XPeng Bets Big on Physical AI as Automakers and Tech Giants Enter the Next F...

Standing on the stage of Technology Day 2025, He Xiaopeng spoke with deep emotion: "No one can beat the trend."

Over the past few decades, we've witnessed the internet connecting computers into vast networks, seen smartphones make information omnipresent, and lately watched AI systems like ChatGPT become creators in the digital realm. But all these revolutions have been confined to the screen.

Now, the tide is shifting toward a much tougher battlefield—the physical world.

He Xiaopeng feels this transition keenly. His team invested in 30,000 GPU cards, burned through over 2 billion yuan in training expenses, and stood on the edge of giving up countless times. Just when hopelessness was setting in, a turning point arrived unexpectedly—their autonomous driving system suddenly "woke up." It began to understand things it was never explicitly taught, like interpreting a pedestrian's hand gesture or the timing of a red light countdown.

This was more than just a technical breakthrough—it felt like a signal. When AI steps out from the virtual world and starts to perceive, comprehend, and control the real environment we live in, "physical AI" emerges as the next unstoppable tech wave barreling toward us.

The Race for a New Frontier

As AI starts to enter the physical world, the entire tech industry seems to have reached an overnight consensus: physical AI is the next battleground.

At the GTC conference, Jensen Huang declared that AI is shifting from "generative" to "agentic," with the next era being "physical AI." Alibaba CEO Wu Yongming also stated that the road to superintelligent AI has three stages, and the third is when AI acquires raw, comprehensive data from the physical world and achieves self-iteration through autonomous learning.

But beneath the consensus, the paths diverge.

Some players, such as Tesla and XPeng, have chosen the path of vertical integration. Taking a page from Apple’s playbook, they are committed to creating a closed-loop system—from chips and sensors to software and vehicle manufacturing—putting all critical technologies firmly in their own hands. Tesla has established its moat in autonomous driving with its fully self-developed FSD chip and vision algorithms. XPeng, on the other hand, leverages its proprietary Turing AI chip, second-generation VLA model, and cloud computing clusters to enable the transfer of these capabilities to flying cars and robots.

On a different path, companies like NVIDIA and Huawei continue to act as ecosystem enablers. For example, NVIDIA is building a powerful computing empire with its Blackwell GPU series and Cosmos platform. Huawei is focusing on supernode interconnection, with Vice Chairman Xu Zhijun emphasizing that this is the key to supporting large-scale physical AI computing power.

Meanwhile, companies like Waymo and Baidu Apollo have long focused on Robotaxi as the ultimate scenario, aiming to prove the ultimate potential of physical AI by conquering the highest level of fully driverless technology.

Though their paths diverge, the technological advancements of these giants all point in the same direction: enabling AI to evolve from a specialist that solves single problems to a generalist that can confidently handle the chaos and uncertainties of the real world.

For these companies, every road test, every algorithm update, and every new product launch has significance far beyond simply building a better car; each move is a crucial step in laying claim to a core ticket into the future.

Four Trump Cards, One AI Brain

The path ahead is clear, but the competition is just beginning. No matter which of these roads a player chooses, it’s a high-stakes gamble requiring tremendous investment. Tesla is betting that pure vision and scale effects can disrupt costs, NVIDIA is wagering that an open ecosystem will define industry standards, and Waymo believes that pushing technology to the extreme can conquer a singular scenario.

Standing on the stage at Technology Day 2025, XPeng founder He Xiaopeng spoke with palpable intensity: “No one can beat the trend.”

The comment, delivered to an audience of engineers, investors, and journalists, encapsulated a vision that stretches far beyond electric vehicles. For He, the next wave of technological disruption will not remain confined to the digital realm—it is coming to the physical world.

Over the past few decades, the tech revolution has unfolded on screens. The internet connected computers into vast networks. Smartphones made information omnipresent. And AI models, such as OpenAI’s ChatGPT, began creating content in the digital space.

These breakthroughs were transformative, but they largely remained virtual. Now, He Xiaopeng argues, the battlefield is shifting: toward autonomous systems that perceive, learn from, and act upon the real world.

At XPeng, this shift has been neither cheap nor easy. The company invested in nearly 30,000 GPU cards, spent over 2 billion yuan in training its autonomous systems, and faced repeated setbacks. Yet, after months of effort, the team reached a breakthrough: its autonomous driving system began to recognize patterns it was never explicitly taught—interpreting hand signals from pedestrians, understanding the timing of traffic lights, and navigating complex urban environments.

“It felt like the system woke up,” He said. “When AI can perceive, understand, and interact with the real world, that is the emergence of physical AI—the next unstoppable technological wave.”

As AI increasingly crosses into the physical world, the tech industry is converging on the idea that physical AI—the ability to act intelligently in real environments—will define the next competitive frontier. At the 2025 GTC conference, NVIDIA CEO Jensen Huang called the transition from “generative” to “agentic” AI the dawn of the physical AI era. Alibaba CEO Wu Yongming outlined a similar vision, identifying three stages to superintelligent AI, with the third stage achieved when AI integrates raw data from the physical world and iterates autonomously.

Companies pursuing physical AI are taking divergent approaches. Tesla and XPeng have embraced vertical integration, aiming to control every layer of the technology stack—from chips and sensors to algorithms and production. Tesla has fortified its moat in autonomous driving with proprietary FSD chips and vision-based algorithms. XPeng, meanwhile, leverages its in-house Turing AI chip, second-generation VLA model, and large-scale cloud computing clusters to expand its physical AI ambitions beyond cars to robots and flying vehicles.

Other players have positioned themselves as ecosystem enablers rather than vertically integrated operators. NVIDIA’s Blackwell GPUs and Cosmos platform, for instance, provide computing infrastructure that supports multiple applications of physical AI. Huawei focuses on high-performance interconnects, with Vice Chairman Xu Zhijun emphasizing the need for “supernode” connectivity to power large-scale AI workloads. Meanwhile, Waymo and Baidu Apollo continue to push fully autonomous Robotaxi services, proving the upper limits of physical AI in controlled scenarios.

Despite these different approaches, the underlying goal is shared: to evolve AI from a narrow specialist into a generalist capable of navigating the uncertainties and complexities of the physical world.

XPeng’s Four-Pronged Physical AI Strategy

XPeng has positioned itself uniquely in this race by pursuing four interrelated scenarios—autonomous driving, humanoid robots, flying cars, and smart vehicles—under a single AI framework. The company’s core model, the Second-Generation VLA, diverges from industry norms by eliminating the intermediate step of translating visual input into descriptive language. Instead, the model generates control commands directly from raw visual data, forming an end-to-end loop from perception to action. This allows the AI to respond faster and learn underlying physical principles without losing information through language conversion.

To support this model, XPeng has significantly upgraded its computing infrastructure. Onboard AI processing reaches 2,250 TOPS, enabling billion-parameter models to operate in vehicles—a scale far beyond the tens of millions of parameters typical in industry competitors’ systems. In the cloud, XPeng’s 30,000-GPU cluster continuously trains and refines the model, ensuring ongoing evolution.

The upcoming XPeng Robotaxi, slated for launch in 2026, illustrates this approach. Purpose-built for autonomy rather than retrofitted from conventional vehicles, the Robotaxi relies on a vision-only system and does not require high-definition maps or LiDAR. By leveraging the Second-Generation VLA’s generalization capabilities, the Robotaxi can adapt to diverse cities and traffic patterns, targeting the industry’s persistent pain points of cost, coverage, and user experience.

In robotics, XPeng’s IRON humanoid robots aim to create machines that are not only lifelike in appearance but capable of independent thought. Flexible joints, realistic skin, and advanced AI allow these robots to operate commercially in settings such as shopping centers and factories, performing guidance and inspection tasks.

In aerial mobility, XPeng HT’s “Land Aircraft Carrier” and “A868” flying cars are designed to accelerate the commercialization of low-altitude transportation. The Land Aircraft Carrier has reportedly received 7,000 pre-orders and entered pilot production, signaling a tangible path toward large-scale deployment.

The common thread among these initiatives is the notion of the “capability ladder.” Each scenario—Robotaxi, humanoid robots, flying cars, and smart vehicles—builds upon the last, enabling XPeng to progressively validate and enhance physical AI capabilities in increasingly complex real-world environments.

Data: The Fuel for Physical AI

He Xiaopeng has repeatedly emphasized that data is the lifeblood of physical AI. High-quality, scenario-specific data—not merely large volumes—is crucial. “Driving 100 kilometers on a straight road produces zero value. The key is long-tail and anomalous data,” he explained. Every interaction, from a Robotaxi navigating city streets to a robot guiding visitors, feeds the AI system, enabling iterative improvements.

This focus on diverse scenarios explains XPeng’s multi-pronged strategy. Each scenario acts as a data mine, enriching the AI’s understanding of chaotic, unpredictable environments.

The second challenge is ecosystem complexity. While collaboration is essential for industry growth, excessive openness can dilute competitive advantage. XPeng addresses this by maintaining in-house control over core components, including chips, models, and data loops, while selectively opening certain capabilities to partners. For example, Robotaxi capabilities are offered as a toolkit for partners such as Amap, and collaborations extend into robotics and chip development.

He Xiaopeng describes this approach as “close the loop at critical points, open up for ecosystem expansion.” This hybrid strategy balances the need for technical autonomy with the benefits of industry collaboration.

As automakers, tech giants, and robotics companies increasingly blur lines between industries, the era of physical AI has shifted from concept demonstrations to real-world deployments. Tesla, NVIDIA, Waymo, and XPeng are all vying for control, yet none has fully solved the puzzle. Success depends not only on technological sophistication but on building commercially viable products that consumers trust and embrace.

For XPeng, the stakes are high. Its vertical integration, multi-scenario strategy, and investment in computational infrastructure represent a bold bet on a future where AI transcends screens to operate in the physical world. If successful, XPeng could emerge as a defining leader in the age of physical AI, setting standards for the integration of autonomous systems into everyday life.

The broader implication is clear: physical AI is no longer an abstract concept—it is a tangible, competitive battlefield. And in this race, those who can seamlessly combine data, computation, autonomy, and ecosystem partnerships are likely to set the pace for the next era of technological disruption.