How much courage does it take to scrap a technology system that took years and billions of dollars to build? He Xiaopeng, founder and CEO of XPeng Group, detailed a little-known internal technological revolution that took place in the third and fourth quarters of 2025 in a podcast interview aired on May 28th. He stated bluntly in the program: "A CEO in the physical world wouldn't dare to make this bet; perhaps I'm bolder. We made a huge gamble. We shut down the old system, a system that cost us several billion yuan." The backdrop for this high-stakes bet was XPeng's announcement in 2025 that it was changing its name to "XPeng Group" and repositioning itself as a "Physical AI" company. Behind the name change was a complete overhaul, from the underlying technology architecture to the organizational structure.
Why scrap the old system? He Xiaopeng referred to the old system as an "AI Frankenstein." He explained that the end-to-end autonomous driving approach, which the entire industry had championed, including XPeng's own efforts, was essentially a product cobbled together using "software engineering, software processes, plus AI algorithms and AI tools." "You're using a software methodology with an AI toolbox to create something, resulting in a more powerful piece of software. I call it an AI Frankenstein," he said. He used a straightforward analogy: "It's like building a house. You might have more materials, more techniques, using AI, but you're still using the old methodology for building a house. What you build is still the same old house, just maybe built a bit faster." The fatal flaw of this system was its low ceiling. He Xiaopeng said the old system's capability ceiling was around 1,000 points, and its shortcomings would be immediately exposed in scenarios like underground parking lots or complex roads. His conclusion was: "I believe it cannot achieve true autonomous driving. I believe it also cannot enable true generalization for robots." This judgment gradually took shape in his mind around early 2025. He said that at the time, he saw many companies in the industry using "simpler methodologies and faster efficiencies" to produce "barely usable autonomous driving." However, "after one or two years, they realize that in order to solve many shortcomings, they limit their own strengths, so they find they can never reach Level 4 or Level 5." His assessment was: "This path is a shortcut, but not a main road; it's a side path."
The new path: A high ceiling, but a low floor In 2025, XPeng was simultaneously developing two new generations of systems, internally referred to as "VAA Gen 1" and "VAA Gen 2." The first generation was an incremental improvement on the existing end-to-end framework, scaling up the model size and reducing the proportion of software rules. The second generation was a completely different path—abandoning the end-to-end logic, driving everything with a larger Foundation Model, and reconstructing the entire autonomous driving system in an AI-native way, known as the VLA system. He Xiaopeng described the gap between them using scores: the old system had a ceiling of 1,000 points and a floor of 900 points, which was "pretty good capability." The new system had a potential ceiling of 100,000 to 1,000,000 points, but its initial floor was only 100 points—"riddled with engineering bugs"—lower than the floor of the old system. "You originally expect to make a product with a ceiling of 1,000 points and a floor of 900 points; that's pretty good capability. But at that time, the floor might have been only 100 points, with a lot of engineering problems," he explained. Faced with this situation, he made what he considered a "very huge bet": firmly betting on the second generation while simultaneously halting the first generation and even the entire old system. He said: "When I decided that the second-generation VAA should be scaled up, and finally firmly decided to even completely stop the first generation, from business to technology, even to organization, that created enormous psychological and material pressure."
Internal resistance: Some "voted with their feet" and left This decision was not smooth sailing within the company. He Xiaopeng revealed that at the time, most non-AI background managers believed that "whether you do A or B, it's probably wrong." Among AI-related managers, some also took a wait-and-see attitude, "not sure if this could achieve a practical effect." The strongest opposition was not voiced but was a silent departure. "The biggest opposing voice was that many people voted with their feet... They didn't believe this could be done, they thought it was impossible, so they left." Regarding this, He Xiaopeng's attitude was: once you've thought it through, you must make a clean cut. "Remember, don't use a small knife to chop down a big tree, chopping slowly. Once you've thought it through, cut it down. Change everything from the organization to the processes, to the direction." By the end of the third quarter of 2025, XPeng had completely restructured the core organizational framework of its entire autonomous driving center. He explained the reason: "Everyone has their old habits; you're used to using past methods, wanting to use the latest tools and technology to create better things... But many times, your entire work methodology has to change."
Why did he dare to bet when others didn't? In the interview, He Xiaopeng mentioned a dilemma he believes CEOs in the physical world commonly face: unlike pure digital AI companies, those who make cars have too many "planks" to manage—quality, cost, regulations, engineering, supply chain. Having one outstanding long plank is useless; a short plank can directly drag down the whole. "A CEO in the physical world either doesn't dare to bet, or feels they have so many planks, what should they do? I think I might be bolder," he said. He admitted he wasn't sure at the time if the new system could fix the shortcomings but placed the bet anyway. "We made a huge gamble. We just felt, I believe in Physical AI. Changing the mindset, changing the process, fundamentally changing the organizational approach might achieve something different." He also cited two principles he set for himself during a low point in 2022 to explain this decision-making logic—"Never admit defeat" and "Accept the outcome of your bet." He said these are not contradictory but a way to maintain momentum under high pressure.
The logic behind the bet: Software value should reach 50% He Xiaopeng's bet on the new system is underpinned by a longer-term business judgment. He believes that in the past era of smart electric vehicles, hardware accounted for far more than 50% of the value users needed. But in the next decade, if AI-native methods can be used to re-drive the software and hardware system, the comprehensive value of software could rise to 50%. He gave a specific example: "Would users be willing to say, I'm buying a 300,000 yuan car, with 150,000 yuan for your hardware and 150,000 yuan for your comprehensive software capabilities? Is there such a car? Not now. But I think it will definitely change within this decade." This is also the core logic behind XPeng's name change from an "AI car company" to a "Physical AI company"—no longer thinking about intelligence in terms of electric or automotive methods, but "changing the logic," adjusting everything from strategic planning to the R&D process. Regarding the outcome of this bet, He Xiaopeng admitted it is still being validated and acknowledged that at the time he "didn't know if it could solve the shortcomings." But his judgment is: the more you hesitate, the more you wait, the lower the probability of success. "The more you hesitate, the more you wait, the more you want to observe, the more you want to say 'let's wait another six months,' the harder it might be to succeed."
What are the results? Preliminary data is emerging He Xiaopeng cited a set of recent data in the interview: at the end of March 2025, XPeng released the first version of the second-generation VAA. In April, overall Chinese car sales fell by about 20% year-on-year and month-on-month, while XPeng's sales in that month increased by about 50% to 70%. "A significant portion of this is related to the second-generation VAA." He stated that whether this path ultimately succeeds still requires time to verify, but he has not wavered in the direction. "I think in today's process, first, each company is walking its own path; it cannot be replicated."
Comments