He Xiaopeng's live "confession": spending 300 million yuan a month, betting on the second-generation VLA, we are also anxious inside

Summary:

How big is this “bet”? He Xiaopeng provided a figure: 300 million yuan per month. For more than a dozen consecutive months. The investment targets range from chips and compilers to software architecture and data closed-loop full-stack self-developed systems.

On the evening of March 16, He Xiaopeng, Chairman of XPeng Motors, and Liu Xianming, head of XPeng’s General Intelligent Center, sat in the live broadcast room. This “Ask Me Anything” session was supposed to be a routine technical communication, but due to He Xiaopeng’s frequent self-disclosures, it evolved into a candid discussion about technical routes and business choices.

“Spending 300 million yuan a month to gamble on this, for more than a dozen months in a row, by that time I was really panicked.”

The “bet” He Xiaopeng mentioned is the upcoming launch of XPeng’s second-generation VLA (Visual-Language-Action) autonomous driving system on March 19. To build this new architecture that completely abandons rules and reconstructs driving logic with AI, XPeng has chosen “the only path” over the past year.

300 million yuan monthly investment, a bold gamble on “security”

“Many companies probably pursue two paths simultaneously, but we ultimately chose one. We are confidently walking one path,” He Xiaopeng admitted during the live broadcast, emphasizing that technical intuition and courage are both indispensable. Liu Xianming added, “Mainly, you also have to dare to bet.”

How big is this “bet”? He Xiaopeng gave a figure: 300 million yuan per month. For more than a dozen months. The investment covers everything from chips and compilers to software architecture and data closed-loop full-stack self-research systems. Liu Xianming explained, “We are developing from scratch across all dimensions—Turing chips, Turing model structures, compilers, the underlying software architecture, plus a large amount of data and AI infrastructure. It’s not accidental; it’s a complete system advancing forward.”

The direct product of this system is the upcoming second-generation VLA. Liu Xianming revealed that after deciding on mass production, the team underwent intensive iterations: media test drives and store versions are “roughly the second major version, with over a dozen minor branches,” while the upcoming mass-produced version is “the fourth major version, with the 28th minor version.” The iteration speed is fast, with each version rigorously tested, but one feature was temporarily halted—park and underground garage roaming.

“We believe it hasn’t reached the stunning level of urban driving, so for the user experience, we won’t push it for now. The next version will catch up,” Liu Xianming said. He Xiaopeng offered a more emotional explanation: “I hoped to release the park feature in the first version last year, but if the highway and urban driving score 90, the park only gets 60, users would find it hard to accept. We want to reach 85 or 90 before pushing.”

A video was played during the live broadcast that instantly froze the atmosphere: on a two-way, two-lane road, four children are lying in the road. The second-generation VLA recognized and slowed down in advance. Liu Xianming’s reaction was unexpected—he didn’t show off but felt “scared.”

“We looked at the backend data for this case. The second-generation VLA slowed down, but the deceleration was not enough to stop,” Liu Xianming said. “This also motivated us—there’s still a gap to perfecting it. We want it to be safe enough in any emergency.”

He Xiaopeng added more concerning details: the driver was a team member who noticed the car “strangely decelerating,” looked at the road, realized there were children, and then pressed the brake. “If he hadn’t paid attention, he might have run over the ‘speed bump,’ with unimaginable consequences,” He Xiaopeng said. “Why do we develop autonomous driving? It’s not just about freeing hands; it’s about greatly improving safety. Nearly 200,000 people die annually in traffic accidents worldwide. If technology can save those lives, it’s worth everything.”

This case also underscores XPeng’s emphasis on “peace of mind.” He Xiaopeng said that when he took his mother for a test drive before, she was so nervous she held the handrail tightly. Now, the goal of the second-generation VLA is “to make even moms willing and eager to drive.” They invited “canteen aunties” to experience it, and the moms went from “not daring” to “daring” in just 15 minutes.

14 seconds of “thinking” and generalization ability

Another impressive scene involved media worker Dayu: navigating a road section under construction with no updated map, the second-generation VLA paused for 14 seconds before turning left and rerouting, then updated the navigation route. This 14-second pause sparked online discussion—does the car “think”?

Liu Xianming explained, “We don’t know exactly what it did during those 14 seconds, but we believe it was reasoning, trying to generate different paths and evaluate feasibility.” He Xiaopeng immediately challenged, “I criticize myself—why wait 14 seconds? I think it should be 2 seconds. That shows we’re not yet good enough, although it’s true that it generalizes and doesn’t have to follow navigation rules, but I want it to be faster.”

This “2 seconds” became another OKR for Liu Xianming on the spot. He Xiaopeng also added multiple goals: increase takeover rate from 5 times to 5-10 times, achieve one takeover per thousand or even ten thousand kilometers; solve fully autonomous parking in unfamiliar parking lots; and even dream of achieving “off-road autonomous driving” in 2027 or 2028—driving in places without roads.

Regarding recent media comparisons between the second-generation VLA and Tesla FSD V13, He Xiaopeng gave an objective assessment: “From V13, we are clearly ahead. But I think it’s because XPeng is in China, and the data is in China, so we are more familiar with Chinese road conditions. We look forward to V14 entering China or us entering Europe, where the comparison will be more meaningful on the same complex routes.”

He emphasized that the second-generation VLA handles “human-vehicle interactions” better—such as delivery workers, pedestrians, narrow roads. “This isn’t just a Chinese feature; Europe has many narrow streets, Southeast Asia too. As we expand into more countries, XPeng may have an advantage.”

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

Skipping L3 directly to L4?

During the live broadcast, He Xiaopeng proposed that we should jump directly from L2 to L4, skipping L3. He explained that in the past, code was written based on rules, requiring hundreds of thousands of rules, which needed incremental development; but today’s AI large models are dynamic reasoning systems. “If we still rely on L3, we risk falling behind in global competition.”

Liu Xianming added from a technical perspective, “L4 requires solving all problems without passing difficult issues to the user. The ultimate goal is to make the model’s generalization ability strong enough to operate comfortably, confidently, and safely in all road conditions.”

To achieve this, XPeng is promoting cross-domain integration: combining chassis, powertrain, autonomous driving, and cockpit into a unified intelligent system to control the entire vehicle. “Ultimately, it’s more like a super-intelligent agent, a robot,” He Xiaopeng said. The same VLA technology is also being tested on robots, “and you’ll soon see XPeng’s robot capable of very flexible actions—sitting, standing, jumping, running, jumping over hurdles, climbing stairs.”

On March 19, the P7 Ultra was launched first; followed by G7 and X9 within a few days—all completed within March; other models (P7+, G9, G6, etc.) will start Ultra/SE versions in April. The Max version will initially feature the Turing chip, with dual Orin Max versions coming in the second half of the year.

“Only after trying it will you have a new understanding of the future,” He Xiaopeng said. “Come to the store for a 15-minute test drive, and you’ll believe that autonomous driving is coming in the next few years, and robots will truly enter homes, changing our lives.”

The two-hour live broadcast ended with a blend of He Xiaopeng’s emotional insights and Liu Xianming’s rational analysis.

And perhaps all the owners waiting for the update will recall He Xiaopeng’s confession: “We’ve placed a big bet, but it’s a bet on the future.”