XPeng's He Xiaopeng Bets the Company on Physical AI, Scraps Billions in Software for Radical New Approach
Zhang Xiaojun Podcast Interview, May 28, 2026
He Xiaopeng made one of the riskiest bets in Chinese automotive history last year, and almost no one outside his inner circle knew about it. In April 2025, the XPeng Motors founder and CEO unilaterally decided to scrap the company's first-generation autonomous driving system—a platform that had already consumed billions of yuan in development costs—and pivot entirely to what he calls "physical AI." The decision triggered an internal exodus, with employees literally voting with their feet, but He pushed forward anyway. A year later, the gamble appears to be paying off, though he readily admits the odds of ultimate success remain just 20 percent.
The radical pivot emerged from He's growing conviction that the entire automotive industry was building what he derisively calls "AI stitch monsters"—systems that bolt AI tools onto traditional software architectures rather than reimagining vehicles from first principles as intelligent physical agents. By mid-2023, XPeng was running two parallel autonomous driving programs internally: a first-generation system that incrementally improved end-to-end capabilities within existing frameworks, and a second-generation approach using dramatically larger foundation models with fundamentally different thinking logic. The first generation offered a manageable upper limit of around 1,000 capability points with a lower limit near 900 points, He explained. The second generation promised an upper limit reaching 100,000 to one million points—but its lower limit sat catastrophically low at just 100 points.
The April 2025 Inflection Point
He describes the turning point as less a formal meeting than a realization that crystallized in his head during early 2025. He had spent months watching competitors pursue what seemed like faster paths to "barely usable" autonomous driving, only to hit inevitable walls when trying to solve edge cases. "I think all companies think with simpler methodologies and faster performance can make one that seems to be just about usable autonomous driving," he said. "But after one to two years of doing it, it will be found that in order to solve many shortcomings, it limits it to its own longboard. So it finds that it can never do level 4 or level 5."
The theoretical breakthrough came from recognizing a fundamental difference between digital and physical AI. Digital AI operates in human language, which is "highly summarized and condensed," He argued. Physical AI must process vastly more data that cannot be adequately captured in words. "The amount of data each of us sees every day, it is simply impossible to summarize, describe, restore and copy in words," he said. This distinction meant that methodologies from digital AI companies—which XPeng's teams had been attempting to adapt—were structurally inadequate for the physical world.
By late third quarter 2025, He executed a sweeping organizational restructure of XPeng's autonomous driving center, completely changing its core structure. The move was brutal: of a 300-person robotics team built using traditional methods, fewer than 60 remained. He acknowledged that "many outstanding classmates" departed, spawning roughly 10 startup teams that have since raised funding. But he judged this creative destruction necessary to escape what he saw as fatal path dependency. "You will get used to using the old methods, enjoy the latest tools and technology to make something better," he said. "I think most of the time it's right. But there are many times when your work methodology needs to be changed."
The Technology Wager Behind the Bet
What exactly did XPeng build instead? He remains deliberately vague on implementation details, citing ongoing development and competitive sensitivity, but the broad strokes involve abandoning software-primary architectures entirely in favor of AI-native systems. The old approach used "software engineering, the process of using software, plus AI algorithms and AI tools including the coding you just mentioned," he explained. "You can think of it as being in a business flow, or software, in some important tool links and process links, it's using AI. But we think it's wrong."
The new methodology treats the vehicle as a holistic intelligent agent rather than a software system enhanced with AI modules. He draws explicit parallels to XPeng's humanoid robot work, where the company is attempting to create systems that think through tasks comprehensively—understanding context, planning multi-step actions, and adapting to novel situations—rather than executing pre-programmed routines with AI assistance at decision points. "How do I throw away the original end-to-end logic, in fact a larger foundation model is used on more thoughts, how to open the upper limit of autonomous driving first, then go to the lower limit of convergence," he said, describing the philosophical shift.
The infrastructure implications are profound. He mentioned that XPeng's physical AI models consume token volumes orders of magnitude higher than digital applications—a single vehicle running the new VLA system for three to four hours daily processes more tokens than most digital use cases, though he declined to specify exact figures. The company has invested nearly one billion RMB annually just in data storage infrastructure, requiring specialized teams to manage which datasets need hot access versus cold storage versus archival. Computing power allocation follows similar patterns, with management focused on efficiency metrics for 30,000 to 50,000 H100 GPU equivalents rather than per-employee token budgets that digital-native companies track.
Early Market Validation
The first commercial proof point arrived in late March 2026 when XPeng released the initial version of its second-generation VLA system. In April, while China's overall passenger vehicle sales declined roughly 20 percent year-over-year, XPeng's deliveries surged 50 to 70 percent, with "considerable part" attributable to the new autonomous driving capabilities, according to He. The company is now projecting what it believes will be Level 4 capability within 18 to 24 months—a timeline He characterizes as realistic rather than the "bragging" he attributes to competitors claiming earlier deployment.
But He remains unsentimental about what even Level 4 achievement would mean. "Even if there are more sales, it does not represent long-term value," he noted. "You see there are different issues here." Success in automotive requires sustained excellence across dozens of dimensions simultaneously—what He calls the "long board, short board, and wide board" problem. A single technical achievement, no matter how impressive, cannot compensate for weaknesses in other areas when customers evaluate complete vehicle ownership experiences.
The Humanoid Robot Parallel Track
XPeng's robot development followed a similar trajectory of false starts and radical pivots. The company's robotics work spans three distinct phases: an initial 2018-2020 period when it acquired an independent team working on quadruped robots; a 2020-2023 phase exploring different approaches using both robotic methodologies and automotive methodologies; and the post-2023 era when He committed to universal humanoid robots driven by AI-first architectures.
The catalyst for the humanoid bet came in 2023 when He concluded that previous approaches could never solve what he calls the "brain" problem—the cognitive capacity required for general-purpose operation. Earlier systems relied on what He dismisses as "spine" or "brainstem" level control: maintaining balance and executing simple gaits. "Walk forward slowly with the same monotonous pace, I don't think it's the cerebellum," he said. "That's the spine. Or your brainstem is just maintaining balance. You are far from even having a cerebellum."
He selected Mi Liangchuan to lead the robot center based on what he describes as "quadrant" alignment—shared thinking frameworks rather than deep domain expertise. Mi had worked on neither robots nor autonomous driving specifically, which He viewed as an advantage rather than liability. "His quadrant is more consistent with my thinking quadrant," He explained. The team Mi assembled skews heavily toward recent graduates from top-tier programs, with nearly 80 master's and doctoral students from Tsinghua and similar institutions hired just in late 2025 and early 2026. "We must be willing to support their long-term exploration," He said. "You have to believe this group of young guys or girls, they can create miracles."
The strategic rationale for humanoid form factor centers on environmental compatibility and long-term societal integration. He walked through detailed reasoning about why quadruped robots fail in human spaces—a robot dog tall enough to be useful (1.1 to 1.2 meters) cannot navigate typical rooms without damaging furniture or itself, while smaller versions lack capability and battery life. Similarly, heavily armored bipedal robots create psychological barriers that prevent close human interaction, limiting their utility in home and commercial settings.
The Iron Bot Controversy
XPeng's robot garnered unexpected international attention in late 2025 when demonstration videos sparked widespread speculation that a human was inside the suit rather than autonomous systems. He described the episode with a mixture of frustration and bemusement. The robot shown was merely an intermediate milestone—one the team knew had thermal management issues and was "hot"—not a product intended for public scrutiny. But when conspiracy theories proliferated claiming the demonstration was faked, He personally intervened over his team's objections to definitively prove the robot's authenticity.
"Many people inside think that we don't have to say anything because we know it ourselves," He recounted. "But in fact 99.99 percent of people in this world have no way to touch him." He forced his team to devise a demonstration strategy overnight, culminating in a video showing the robot's left leg disassembled to reveal internal mechanisms. The choice was tactical: "Everyone can see the left leg. It walks from left to right. The left leg is the easiest to see. This is a real robot."
The public reaction revealed unexpected market segmentation that has since influenced XPeng's commercialization strategy. Social media analysis showed younger users primarily interested in robots for household tasks and work assistance, while middle-aged and older demographics—people who "are not posting on social networking sites"—focused intensely on eldercare applications. "I think this is something to look at from different perspectives," He said. "I haven't experienced such a big point of view before. He does not have such a big data."
The GX Flagship and Automotive Strategy
On the automotive side, XPeng is launching its GX flagship SUV as the company's first full-size six-seat model and most expensive vehicle to date. He positions the GX as synthesis of capabilities from across XPeng's portfolio—flying cars, robots, and conventional vehicles—rather than a conventional large SUV play. The vehicle incorporates six-way safety redundancy inspired by aircraft design, meaning it can continue operating even with power supply failures or severed wiring harnesses. "It's like an airplane, it allows you to have redundancy," He explained.
The integration goes deeper than hardware. XPeng's new wire-controlled chassis links directly to the VLA autonomous system through a redesigned electrical/electronic architecture, enabling what He describes as raising the lower limit of VLA execution—higher safety levels, lower latency, and improved control sensitivity by "nearly dozens of percent." Future versions will incorporate task-oriented AI from the robotics program, allowing the vehicle to understand and execute complex multi-step commands similar to how a humanoid robot processes instructions like "get me a bottle of water."
He deflected direct questions about whether the GX risks repeating the G9's troubled 2022 launch, which was marked by pricing confusion and market missteps. "The thinking and logic at that time, think about logic at this time, I think the difference is very big," he said. "If your ability reaches a higher dimension, you can probably analyze it all: where were the previous problems, and there is a high probability that what is the problem now." The implication is that XPeng's organizational and analytical capabilities have matured sufficiently to avoid similar execution failures, though He acknowledged the vehicle incorporates lessons learned across product planning, customer understanding, and business logic.
The target customer is described as over 30, "relatively brave," professionally successful, and willing to explore new technology—essentially He's peer group. "Our car, more expectations are for family-oriented technological luxury," he said. Pricing and detailed specifications are scheduled for announcement May 21, 2026, positioned as pre-sale ahead of formal market launch.
Industry Consolidation and Supplier Relationships
He doubled down on his prediction that China's automotive market will consolidate to roughly five large-scale companies by 2030, though he emphasized this doesn't mean other manufacturers will necessarily disappear—just that reduced scale will make ultra-high-intensity competition increasingly difficult. This consolidation thesis directly contradicts the optimistic scenario underlying Horizon Robotics' business model, which depends on numerous OEMs adopting third-party solutions rather than vertically integrating AI capabilities.
He carefully praised Horizon founder Yu Kai personally and the company's products, but expressed skepticism about the strategic viability of the tier-one supplier model for AI systems. "If you look at something from the user's perspective, because we look at a thing from a business perspective, the user perspective may be just one of our dozens of perspectives," He said. "Everyone feels the need for more car companies. There are more products from friendly merchants to choose from. There are more different things. But in this world, cars are too complicated."
The complexity argument centers on cross-domain fusion—integrating hardware, software, manufacturing, design, service, and operations in ways that defy clean modularization. He argued that while partnering with three companies to deliver a total solution is feasible, "you have to work with 300 companies to provide a personalized solution, you will be in great pain." XPeng's strategy of developing 80 percent of hardware in-house, including hand chips and joints for robots, reflects this integration philosophy extended to the physical AI domain.
Competitive Dynamics and the Sea of Blood
Asked whether XPeng has emerged from the "sea of blood" he famously described in prior interviews, He was unequivocal: "I think they are all swimming. I don't think anyone came out now." But he suggested robotics may follow a different competitive trajectory than automobiles. Unlike cars, where homogeneous products inevitably lead to brutal price competition, robots' high software value and lack of strong open-source alternatives should enable more differentiation and healthier economics.
He estimated there are now more than 200 established robotics companies in China, potentially double the peak number of automotive startups, though he noted robot categories are far more diverse than vehicle segments. "I think there may be many solutions in this world," he said. "That robot's winning rate is quite high. I think it's higher than a passenger car." But he was brutally honest about survival odds for companies pursuing universal humanoid robots specifically: "I personally think that if we adopt a universal humanoid robot, 99.99 percent will die."
XPeng's own odds, by He's assessment, sit at roughly 20 percent—a figure he characterizes as "the highest winning rate" among Chinese companies attempting this category. The bottlenecks to commercial robotics are threefold: achieving reliable and stable hardware, successfully integrating multiple large models at high level, and proving commercialization viability in real deployments. XPeng expects to reach automotive-style SOP (start of production) for robots by year-end 2026, with 2027 representing "the first year of commercial mass production" for high-level robots across China and the United States.
The difficulty exceeds automotive by He's estimation—20 to 100 times harder, with 20 representing the minimum multiplier. "Even with Xiaopeng's ability, after the marketing of this robot appeared at the end of last year, I think this has increased our probability of success a little bit," he said. The implication is that even for a company with XPeng's resources and technical depth, success remains far from assured.
Philosophical Shifts and Organizational Anxiety
He spoke candidly about the psychological challenges of leading through technological discontinuity. "Of course you are anxious," he said when asked about navigating the AI transition. "Because you will discover a lot of your previous logic theory, even the theorems and paradigms at the bottom of your logic theory, even down to the axiomatic level—your values, outlook on life, and world view. You will all feel under the changes of these things, it seems that the previous set of logic did not work so well."
The existential uncertainty extends beyond personal frameworks to organizational capabilities. He noted that applying token-usage metrics borrowed from digital AI companies makes little sense for physical AI businesses, and that most conventional wisdom about AI transformation is "not comprehensive and sufficient enough" when extended to the physical domain. "Copy their methodology to physics AI, I think there are many places that are not suitable enough," he said. "Maybe some of the business logic is correct, but they are not comprehensive and sufficient enough. This is the biggest problem."
On knowledge acquisition, He has largely stopped reading business books in conventional ways, finding that summarized rules become obsolete too quickly as the world changes. "I might just turn one chapter and pass it," he said of books on his desk. "The rules summarized in the book, the current combination changes too quickly, it's hard to summarize. After he writes a book and prints it, the world may have changed by another 10 percent." Instead, he emphasized learning through practice and rapid iteration cycles, particularly in AI systems where human involvement in the improvement loop should be minimized.
Despite the immense pressures and long odds, He expressed no regrets about past decisions, including the costly pivot away from proven technologies. "There is no need to think about regretful decisions in this world because I made too many mistakes," he said. "Why should you regret making a mistake? We often feel that it turns out that something was done wrong at this time. Then let's think about why it's wrong now. But don't regret it."