He Xiaopeng: For smart car companies to make robots, it is not only a technological origin but also an inevitable exploration of AI

"In the future, AI automotive companies will eventually become robotics companies." XPeng Motors founder He Xiaopeng made this thought-provoking industry judgment at the 2025 Tencent ConTech Conference and Tencent Technology Hi Tech Day hosted by Tencent News. This is the second time in a month that He Xiaopeng has expressed a similar view. At the XPeng Technology Day in early November this year, He Xiaopeng announced that "XPeng Motors' positioning has been upgraded to an explorer of mobility in the physical AI world, a global embodied intelligence company," which well explains why XPeng Motors is transforming. XPeng is not the only company exploring physical AI with robotics. According to incomplete statistics, by December 2025, about 18 automotive companies worldwide have announced their entry into the humanoid robotics sector, including Tesla, Xiaomi, GAC, BYD, Hyundai, Toyota, and others. In He Xiaopeng's view, the core reason for intelligent automotive companies crossing into the robotics sector lies in the natural advantages of technological commonality and the underlying demand for AI exploration in the physical world, which will ultimately reshape the industrial landscape of the intelligent era under the catalysis of the "ant colony effect." Why are automotive manufacturers rushing to enter embodied intelligence? He Xiaopeng believes this is inevitable. There is a significant overlap in the technological systems of automobiles and robotics, and this commonality makes technological complementarity possible. Over the past decade, more and more automotive companies have built a complete technological system from underlying chips to physical world models and upper-level applications through self-research, and this system just happens to provide a mature technological foundation for robotics development. From a hardware perspective, the core hardware of intelligent vehicles highly aligns with the hardware requirements of robots. He Xiaopeng introduced that the XPeng robot is equipped with three self-developed Turing AI chips, providing 2250 TOPS of effective computing power. The development experience of such high-computing power chips directly comes from the technological accumulation of intelligent vehicle autonomous driving chips. In terms of software and algorithms, the autonomous driving algorithms of intelligent vehicles need to process real-time data in complex road conditions to achieve functions such as path planning and obstacle avoidance, and the logical algorithms of these functions are highly consistent with the autonomous movement and task execution requirements of robots in indoor and outdoor environments. More importantly, the industrial capabilities of the automotive industry provide crucial support for the mass production of robots. The large-scale manufacturing experience, supply chain management systems, and quality control standards of intelligent vehicles can be directly applied to robot production. This interconnection of technology and industrial capabilities allows intelligent automotive companies to develop robots without starting from scratch, but rather to iterate and upgrade based on existing foundations, forming an efficient technological complementarity effect. 01 The "huge" data source of the physical world is the next explosive point for AI If technological commonality is the "hard foundation" for intelligent automotive companies to develop robots, then the continuous exploration of AI in the physical world is the "soft power" and core demand. He Xiaopeng believes that the development of AI in the digital world over the past 30 years has validated the value of data, computing power, and models, but the growth of valuable data in the digital world is slowing down, while the physical world contains a vast amount of data sources, which is precisely where the next explosive point for AI lies. ** Smart cars themselves are an important carrier of AI in the physical world, but robots can enable deeper and more comprehensive physical exploration of AI. He Xiaopeng pointed out that human thinking logic is often constrained by language, and existing large models have the same limitations. The intelligence of the physical world needs to be centered around multi-modal experiences such as perception and touch. The AI systems of smart cars mainly revolve around the "mobility" scenario, while robots can cover more scenarios such as home services, commercial services, and industrial production, allowing for the collection of richer physical world data — including tactile feedback, spatial interactions, task execution data, and more in different scenarios. This vast amount of physical world data will feed back into the iteration and upgrading of AI models — the scale of data in the physical world far exceeds that of the digital world, and robots, as "data collection terminals" of the physical world, can provide richer training materials for AI models. 02 When intelligent agents learn to think independently, the "ant colony effect" will emerge He Xiaopeng proposed the "ant colony effect" in his sharing, providing a long-term era annotation for smart car companies' layout in robotics. In the industrial era, scale effects were core; in the internet era, network effects were key; and in the intelligent era where AI is deeply coupled with the physical world, the "ant colony effect" will become a new industrial logic. The so-called "ant colony effect" refers to decentralized intelligent agents forming efficient, robust, and adaptive systems through independent thinking, decision-making, and collaboration. Ants do not need to report to the queen after discovering food; instead, they call nearby companions to collaborate. This decentralized collaboration model is an important feature of the future intelligent world. Smart cars and robots are essentially intelligent agents of the physical world. When these intelligent agents possess autonomous decision-making and collaboration capabilities, they will form the "ant colony effect." "This ecological layout will fundamentally change the existing industrial pattern." He Xiaopeng predicts that the next generation of large companies will have 10 million intelligent agents and 1 billion GPUs, covering both the digital and physical worlds, achieving efficient collaboration through the "ant colony effect." Smart car companies, leveraging their first-mover advantage in the development of intelligent agents in the physical world, will become builders of this ecosystem — cars, as the largest mobile intelligent agents, will provide data, computing power, and scenario support for other intelligent agents, while robots and other intelligent agents will enrich the application scenarios of the ecosystem, forming an industrial ecology of data sharing, technological interconnection, and scenario complementarity. Under the influence of the "ant colony effect," the boundaries between smart cars and robots will gradually blur, creating a situation where "everything is an intelligent agent." Users will no longer face isolated car or robot products, but rather an intelligent ecosystem that can meet various needs such as mobility, living, and working. The core competitiveness of this ecosystem lies in the collaboration capabilities and data intercommunication abilities among intelligent agents, and smart car companies have already seized key nodes in ecosystem construction through technology-aligned layouts From the natural advantages of technological homology, to the underlying demands of AI physical exploration, and then to the era trend of the "ant colony effect," it is not a cross-industry move for intelligent automotive companies to engage in robotics, but an inevitable choice for industrial development. He Xiaopeng's sharing reveals a core logic: the competition in the intelligent era is no longer a competition of single products, but a competition of intelligent ecosystems. XPeng's layout aims to start with automobiles and, through the collaboration of robots, flying cars, and other intelligent entities, become an explorer in the physical AI world. In the future, when robots enter commercial scenarios and households, when autonomous driving achieves full popularization, and when flying cars shuttle through urban skies, these intelligent entities will reshape our way of life through the "ant colony effect." Intelligent automotive companies will also complete their transformation from "mobility service providers" to "embodied intelligent companies" in this revolution, promoting human society into a new era of deep coupling between physics and AI. The following is a transcript of the speech, with some adjustments made without changing the original meaning: I am very happy to share with you today at Tencent Technology about XPeng's thoughts on the future of physical AI and related considerations, as well as some practices we are currently undertaking. Just now, Academician Chai spoke very well about the changes in the industrial world. In fact, because I have started businesses in different industries, both in the internet and in the physical world, I have seen a huge change. Over the past 30 years, we have seen the emergence of computers, the internet, mobile internet, and today we are witnessing changes brought by AI. At the same time, if we look back over a longer period, more than 100 years, we can see the birth of the automobile and the invention of the airplane from the 1860s to the 1890s, as well as many changes in the physical world over the past 100 years. From my perspective, these two changes are very interesting. One is centered around physical engines, or chemical engines, including electrochemical engines. We can see the transition from oil to electricity, from petrochemical engines to electric motors. In the digital world, it represents another type of energy, another type of engine. I believe this energy is increasingly seen as data-driven energy, where large computing power and large models serve as a new engine for propulsion. Therefore, I firmly believe that in the next 20 to 30 years, we will see a large amount of coupling between physics and AI entering a new era. In this new era, we can see that each era has its own capabilities. In the industrial era, we know how to manufacture cars, mobile phones, and many other things, where scale effects are the most important. In the internet era, we all know about network effects, which come in various forms: unidirectional network effects, multidirectional network effects, and network ecological effects. The core logic of network effects is that as the network grows larger, each person within that network creates or generates value, content, and capabilities, thereby enlarging the network effect. I believe that in the next era, which is the entire AI era, if we couple physics and AI together, it will produce new effects Let's take a look at what is called the agent effect. I saw the form of performance, and I first want to use a small case from Professor Zeng Ming, who mentioned a "black hole effect." I believe the "black hole effect" is a progression of the data "Matthew effect," which means that AI is continuously consuming knowledge and turning it into a black hole, emerging with entirely new capabilities. We have all seen this change in the digital world, and I particularly want to emphasize that this change will be even more significant in the physical world. How much data is there in the digital world today? If I remember correctly, it is about hundreds of TB; I can't quite recall if it's tens or hundreds. However, we see that the scale of valuable data in the digital world is slowing down, meaning it is amplifying at a slower rate, but the physical world is completely different. We see that in robotics, the data in the physical world far exceeds that in the digital world. So from another perspective, more data, larger models, and better computing power lead to further intelligence of agents, attracting more users, obtaining more valuable data from both the physical and digital worlds, and absorbing it. We see this as a huge, entirely new effect in the intelligent era. Let me share another effect we see in the intelligent era, called the ant colony effect. Previously, we knew that network effects were centralized or semi-centralized, but today a decentralized effect has emerged, which we call the ant colony effect. What is the ant colony effect? Everyone knows about ants. If an ant finds a piece of bread on its way to find food, it doesn't actually need to return to the ant nest to report to the ant king or queen; it can issue a nearby call, and the ants that come in will cooperate with it. This is a decentralized, very robust system, and it has a high degree of adaptability. Therefore, we see an interesting phenomenon: in the future world, when an agent can think, decide, and act on its own, a new effect called the ant colony effect will emerge. We see that future humanoid robots will exhibit the black hole effect, or the ant colony effect, or in future practices, there will be more similar, entirely new effects of the intelligent agent era. So let's boldly predict that I believe large enterprises will undergo a new change in the next few decades. In the past 100 years, we saw that large companies had 100,000 people, and on average, each person had one tool. The ability to integrate with tools was the most important, and this tool formed the industrial era. In the last 10 to 20 years, we have seen large companies with 100,000 people managing and using 1 million servers, serving possibly 1 billion users. However, in the next generation of large companies, there may be an additional 10 million agents, which could be agents from the digital world or the physical world. But more importantly, it will shift from 1 million servers to 1 billion cards. Why are there so many cards? How can so many agents be managed effectively? This is a challenge that new enterprises will face XPeng is transitioning from the last decade to a new decade, focusing on full-stack self-research and overall physical AI. We are developing both the physical world and the AI world simultaneously. Therefore, from the underlying chips to the foundational physical world models, to the upper-level Robotaxi and humanoid robots, we have been deeply researching for the past decade. In our research at XPeng, we have observed something very interesting: we directly transform visual signals into planning and control, which is a significant change. When we were young, we grew up learning from books and acquiring knowledge. But have you ever thought about a question? For example, if there is a book that teaches you how to walk, walking is a very complex action for a person. In developing robots, we have started to truly decompose and analyze how a person walks well. Today, when you see all the robots in the world walking, they may only have one to three postures, and they are not smooth and coherent. However, a person can walk at one year old and perform many actions by two years old. Is there a book that teaches a person the act of walking? Thus, we find ourselves trapped in a huge incomplete cycle from childhood learning, where knowledge, or language and the world, is constrained by the language of the world. Many people know the philosopher Ludwig Wittgenstein, and what he said, in a certain sense, is that each of our logical thoughts is confined by the language of this world. So today, our knowledge is like this, and our large models are like this as well. But from another perspective, if we want to learn to swim today, it is best not to look at it from a capital perspective, but rather to have someone get into the water and swim with you for a while; this will speed up your learning to swim. Therefore, the model of the physical world and the operating system of the physical world are fundamentally centered around perception, touch, and other multimodal elements, forming a new generation of models. Last month, in November, you could see the launch of XPeng's new IRON, which gained public and industry attention with a very human-like gait. This time, we saw a significant breakthrough; many of my friends' children are discussing it, and they are very interested in technology. We also noticed a lot of overseas friends who can feel that the world of robots has undergone such rapid and significant changes. Many friends have asked us, in the past 7 or 8 years, why XPeng is now working on the eighth generation of humanoid robots while we are currently seeing the seventh generation. Why did we start with humanoid robots? Here, I would like to share that in the past seven generations that have been completed, we have developed four generations of quadrupedal robots. However, we found that quadrupedal robots have many issues and are difficult to use commercially. As you know, in this world, there are not only rational values but also emotional values; in addition to public utility value, commercial value must also be considered. Secondly, all environments in this world are designed, used, and operated from a human perspective, making it very easy to integrate into our environment For example, when we were working on quadruped robots, we found that if they enter a household, everyone knows what a bedside table is. If there is a dog in the house, it needs to turn around in place at the bedside table, but it is very difficult for a quadruped robot to turn around in place; only humanoid robots can do this more easily, as they can integrate into this world more seamlessly. Additionally, many people believe that it is impossible to create humanoid robots without models. So where does the data come from? It can only come from humans; it is impossible to fully simulate the physical world relying solely on digital twins. Therefore, from another perspective, data acquisition is very important. Moreover, many people ask why humanoid robots do not combine robotic capabilities with tools. Many forget that many tools in this world are developed for humans; a person can use 1,000 types of tools or even 10,000 types of tools. Therefore, in future robots, they can also use a significant proportion of tools. If a robot can universally use tools, it may not outperform a specialized robot, such as a vacuuming robot, in a single scenario, but it will certainly be more comprehensive and versatile, which is undeniable. From another perspective, this is an important point for why XPeng is developing humanoid robots. I believe that in the next 20 years, we will see tens of thousands of robots in this world, possibly with various shapes, both humanoid and non-humanoid. However, from another perspective, humanoid robots will definitely be our main companions in life. Cars and robots share the same origin. We believe that in the future, AI car companies will all become robot companies because there are two-legged robots, four-legged robots, and robots with wings in this world, which create different capabilities. The entire automotive system has a lot of common origins with robots in terms of hardware, software, processes, manufacturing, and commercialization. By 2025, XPeng and its ecosystem will have invested approximately 11 billion RMB in research and development in physics and AI, and we can see very rapid progress in humanoid robots. Therefore, our humanoid robots are different from many others; we use a very large computing power and have developed three self-researched Turing AI chips with 2250 TOPS of effective computing power. We have created a new combination of VLA for movement, VLM for interaction, and VLT for tasks, integrating different brains, cerebellums, left brains, right brains, and brain stems. Moreover, our IRON can also operate in airplanes. We hope that in about a year, our new IRON will begin to provide mass production, initially offering services in commercial scenarios. I believe that entering the industrial sector in Europe and the United States is the most suitable, while in China, commercial applications may be more appropriate. However, many people ask me when we will enter households. I think it will be very difficult to enter households in the short term, as households present significant challenges in terms of safety, reliability, and generalization. But I believe that in the not-too-distant future, we will definitely see robots entering households Finally, let me share with you about XPeng's endeavors in the flying car sector. We are developing two types of flying cars: the first type integrates an airplane into the trunk of a car. We plan to officially mass-produce and launch this in the second half of next year. Recently, many friends have visited our aircraft factory in Guangdong to see for the first time how airplanes are produced on an assembly line. This airplane has a very short flight distance, only a few dozen kilometers, but it is very convenient to fly. The second type is a long-range vertical takeoff and landing fixed-wing aircraft capable of flying at speeds of around 500 kilometers or 400 kilometers per hour. We boldly imagine that in 1970, when our parents got married, the "three major items" were a bicycle, a watch, and a sewing machine. In 1990, twenty years later, the "three major items" that many young people needed were a color TV, a refrigerator, and a washing machine. By 2010, many people believed that cars, houses, and diamond rings were the "three major items." I believe that in the next ten years, it is very likely that robots and low-altitude aircraft will enter our daily lives, which is also what XPeng is working on. Therefore, XPeng hopes to become an explorer of mobility in the physical AI world. Additionally, we will firmly move towards globalization and become a globally integrated intelligent company. That concludes today's sharing, thank you all! Risk Warning and Disclaimer The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investment based on this is at one's own risk