Does He Xiaopeng Burn 300 Million Per Month? Can He Really Make Autonomous Driving Accessible to Ordinary People?

On the evening of March 16, He Xiaopeng rarely revealed the details.

Xpeng’s second-generation VLA, defined by He Xiaopeng as “the first version built for L4 capability,” centers on solving problems entirely with AI, not rules. During the Two Sessions, delegates also called for accelerating the revision of the Road Traffic Safety Law.

To supply enough computing power and data for the second-generation VLA, Xpeng invests 300 million yuan monthly, continuing this effort for over a year.

After merging the autonomous driving and intelligent cockpit business lines, Liu Xianming’s team at the General Intelligence Center has a clear mission: turn this investment into user trust. The test drive traffic at terminal stores, driven by interest in intelligent driving, is indeed picking up.

But at this moment, a video of “four children lying across the road” sparked online heated discussion, pushing the boundaries of autonomous driving capabilities into the spotlight. The system recognized the anomaly and slowed down to avoid, but ultimately manual intervention was needed to fully stop the vehicle.

The focus of the autonomous driving competition has shifted: it’s no longer about parameter dominance, but about helping the public understand where the safety bottom line of machines lies and when humans should take over.

Reactions and Boundaries of Autonomous Driving

During the live broadcast, a recent viral video was discussed: several children lying in the middle of the road. A Xpeng vehicle with autonomous driving activated recognized the anomaly and slowed down, then the driver intervened to brake.

Liu Xianming retrieved backend data and provided analysis. The data showed the vehicle did detect the obstacle and triggered deceleration, but he also pointed out that the deceleration was insufficient to fully stop the vehicle at that moment.

He Xiaopeng added that the driver was an internal Xpeng employee testing the system, who noticed the vehicle slowing down abnormally and then took action to brake.

The system detected the problem, but ultimately a human had to make the decision. This reflects the current state of autonomous driving: capable of perception and warning, but extreme scenarios still require human judgment.

Different technical approaches have different strategies—some focus on algorithm generalization, others on sensor redundancy—but all face challenges with long-tail scenarios.

To make ordinary people willing to drive, the key is building trust. This case exemplifies that process.

In the live broadcast, He Xiaopeng again proposed the concept of “National Autonomous Driving,” defining it as “autonomous driving that even moms love,” emphasizing that the system should return to safety and ease of use. He mentioned that after inviting employees without technical backgrounds to experience it, acceptance significantly improved.

Data is more intuitive: the second-generation VLA’s hard braking incidents decreased by 99%, sudden accelerations by 98%. No more abrupt braking or sudden acceleration, greatly reducing passenger discomfort—this is a crucial step for autonomous driving from “usable” to “willing to use.”

Without relying on high-precision maps, the second-generation VLA can cover more unstructured roads. However, Liu Xianming admitted that the current version occasionally does not follow navigation perfectly.

He explained this as a necessary phase in shifting from rule-based to reasoning-based systems—the system is trying to understand the environment and make autonomous decisions, but users need to give it more time to adapt in certain scenarios.

He Xiaopeng also set a boundary: in extreme weather, dead-end roads, or other situations that even human drivers find difficult, autonomous driving is not recommended.

This aligns with industry consensus: systems are assistive tools, not replacements.

The competition in autonomous driving is shifting from parameter dominance to trust.

In the past, the focus was on the breadth of functionality; now, it’s about who can truly make users dare and want to use.

Invest 300 Million Monthly, Returns Are Coming

“Invest 300 million yuan every month to gamble on this, for over a year. By then, I’d be pretty nervous,” He Xiaopeng admitted rarely during the live broadcast. Where is this “bet” placed? Liu Xianming responded: from chips, compilers, to software architecture and data loops—full-stack self-research and development. “Mainly, you have to dare to bet.”

Why dare to gamble like this? He Xiaopeng made a highly controversial judgment—that Chinese autonomous driving should leap directly from L2 to L4. Staying at L3 would likely lead to defeat in global competition.

The core of this leap is responsibility division. Liu Xianming explained that L4 requires the system to solve all problems without passing difficult issues to the user. He Xiaopeng revealed that the second-generation VLA is the first version designed for L4 capability, with the core logic of solving problems entirely through AI, not rules.

The market is also giving positive feedback.

He Xiaopeng disclosed that since March 11, when the second-generation VLA test drives were opened at 732 Xpeng stores nationwide, the test drive rate doubled, with many users coming specifically to experience this system. More directly, the order structure has shifted: the Ultra version equipped with this system has seen a significant increase in sales.

Regarding the rollout schedule, starting March 19, the system will be gradually pushed, prioritizing the Xpeng P7 Ultra, followed by G7 and X9 Ultra. Users of these models will receive updates successively within this month. More models will be pushed in April.

On the issue of version differences, He Xiaopeng clarified: the Ultra version is built for L4-level capabilities, supporting full-scenario driving; the Max version mainly covers highways and main urban roads.

Beyond immediate commercial returns, Xpeng’s goal is global. By 2025, the Xpeng autonomous driving team has set a goal to compete with Tesla.

Recently, media comparisons between the second-generation VLA and Tesla FSD V13 have intensified. He Xiaopeng shared his view.

“From V13, we are clearly ahead. But I think it’s because Xpeng is in China, with data in China, and more familiar with Chinese road conditions.” He emphasized that the second-generation VLA handles “human-vehicle interactions” better, such as delivery workers, pedestrians, narrow roads. “This is not just a Chinese characteristic; Europe has many narrow streets, Southeast Asia as well. As we expand into more countries, Xpeng may have an advantage.”

Liu Xianming was more cautious: “Honestly, we don’t know how Tesla is doing. It’s more like feeling our way across a river, stepping over many pitfalls, and wasting a lot of money. But we believe the ultimate solutions may converge.”

He Xiaopeng believes China and the US are both in the first echelon of autonomous driving, but China’s roads are ten times more complex—beyond highways and urban roads, there are rural paths outside third- and fourth-tier cities, where cows, sheep, and chickens can suddenly appear.

“Autonomous driving is a comprehensive contest of hardware, engineering capability, and scale. Currently, both China and the US are in the top tier. But China’s roads are more complex. Only by first solving these difficult problems and improving AI large model generalization can the second-generation VLA truly achieve global deployment,” He Xiaopeng said.

Any technology must be tested by users. The industry now believes that the autonomous driving inflection point has arrived, but the real test is whether users are willing to recommend it to others.