Stocktwits. More varied corner cases, he said, give XPeng more chances to expand internationally. The strategy is already monetized: Volkswagen announced a strategic partnership with XPeng in 2023 to license the Chinese EV maker's self-driving tech for its own China-market models, while Tesla has yet to find interested rival carmakers for FSD — a fact Musk has often expressed frustration about. XPeng also positions itself as a 'physical AI' company first and an automaker second, in Liu's telling.

Liu wants the same foundation models that drive XPeng's cars to generalize far beyond them — to robots and flying vehicles, where the company is also investing, Stocktwits reports. He likened the approach to Google's Pixel phones: devices built to demonstrate what the underlying AI can do. "We need physical devices in the real world to make sure we get feedback, we get data," Liu said. Meanwhile, the Tesla comparison keeps escalating on both fronts: an EU-wide vote on FSD approval has slipped from October to at least December, with countries like France conducting additional testing, per Reuters via The Indian Express, and eight EU countries already allow the system with Slovakia expected to become the ninth.