He Xiaopeng burns 300 million a month—can it really make ordinary people dare to use autonomous driving?
```
On the evening of March 16, He Xiaopeng unusually revealed the truth.
Xpeng’s second-generationVLA was defined by He Xiaopeng as“the first version built forL4 capabilities”, with the central logic being to solve problems entirely withAI rather than rules. During the Two Sessions, representatives have been intensively calling to accelerate revisions to the "Road Traffic Safety Law".
To feed the second-generationVLA with sufficient computing power and data, Xpeng invests 300 million yuan purely each month, and has been burning cash like this for over a dozen months consecutively.
After merging the business lines of autonomous driving and intelligent cockpit, the General Intelligence Center led by Liu Xianming has a clear task: Turn this money into users’ trust. Terminal stores, thanks to intelligent driving, are indeed seeing a rebound in test-drive customer traffic.
But just then, a video of“four children lying across the road” went viral online, pushing the boundaries of intelligent driving capabilities into the spotlight. The system recognized the anomaly and slowed down to avoid, but ultimately needed manual intervention to fully stop.
The focus of the intelligent driving competition has shifted: it's not about parameter superiority, but about letting the public understand where the machine’s safety baseline lies and when humans should take over.
Intelligent Driving's Response and Boundaries
In the livestream, they discussed a recent viral video: Several children lying in the middle of the road, a Xpeng car with intelligent driving detected the anomaly and slowed down, then the driver intervened to brake.
Liu Xianming accessed backend data during the livestream and gave an interpretation. Data shows the vehicle did indeed detect the obstacle and triggered deceleration, but he also pointed out the deceleration was not enough for the car to fully stop.
He Xiaopeng added that the driver was an internal Xpeng employee testing the system, and noticed the situation on the road because the vehicle decelerated abnormally, thus took action to brake.
The system identified the problem, but in the end, people still had to make the decision. This is the current reality of intelligent driving: it can provide perception and warning, but extreme scenarios still require humans to decide.
Different technical paths have different response strategies—some focus on algorithm generalization, some on sensor redundancy, but all face the challenge of long-tail scenarios.
The key to getting ordinary people to use it is trust. The above case is precisely about building that trust.
In the livestream, He Xiaopeng again raised the term"national intelligent driving", defining it as "intelligent driving that moms love to use", emphasizing the system should return to safety and ease of use. He mentioned the company invited employees without technical backgrounds to experience it, and acceptance has significantly improved.
The data is more intuitive: second-generationVLA reduced hard braking by 99% and rapid acceleration by 98%. No more abrupt braking or fast acceleration, greatly reducing passenger discomfort—this is the key step for intelligent driving from "usable" to "willing to use".
After removing high-precision maps, the second-generationVLA can cover more unstructured roads. But Liu Xianming admits the current version occasionally does not fully follow navigation instructions.
He explained this as a necessary stage in the transition from rule-driven to inference-driven—the system is trying to understand the environment and make decisions autonomously, requiring more time for users to adapt in certain scenarios.
He Xiaopeng also drew a line: in extreme weather, or when there’s no road to follow—scenarios even human drivers have trouble with—it’s not recommended to use intelligent driving.
This is also industry consensus: the system is an auxiliary tool, not a replacement.
The competition in intelligent driving is shifting from parameters to trust.
In the past, they competed on the breadth of functional coverage. Now, it’s about who can make users truly dare to use and willing to use.
300 million invested monthly, returns are coming
"Spending 300 million each month to bet on this, for over a dozen months, at that time I was really nervous," He Xiaopeng admitted unusually in the livestream. Where is this "bet" placed? Liu Xianming's answer: From chips, compilers to software architecture and data closed-loop, all stack self-developed. "The main thing is to dare to bet."
Why dare to bet this way? He Xiaopeng raised a highly controversial view—China’s intelligent driving should leap directly from L2 to L4. If we stay at L3, it’s easy to lose in global competition.
The core logic of this leap is responsibility allocation. Liu Xianming explained,L4 requires the system to hard solve all problems and not throw difficulties to users. He Xiaopeng revealed that the second-generation VLA is precisely the first version built for L4 capability, with the core logic being to solve everything using AI, not rules.
The market is giving positive feedback.
He Xiaopeng revealed that sinceMarch 11, the second-generation VLA test drive was opened to 732 stores nationwide, the test drive rate doubled, and many users came specifically to experience the system. The more direct return is in the order structure: the Ultra version with the system has seen a significant sales proportion increase.
For rollout rhythm,gradual rollout started March 19, prioritizing Xpeng P7 Ultra, then G7 and X9 Ultra, with users of these three models receiving the update within the month. More models will be rolled out in April.
Addressing concerns about version differences, He Xiaopeng offered clear definitions:Ultra is for L4 capability in all-scenario traffic; Max mainly covers high-frequency scenarios like highways and main city roads.
Beyond immediate commercial returns, Xpeng's goal is global.In 2025, Xpeng's autonomous driving team has set a flag to compete with Tesla.
Recently, the media has been comparing the second-generationVLA to Tesla’s FSD V13, and He Xiaopeng gave his judgment.
"From V13, we clearlyhave the advantage. But I think it’s because Xpeng is in China, the data is in China, and we're more familiar with Chinese road conditions." He emphasized that the second-generation VLA is better at "person-vehicle game",like handling delivery guys, pedestrians, and narrow roads."This isn’t just a China feature; Europe has many narrow roads, Southeast Asia too. As we enter more countries, Xpeng may have more advantages."
Liu Xianming is more cautious:"Actually we don’t know how Tesla does it. It's more like crossing the river by feeling the stones,we stepped intoa lot of pits, and wasted a lot of money. But we believe, ultimately, the solutions may converge."
He Xiaopeng believes China and the US are both in the first tier for intelligent driving, but China’s roads are ten times more complex than America’s—not only highways and city roads, but also rural lanes in third- and fourth-tier cities, where you might run into cows, sheep, and chickens.
"Autonomous driving is a comprehensive competition of hardware, software, engineering, and scale capabilities. Currently, China and the US are both in the first tier. But China’s roads are more complex—only by tackling these tough problems and improving AI model generalization can the second-generation VLA truly globalize," He Xiaopeng said.
Every technology mustbe tested by users.The industry's judgment now is that the turning point for autonomous driving has arrived, but in practice, it still depends on whether users who've tried it are willing to recommend it to more people.
Risk Disclosure and DisclaimerThe market bears risks and investment requires caution. This article does not constitute personal investment advice, nor does it consider individual users' unique investment objectives, financial situations, or needs. Users should consider whether any opinions, viewpoints, or conclusions in this article fit their own situation. Invest accordingly, and at your own risk. ```