NextFin News —For most of the past decade, autonomous driving systems improved by refining separate pieces of software. Cameras and sensors spotted objects. Another program guessed what those objects might do. A third program chose a path. The approach worked in many everyday situations, yet information was lost each time it passed from one piece to the next. The car could react well to familiar conditions and still struggle when something unexpected appeared.

That design is now giving way to systems that try to understand and anticipate the physical world as a whole. Industry leaders call the shift physical AI: software that does not merely label what it sees but reasons about motion, cause and effect, and likely futures before it acts. At the 2026 Consumer Electronics Show, NVIDIA chief executive Jensen Huang stated the change directly: “The ChatGPT moment for physical AI is here—when machines begin to understand, reason and act in the real world.” He has since described the broader opportunity as multi-trillion-dollar in scale, with self-driving vehicles among the earliest large applications.

The technical steps have been gradual. Tesla’s Full Self-Driving version 12, released in North America in 2024, replaced large amounts of hand-written rules with a single neural network that took camera images and produced steering and braking commands. The change showed that a unified approach could work across a large fleet. Chinese suppliers and carmakers followed with similar systems through 2024 and 2025. Those early versions made driving smoother and reduced certain failures, yet they remained hard to examine when something went wrong.

Two ideas then gained ground. One adds a layer of broader understanding so the system can interpret a scene more like a person would—recognizing not just a pedestrian but the likely intention behind the movement—and, in some designs, respond to spoken goals. The other, often called a world model, tries to learn the basic rules of how the physical environment behaves. With that knowledge the car can run short internal simulations of what might happen in the next few seconds and choose the safer path. By 2026 most leading programs had stopped treating these two approaches as rivals. Understanding the present and predicting the near future are now usually combined.

Real deployments show the change in practice. Momenta’s R7 system entered production cars in 2026 and has been extended to driverless parcel vans operating in parts of Suzhou, including overnight routes. Company materials state that the same core model, trained on data drawn from more than twelve billion kilometers of real driving, supports both passenger and delivery vehicles without needing detailed pre-made maps of every street. At a major computer-vision conference in 2026, XPeng described a system that builds an internal sketch of how traffic is likely to evolve before the car commits to a maneuver. NIO has rolled successive versions of its own predictive model across a large share of its fleet. Huawei’s driving software uses large cloud systems to generate difficult practice scenarios and smaller onboard models for real-time decisions. Tesla continues to expand its unified networks while adding more reasoning capacity and closed-loop practice in simulation.

The competitive terms have shifted with the technology. Earlier races focused on the number of sensors, the power of the chip inside the car, and the sheer volume of labeled examples. The new race centers on the quality of the core model that understands physical behavior, the variety of real driving data used to train it, and the realism of the digital practice environments where rare situations can be rehearsed. Companies that already run large fleets hold an advantage in data. Those with substantial capital can buy more simulation capacity and more computing power for training. Smaller teams face steeper barriers.

The practical limits are still clear. Even tens of billions of kilometers of real driving leave many unusual situations poorly covered. Digital practice multiplies the number of scenarios that can be tested, yet lessons learned in a simulated world do not always transfer perfectly to messy public roads. The models must also run quickly and efficiently inside the power and cost limits of a production vehicle. Checking safety becomes harder as the systems grow less transparent. Public talk of physical-AI companies and foundation models therefore describes a direction more than a finished product. Higher levels of automation continue to expand in carefully defined or mapped areas, yet fully unsupervised driving on open roads remains limited to selected trials.

The deeper change is architectural. Once a model can simulate physical consequences, the same core software can, in principle, support driving, cabin features and, later, robots that handle objects. Several carmakers have already brought driving, cabin and robotics research teams closer together so they share model development. The near-term contest is still about safer and more capable vehicles. The longer contest is about who builds the most useful model of how the physical world works.

Physical AI has not removed the hard problems of autonomous driving. It has reframed them. Success now depends less on writing ever more detailed rules and more on gathering the data, computing power and engineering care required to teach machines how the world actually moves. The companies that master that teaching will set the pace for the next phase of the industry. The gap between today’s supervised systems and reliable unsupervised ones remains wide, and the resources needed to close it continue to grow.