XPeng Transcript: He Xiaopeng on Abandoning XPeng's Multi-Billion Dollar ADAS System for Physical AI and the High-Stakes Race for Humanoid Robots
Zhang Xiaojun Video Podcast Interview, May 28, 2026
The Role of AI Tools and Coding in Modern Enterprises
Zhang Xiaojun: Hello everyone, I am Xiaojun. Unfinished Date is a first-class program co-produced by Weibo Finance and Language is the World Studio. Today, our guest is He Xiaopeng, the founder and CEO of XPeng Group. We will discuss his exploration of physical AI, though he still has many secrets he cannot share. Before we officially start, I want to do a quick survey about your recent experience with AI. What are the AI products you have used most recently?
He Xiaopeng: To be honest, I really do not use many. I still use very traditional AI products like Tongyi Qianwen and Doubao. However, within our team, we use AI a lot for coding. Personally, I do not want to use it too deeply. When my team asked me why I did not want to use it myself, I gave them an interesting example from when we were building internet products. If you use a product every day, you quickly get bogged down in the minute details. You focus too much on its immediate shortcomings and functions that do not work, rather than looking at its long-term potential. Once you start using it constantly, you focus all your energy on how to solve its current bugs, which prevents you from looking into the distance. So, while I am deeply concerned about the rapid changes in technology, I believe the person in the number one position should not use these tools too deeply. At the grassroots level, however, we must encourage and even force its adoption, and eventually normalize it.
Zhang Xiaojun: Does AI-assisted programming make sense for the number one position of an enterprise? Do you think it will benefit XPeng, the automotive industry, or the intelligent driving industry as a whole? Will it bring about real changes?
He Xiaopeng: I think it is an excellent tool for junior programmers, but it remains an auxiliary tool for now. Perhaps in two to three years, it will force junior programmers to quickly transition into senior programmers. However, for things like intelligent assisted driving or other advanced, specialized AI capabilities, its direct help is relatively small. It is just one of many tools. The entire underlying infrastructure must be built first. Once the entire system is established, AI coding really only helps at the application layer. If you are working at the kernel layer, such as writing an operating system, the core value lies in the overall system infrastructure rather than the coding itself.
The Financial and Compute Realities of Large-Scale Physical AI
Zhang Xiaojun: How many tokens does your company use every month? How much attention do you pay to that metric?
He Xiaopeng: We do not pay much attention to token usage. Over the past year, many people have brought this up. I think digital small and medium-sized companies might need to focus heavily on tokens, and medium-to-large digital companies should pay appropriate attention to it, but it should not be a comprehensive metric. Some companies' core support businesses are not fully digitalized. We actually ran an interesting statistic on our cars. If our new-generation Vision-Language-Action, or VLA, model only runs for three or four hours a day, how many tokens will it use? I do not remember the exact number, but it is processed as an inner loop. In digital AI, the number of tokens used is far lower than the number of tokens required to parse the physical world. However, comparing the two is ultimately meaningless because a self-driving car operates as an automated machine. It uses as many tokens as it needs to process the environment. In the physical world, token usage will eventually be about how much value a machine can generate for people and enterprises, which is a completely different dimension from digital AI.
Zhang Xiaojun: What do you think is a reasonable budget for tokens, and how do you manage those costs internally?
He Xiaopeng: I do not know. Internally, I try not to restrict token usage for everyone. Many executives ask me about controlling token expenses, claiming they can save a year's worth of expenses in a single quarter. I believe that if someone can really generate value, my job is only to monitor and manage the most abnormal cases, such as the top ten outliers in terms of spending. The rest should be left open. If an employee spends a thousand or ten thousand RMB a month but generates massive value, why should we restrict them when their salary far exceeds that amount? Currently, our token distribution is concentrated in our General Intelligence Center, which is the merged team of our autonomous driving and cockpit divisions. They are a massive team. We measure computing power rather than just tokens. For example, we look at how efficiently a business uses an allocation of thirty thousand or fifty thousand NVIDIA H100 GPUs.
Zhang Xiaojun: You mentioned managing outliers. What kind of data outliers have you focused on recently?
He Xiaopeng: We recently established strict controls around data. Many people talk about the value of data, but very few companies realize the astronomical costs associated with managing it. In the digital AI field, the data size for training is relatively small, often just dozens of terabytes. But when we train physical AI models, we are processing dozens to hundreds of terabytes in a single run. Storing and managing this data is incredibly expensive. Our direct, rigid costs for data investment are close to more than one billion RMB every year. We have to analyze which data is valuable, which data is temporarily useful, which data must be accessed rapidly, and which data can be warmed up and reused. Optimizing these workflows can save tens of millions of RMB and bring massive efficiency gains. We have dedicated teams managing data and computing power specifically for this reason, which is why I do not worry too much about individual engineer token counts. The computing power required during training for physical models is where the real cost lies.
Why Automating the Role of a CEO is Far Harder Than Autonomous Driving
Zhang Xiaojun: If we were to train an AI model to replicate your skills—a "He Xiaopeng skill"—what training data would we need to feed it, and what would that skill look like?
He Xiaopeng: Today, digital and physical models make it much easier to automate the skills of basic white-collar workers than basic blue-collar workers. However, if we look at the logical progression of deskilling, once higher-end blue-collar and white-collar roles are automated, there will be massive societal risks. A CEO could technically be replaced in a few decades or a century. By then, my capabilities could indeed be packaged into an AI skill. But by that time, human CEOs will also have developed stronger, more comprehensive capabilities. We have thought deeply about this because we are building robots. There are two major contradictions here. Most people think about how to turn their existing abilities into skills. But from a model perspective, how does the system know that a digitized skill is correct and continuously improve through continuous learning? In a physical simulation model, this can be strengthened. This is very different from coding or autonomous driving. In coding and autonomous driving, it is clear when a system makes a mistake. But if you try to make He Xiaopeng's decision-making into a skill, it is extremely difficult to judge whether a strategic decision is correct or incorrect in real-time. That is why we are focused on building the underlying system itself rather than just using a system built by others.
Zhang Xiaojun: Intuitively, what would be the main shortcoming of a digitized He Xiaopeng skill?
He Xiaopeng: I think once anyone's abilities are fully digitized, everyone will realize that this person actually has many shortcomings. As a CEO, I am certainly no exception.
Zhang Xiaojun: As a CEO, do you think your company's focus on AI is sufficient today?
He Xiaopeng: I think digital-world companies can focus heavily on AI, perhaps dedicating dozens of percent of their resources to it. In the physical world, XPeng is a company with tens of thousands of people, and I believe we should spend fifteen to twenty percent of our total resources on what I call pan-AI. This includes our autonomous driving division and our robotics division. This is a reasonable and sufficient proportion.
The Pivot from AI Car Company to Physical AI Powerhouse
Zhang Xiaojun: If we place artificial intelligence companies on one side and manufacturing companies on the other, how do you balance the two? Do you have an internal struggle over whether XPeng should be an AI company or an automotive manufacturing company?
He Xiaopeng: I do not look at it that way. I believe car development consists of several types of research and development. First is hardware R&D, second is software R&D, and AI is just one component of the software. Then there is manufacturing R&D. Hardware R&D and manufacturing are two completely different skill sets. For example, you might have the design capability to sketch a beautiful table and chairs, but you probably do not have the manufacturing capability to actually build them. In an automotive company, mastering these different R&D capabilities is the core foundation. In the future, there might even be a fifth type of capability.
Zhang Xiaojun: Last year, you were still calling XPeng an "AI car company." This year, the company officially rebranded to XPeng Group, and you stated that XPeng is now a "physical AI company." What was the strategic reasoning behind this name change, and what are the practical implications?
He Xiaopeng: This transition has just begun, and there are still many details we cannot fully disclose, but I can share some context. Over the past decade, XPeng focused on smart electric vehicles. We designed our first cars, mass-produced them, sold our first hundred thousand units, and entered a standard manufacturing cycle. When we started in 2014, basically no one believed in vehicle intelligence. Only a few believed in electrification, and most people just saw cars as a traditional, large-scale business. By last year, in 2025, everyone accepted that electrification was the future because of the global shift toward new energy. However, even though intelligence has developed rapidly in the automotive field, the progress has still been unsatisfactory. From Toyota and Google to Baidu, Tesla, and XPeng, we have all poured immense resources into autonomous driving. We achieved significant results, but we did not reach the height we truly wanted. The systems we built were what I call Stitch Monsters—a combination of software rules and AI algorithms, rather than an integrated AI driver built entirely on a unified AI model. It was still software-based logic.
The Multi-Billion Dollar Bet: Dropping the Stitch Monster for the VLA Foundation Model
Zhang Xiaojun: What happened within XPeng last year to change this paradigm?
He Xiaopeng: Last year, we encountered a major inflection point. We were working on two generations of our intelligent assisted driving system simultaneously. Internally, we called them our first-generation VLA and second-generation VLA. The first-generation VLA enlarged the model to lower the influence of traditional software rules, enhanced the backend capabilities, and strengthened post-training. The second-generation VLA took a completely different direction. We discarded the traditional end-to-end logic and utilized a much larger foundation model to unlock the absolute upper limit of autonomous driving capabilities. We then worked to converge the lower limit, which means minimizing critical errors and ensuring the car behaves exactly as expected. Generalization in autonomous driving remains a massive problem. Today, no autonomous driving company can drive smoothly in an unfamiliar underground parking lot without relying on memory-assisted driving, where the car must first run the route once to map the space. This shows that the system's actual understanding of the physical world is extremely low. The upper limit of the old architecture was simply too low. To do it well, you need to open up ten thousand possibilities, but the old systems were capped at a thousand. Around this time last year, our second-generation VLA opened my eyes to a massive shift. Its upper limit could reach a hundred thousand or even a million points, but its lower limit at the time was terrible. Under the old system, you might have an upper limit of one thousand and a lower limit of nine hundred, which is very stable. With the new model, the lower limit was initially much lower than our existing commercial products. We faced a choice: keep using software engineering methodologies and AI as a toolbox within a traditional business flow, or make a massive gamble. We realized that using traditional software methodologies with AI tools just creates a more powerful software Stitch Monster. So, around April of last year, we made a massive bet. We stopped development on our previous system, which had cost us several billion RMB to build.
Zhang Xiaojun: Why did you decide to kill a multi-billion RMB system? What was the trigger?
He Xiaopeng: Because I realized the old system would never achieve true driverless autonomy, nor would it allow robots to generalize. If a robot walks into an unfamiliar room, recognizes you, sits down, and responds naturally when you decline a glass of water, that cannot be achieved using strong rules and minor AI algorithms. I realized that while we wanted to build an incredibly smart car, our existing software methodologies would never allow it to become infinitely smart. It was a shortcut, but not the actual road. We needed to find the real path. We had to trust that the value of hardware and software would eventually split fifty-fifty, where customers would be willing to pay one hundred and fifty thousand RMB for the hardware of a three hundred thousand RMB car, and another one hundred and fifty thousand RMB for its comprehensive software and intelligence capabilities. We had to completely restructure our strategic planning and R&D processes from the bottom up.
Zhang Xiaojun: Were there any landmark meetings within the company when you made this massive bet? How did you execute such a major organizational change?
He Xiaopeng: There was no single dramatic meeting; I made the final decision in my head and we went all out. By the end of the third quarter of last year, we made our move and completely restructured our autonomous driving center. In any organization, talented people have inertia. They get comfortable using old methodologies even when they are handed the latest AI tools. But to truly innovate, you have to change your core methodology and mindset. I cannot share the exact operational steps we took because they are highly strategic and proprietary. We adjusted different business units at different tempos, focusing heavily on leadership and organizational structure. It is extremely difficult. If you do not change the underlying logic, you just end up using AI to build a house faster, but you are still just repairing an old house. We wanted to build something entirely new. As a CEO, you have to place the bet, accept the feedback, and guide the organization through the transition. XPeng is a large startup with tens of thousands of employees, so managing this organizational change is a completely different challenge compared to a small team. When you are the one holding the overall picture, you receive advice from many different angles, but most of it is incomplete. Ultimately, you have to integrate those perspectives and make the call.
Zhang Xiaojun: I heard that you have become much more willing to gamble on big decisions in the past three years.
He Xiaopeng: It is not about being a gambler; it is about knowing when you must place a bet and doing it early. By the end of 2022, when XPeng was facing massive challenges, I established two core philosophies for myself. First, never admit defeat. Second, be willing to admit defeat. Balancing these two means that even when faced with immense difficulties, you must persist and push through. But you must also maintain the objective clarity to shut down projects that are no longer viable, just as we did when we completely stopped our first-generation VLA. Hesitating, waiting, and observing only delays your timeline and makes success impossible. In a few years, once we are more successful, we can discuss the exact details. Every company must find its own unique path. Today in China, everyone talks about copying Company A or Company B, but that is a mistake. What works for digital AI cannot simply be copied to physical AI. It is a completely different world. When digital AI companies try to define the physical world without ever having run a physical business, they build physical AI in a very narrow sense. The actual physical world involves complex human interactions, environmental variables, regulatory compliance, and commercial viability. You have to think far broader than just having a couple of strong features. That is why digital AI CEOs cannot easily explain the transition to physical AI; they are still discovering if they can even succeed.
The Rebirth of XPeng's Humanoid Robot IRON
Zhang Xiaojun: Let us talk about your humanoid robot, IRON, which generated massive attention last year. How did this product come about, and why did you pivot to universal humanoid robots in 2023?
He Xiaopeng: XPeng's robotics journey actually spans three distinct stages. The first stage was from 2018 to 2020. It was an independent team of about four or five companies in China exploring quadruped robotics. The second stage was from 2020 to 2023. Over those four years, we built three different milestones. We tried making robots using traditional robotics methods and even tried making robots the way we make cars, stitching various elements together with mixed success. The third stage began after 2023. When we saw the progress of foundation models in late 2022, our entire logic changed. Previously, we believed it was impossible to build a successful robot brain because the complexity of the cerebellum—maintaining physical balance and motion control—was too high. Today, many companies claim they have developed a robot cerebellum because their robot can walk forward slowly at a monotonous pace. That is not a cerebellum; that is just a basic spine or brainstem maintaining balance. They are far from achieving true cerebellum functionality.
Zhang Xiaojun: What major organizational changes did you make in 2023 to realize this new vision?
He Xiaopeng: In 2023, we decided to pivot completely from four legs to two legs. We abandoned our old assumptions and focused on a brand-new design driven directly by the robot's brain, combined with our engineering expertise from the automotive sector. Having good technology does not guarantee a good product, and having a good product does not guarantee you can scale production. Cars have a highly mature process spanning from planning and design through ET, PT, SOP, and SOD. By the end of this year, we hope to transition our robots into an automotive-grade SOP process. By 2027, we expect high-level robots to enter their first true year of commercial mass production globally. At that point, traditional motion-controlled robots will begin to decline as advanced physical AI robots take over. The core value of physical AI lies in its ability to generate both emotional and physical value for humans by performing actual work. In 2023, we realized that our existing robotics team, despite being highly familiar with traditional robotics, was incapable of building this new generation of physical AI robots. At that time, our robotics division, led by LC, had around three hundred people. We broke the team up, leaving fewer than sixty core members. Many of those who left went on to start their own startups and raised multiple rounds of funding. But to reconstruct the logic of the entire robot, we could not rely solely on traditional automotive engineers or traditional robotics experts. We needed a completely new team with a unified understanding of AI, automotive engineering, manufacturing, and robotics.
Zhang Xiaojun: Why did you choose LC to lead this effort when he was neither a traditional robotics expert nor an intelligent driving specialist?
He Xiaopeng: Sometimes it is down to fate. His strategic thinking and mental quadrant aligned perfectly with mine, making him the right person to drive this new way of thinking. Many of the robot demos you see from other companies today are still utilizing what we consider third- or fourth-generation technology stacks. They are just running basic tests. At XPeng, we have the patience and courage to invest for the long term. A quick demo means very little. It is just like back in 2017 when China had hundreds of autonomous driving startups showing off level 4 data; very few of those technologies ever translated into real, commercialized value.
Zhang Xiaojun: What is LC's recruitment and talent strategy for this rebuilt team?
He Xiaopeng: LC focuses heavily on what he calls talent density, which I view as talent potential. He recruits the absolute best minds. From the end of last year through the first half of this year, our robotics department alone hired nearly eighty master's and doctoral graduates from top-tier institutions. They are incredibly expensive, but we are fully committed to supporting their long-term exploration. We believe that we must use super-smart people to solve super-difficult problems, rather than relying on rigid processes and predefined tools. LC is incredibly ambitious—he often tells me he wants to create artificial humans rather than just build commercial robots. He wants robots to have a genuine sense of participation in our society and to connect deeply with human emotions.
Overcoming the Pitfalls of Bipedal Humanoid Robotics
Zhang Xiaojun: Why did you commit so firmly to universal bipedal humanoid robots over other forms?
He Xiaopeng: It is the most challenging path, but it is the one that will have the most profound impact on human society. Over the next few decades, universal humanoid robots will become deeply integrated into human lives. While digital AI is limited to assisting in a few dozen white-collar roles, physical AI robots can address hundreds of physical roles, especially in aging societies. I believe the two most critical factors for the future of humanity are medical AI to help the elderly live longer, healthier lives, and physical AI robots to provide care and companionship. For many elderly individuals, a robot may eventually become their primary support system. We chose this difficult path because we analyzed the fundamental flaws of other forms. For example, we previously developed quadruped robots—dogs and horses. But if you bring a robot dog that is over a meter tall into a standard home, it cannot function. It cannot turn around next to a bedside table without scratching the walls or damaging the bed. Unlike a real golden retriever whose tail and body are soft, a rigid robot dog will inevitably cause damage. If you make the robot dog smaller, its battery life becomes too short, and its utility drops to basic companionship.
Zhang Xiaojun: What about bipedal robots? What are the human-centric design challenges there?
He Xiaopeng: If you build a massive, 1.8-meter bipedal robot covered in heavy metal armor, even its own designer will feel an intense sense of physical oppression standing next to it. You will naturally worry about it falling, overheating, exposing high voltage, or simply being dirty. If adults feel that way, how will children and the elderly react? How do you resolve those safety and legal regulations in a household setting? While industrial robots can utilize that design because they operate in controlled environments, household robots must be designed differently. They must be comfortable for humans to interact with. That is why our current generation of robots is designed to be around 1.69 to 1.70 meters tall—a height that is comfortable for both men and women. They are designed to wear clothes and can even have hair, but they must not have a realistic human face to avoid the uncanny valley effect and other complex sociological issues.
Behind the Scenes of the Great Humanoid Robot Controversy
Zhang Xiaojun: During your press conference last year, you mentioned that your team was deeply conflicted about whether to prove that the robot on stage was fully autonomous and not operated by a human. Why was that such a struggle?
He Xiaopeng: Personally, I did not care, but my team was incredibly anxious. They argued that the more we tried to explain, the more skeptical people would become, as online public opinion is often cynical. They wanted to wait twenty-four hours to see how the public reacted before making any statement. But after a few hours, I could not stand the spreading rumors anymore. In China and globally, the speculation was spreading at lightning speed. If we waited twenty-four hours, the narrative would be completely out of control. We knew internally that it was a real robot and that its thermal management was still a work in progress, making it run hot. But it was an important milestone for us, and we knew future versions would be significantly better. So, that night, I called the team and told them to prepare a demonstration for the next morning. I insisted we show everyone that it was a real robot operating autonomously, rather than letting the skepticism linger.
Zhang Xiaojun: How did you prove it to the public?
He Xiaopeng: We decided the simplest and most indisputable method was to have the robot walk while we removed its left leg. Because the robot walks from left to right, the left leg is the most visible to the audience. Removing the leg while it was functioning demonstrated the internal mechanical and software architecture clearly, proving it was not a human in a suit. This quickly cleared up the public relations crisis.
The Market Potential, Commercialization, and Future Timeline of Humanoid Robots
Zhang Xiaojun: How has your robotics strategy evolved since that public event?
He Xiaopeng: It accelerated our recruitment. We have attracted an incredible array of talent to our robotics division. Starting a robotics business is vastly different from starting a car company. I believe starting a robotics company is twenty to one hundred times more difficult than building a car company. Even with XPeng's established manufacturing and AI capabilities, our success last year only increased our overall probability of success by a small margin. Today, there are over two hundred robotics startups established in China, which is double the number of electric car startups we saw during the peak of the automotive boom. However, unlike cars, which are primarily categorized into passenger, commercial, or special vehicles, robots will have countless classifications, ranging from medical and freight transport to cargo inspection. Many of these specialized robots do not need to be humanoid. But on the path of universal bipedal humanoid robots, ninety-nine percent of companies will fail. The software complexity is simply too high, and there is no open-source platform capable of helping a robotics startup build high-fidelity physical AI software. We are focused on navigating these pitfalls ourselves because you only truly understand the engineering challenges once you step on the traps.
Zhang Xiaojun: Who do you consider your primary competitors in the universal humanoid robotics space?
He Xiaopeng: Right now, universal humanoid robots have no true rivals. It is entirely a race against ourselves. Every robotics company must focus on making itself as strong as possible, building the underlying organization, the hardware infrastructure, and the overall system integration. Last year, our robot went viral because of its highly lifelike motion control. Traditional automotive motion control has been largely commoditized over the last century, with car companies purchasing motor controllers from Tier 1 suppliers. But to make a robot truly capable, its motion control must be far more integrated and responsive than a car's. In a car, different control domains are isolated. This works for standard driving, but if a car has its left tires on snow and its right tires on grass, executing a forty-seven-degree turn with a person suddenly appearing in front of the vehicle is incredibly difficult. Managing four-wheel traction, movement balance, and latency under those conditions remains a challenge. For a robot, the complexity is multiplied. A human has over two hundred joints, allowing for an infinite loop of potential movement combinations. Replicating this full-pose personification using AI-driven motion control rather than rule-based software is incredibly difficult. We want to give our robots the same physical instinct a human has when walking on snow, ice, or grass, sensing friction and adjusting balance instantly. This is why we develop eighty percent of our robot hardware in-house, including our actuator joints and hands, while cooperating with upstream Tier 2 suppliers to scale quality.
Zhang Xiaojun: What are the main engineering and commercial bottlenecks you are facing as you prepare for mass production?
He Xiaopeng: The first bottleneck is ensuring our in-house hardware is exceptionally reliable and stable. The second is achieving a seamless fit between our high-level foundation models and the physical actuators. The third is proving the commercial viability of the product. The market is waiting for its iPhone 4 moment. The first commercially produced robots will probably not even be as polished as the original iPhone 1, but they will represent a massive paradigm shift. Once a robot achieves generalized software capabilities, its market adoption and production scaling will happen far faster than cars. Cars took a century to scale because we had to build global road infrastructure, establish traffic regulations, and manage highly complex manufacturing logistics. Robots can be deployed instantly into existing human environments. If the software is ready, physical AI will scale rapidly.
Introducing the GX: XPeng's Six-Seater Flagship SUV and Front-Mounted Robotaxi
Zhang Xiaojun: Let us talk about your automotive business. You are releasing several new cars this year, including the GX. This represents XPeng's return to the high-end market, correct?
He Xiaopeng: Yes, the GX is our first full-size, six-seater flagship SUV. What makes the GX unique is that we have integrated many of our advanced R&D capabilities from our flying car and robotics divisions into this vehicle. For example, we took the safety redundancy systems developed for our flying cars' flight-critical components and built them into the GX. The GX is China's first mass-produced passenger vehicle equipped with a front-mounted Robotaxi architecture, featuring six full safety redundancies. Even if the primary power supply fails in the wilderness, the car can still be driven. Even if a mouse chews through a wiring harness, the system remains operational, much like the safety systems on commercial aircraft.
Zhang Xiaojun: How does the software architecture of the GX integrate with its physical chassis?
He Xiaopeng: We have connected our wire-controlled chassis to our new Electronic and Electrical Architecture, or EEA, and our VLA autonomous driving system. This integration allows the chassis to execute VLA decisions with significantly shorter latency, raising the safety lower limit and improving control sensitivity by dozens of percent. We have also brought robot task-planning logic into the vehicle. In future in-car interactions, the vehicle will process tasks much like a humanoid robot. When you give the system a task, it will first identify who is speaking, determine their authorization level, and plan a sequence of actions to execute the command. Furthermore, we designed the interior space to be highly versatile. The third row can be folded completely flat, instantly transforming the SUV from a six-seater into a highly spacious four- or five-seater vehicle. We customized every detail, co-developing premium privacy glass with Fuyao Glass and a new-generation in-car refrigerator with Midea. Building a vehicle with this level of integrated capability is the only way to win in today's highly competitive market.
Zhang Xiaojun: How did you determine the pricing for the GX? Are you worried about repeating the strategic mistakes of the original G9 launch?
He Xiaopeng: I am not worried at all. Our strategic thinking, pricing logic, and organizational capabilities are in a completely different dimension compared to when we launched the G9. Over the past three and a half years, we have completely overhauled our product planning, organizational structure, customer understanding, and core business logic. When your systemic capabilities reach a higher dimension, you can easily identify past errors and prevent them from happening again. Our official launch event on May 21st will showcase our configuration and pricing clearly. The GX is designed as a family-oriented, technologically luxurious vehicle for professionals, managers, and entrepreneurs over thirty who are eager to experience the luxury of cutting-edge technology.
Consolidation and the Battle in the Automotive Sea of Blood
Zhang Xiaojun: How do you view the fierce competition with other flagship SUVs like Li Auto's L9 or Nio's flagship SUV series?
He Xiaopeng: I see them as allies as well as competitors. We are all friends. During the Beijing Auto Show, I visited Li Xiang's booth to check out the new L9, and I visited William Li's booth to see Nio's flagship SUV series. We all have our own unique understandings and solutions for this segment. Aesthetics is ultimately subjective, but I am highly confident in the visual design of our vehicles. We are on a path where our designs are becoming more and more beautiful. Some people ask why we do not focus on making fewer models rather than launching four cars this year. Every company has its own strategic path. For us, launching these models is aligned with our broader business scale and technology integration strategy.
Zhang Xiaojun: Where are you spending most of your personal time now? Is it in cars, robots, AI, or organizational management?
He Xiaopeng: I participate in all of those areas, but I spend the majority of my time on long-term strategy and planning. In the automotive industry, scale alone does not guarantee permanent success, and short-term profits can often be painful and temporary. My job is to integrate our technical capabilities, organizational structure, and market positioning into a cohesive, long-term strategy. Cars are highly complex, composite systems. Having a single longboard is useless if your shortboards are failing. For example, some companies brag about achieving level 4 autonomy this year, which is simply unrealistic. I believe true Level 4 autonomy will begin to see actual implementation in eighteen to twenty-four months. When it is implemented, it will definitely drive car sales, though the exact multiplier remains to be seen. In late March, we released the first version of our second-generation VLA. While China's overall car sales in April saw a year-on-year decline of around twenty percent, XPeng's sales increased by fifty to seventy percent, a considerable part of which was directly driven by our second-generation VLA. True success for an automotive company requires a master class in multi-system integration, combining hardware, software, manufacturing, design, and operations.
Zhang Xiaojun: Yu Kai, the founder of Horizon Robotics, often says that you are his most challenging partner because you insist on in-house development. He believes car companies should ultimately rely on third-party suppliers for software and intelligence. What do you think of his view?
He Xiaopeng: Yu Kai is a close friend, and Horizon Robotics has done an excellent job. However, his strategic path depends on whether there will be more or fewer automotive and robotics manufacturers in the future. If the market concentrates, his addressable market shrinks, making his path more challenging. I believe the market will become highly concentrated. By 2030, China may only have five large-scale automotive groups left. While other small players might survive in niche segments, their lack of scale will make it increasingly difficult to compete. Historically, many car companies did not conduct genuine in-house research; they specialized in assembly and integration. If your goal is short-term tactical execution, relying on third-party integration is correct. But if your goal is long-term survival, you must master self-research. Ten years from now, software will represent more than fifty percent of a car's value. A Tier 1 supplier cannot easily help hundreds of different partners achieve that level of deep, cross-domain fusion. You have to integrate hardware, software, manufacturing, and user operations into a single, cohesive system.
Zhang Xiaojun: Have you finally swam out of the automotive "sea of blood" today?
He Xiaopeng: No, we are all still swimming. I do not think anyone has truly made it out of the sea of blood yet. The robotics market will be slightly better because the software barrier is exceptionally high, preventing the immediate, homogeneous price wars we see in the automotive sector. My daily schedule remains intense, much like any other entrepreneur. I start my day late, work very late into the night, and make rapid-fire decisions across all business units. I do not read many business books anymore because the physical world is changing too rapidly; by the time a book is printed, its insights are already outdated. I prefer absorbing knowledge through direct, hands-on practice, traveling, and constant communication. Looking back, I do not harbor any regrets about my past decisions. We make far too many mistakes in business to spend time regretting them. The key is to analyze why a mistake occurred, learn from it, and keep moving forward.