Cao Xudong: Autonomous Driving Is the Prologue to Physical AI

Introduction: When Autonomous Driving Meets the Physical AI Wave

In 2025, the AI industry's focus is rapidly shifting from the digital world to the physical world. "Physical AI" concepts such as embodied intelligence, autonomous driving, and robotics have become the new frontier for capital and technological competition. Amid this wave, Momenta CEO Cao Xudong has offered a clear and profound judgment — autonomous driving is not the destination, but the prologue to physical AI.

In a recent in-depth conversation, Cao Xudong systematically laid out his thinking on the relationship between autonomous driving and physical AI, and put forward a viewpoint that has sparked heated industry discussion: "Physical AI requires a ticket, and that ticket is having a business with cash flow."

Core Viewpoint: Physical AI Requires a 'Ticket'

In Cao Xudong's view, physical AI is not a track that can be started from scratch. Unlike large language models that can iterate rapidly by leveraging internet data, physical AI confronts the complexity of the real world — it demands massive amounts of physical-world data, continuous engineering iteration, and extremely high safety standards. All of this translates into enormous capital investment and lengthy R&D cycles.

Therefore, Cao Xudong emphasizes that to truly gain a foothold in the physical AI space, companies must first possess a core business capable of generating stable cash flow. For Momenta, this "ticket" is its mass-production autonomous driving solutions that have already been deployed at scale.

Momenta has established deep partnerships with multiple mainstream automakers, with its mass-production intelligent driving solutions installed in dozens of vehicle models, covering scenarios ranging from highway navigation to urban intelligent driving. These mass-production projects not only bring the company sustained revenue but, more importantly, continuously accumulate real-world driving data — which is precisely the scarcest resource for training and optimizing physical AI models.

In-Depth Analysis: The Evolutionary Logic from Autonomous Driving to Physical AI

Cao Xudong's judgment reveals a deep paradigm shift currently underway in the autonomous driving industry.

First, the data flywheel is the core engine of physical AI. Autonomous vehicles drive on real roads every day, collecting data that covers every aspect of the physical world, including weather changes, traffic flow, and pedestrian behavior. After cleaning, labeling, and training, this data feeds back into AI models, making their understanding of the physical world increasingly precise. Once this "data flywheel" starts spinning, it creates a formidable competitive moat.

Second, a closed commercial loop determines the ability to survive. Looking back over the past few years, many autonomous driving companies have fallen into difficulty or even gone bankrupt due to over-reliance on funding and a lack of self-sustaining revenue. Cao Xudong's "ticket theory" is essentially a sober reminder to the industry: while technological idealism is important, technology exploration without a closed commercial loop to support it is ultimately unsustainable. Only companies that can achieve commercial deployment today are qualified to participate in the long-term race for physical AI.

Third, autonomous driving is the most mature testing ground for physical AI. Compared to directions still in early exploration stages, such as humanoid robots and industrial embodied intelligence, autonomous driving has higher technological maturity across perception, decision-making, planning, and control, a more complete industrial supply chain, and clearer deployment scenarios. The world models, multimodal perception capabilities, and real-time decision-making algorithms forged through autonomous driving are expected to transfer to broader physical AI applications in the future.

Notably, this line of thinking is highly consistent with the current evolutionary trends in the global AI industry. Whether it is Tesla extending from autonomous driving to its humanoid robot Optimus, or NVIDIA expanding its autonomous driving simulation platform into a general-purpose physical AI infrastructure, industry leaders are all validating the same logic through their actions: autonomous driving is the most solid starting point for physical AI.

Industry Outlook: The Dawn and Challenges of Physical AI

Standing at the juncture of 2025, the development of physical AI is exhibiting several clear trends.

First, end-to-end large models are reshaping the technical architecture of autonomous driving. Traditional modular approaches are gradually giving way to end-to-end neural network solutions, bringing autonomous driving systems closer to a state of "truly understanding the physical world" and paving the way for evolution toward broader physical AI.

Second, World Models have become a key technological direction for physical AI. By simulating the operating principles of the physical world in virtual environments, AI can complete extensive training and testing before being deployed in real-world scenarios, significantly reducing the cost of trial and error. Momenta's positioning in this direction is an important step in its journey from autonomous driving toward physical AI.

However, challenges cannot be overlooked. Long-tail problems in the physical world are far more complex than those in the digital world, and safety verification standards are far higher than those for internet applications. How to accelerate technological iteration while ensuring safety, and how to achieve capability transfer across different physical scenarios — these are common challenges facing all physical AI participants.

Cao Xudong's "ticket theory" provides the industry with a pragmatic thinking framework: while pursuing the grand vision of physical AI, every step of commercial deployment on the ground is critically important. As the prologue to physical AI, autonomous driving's story has only just turned to its most exciting chapter.