From Virtual Computing Power to Physical Capability: The Automotive Industry Enters the Physical AI Era in 2026

If one word were to define the core of automotive intelligence competition in 2026, it would no longer be "computing power" or "large models," but "Physical AI." This concept has been frequently mentioned at recent AI-powered automotive technology seminars. The industry consensus is that future intelligent driving companies will essentially be "mobile physical AI companies."

This means that AI must break through the limitations of data processing and content generation, learning to understand physical laws such as gravity, friction, and materials, and interact with the real world in a safe and efficient manner. The automobile, as the most complex movable physical entity, becomes the main battlefield for this transformation. The focus of competition is no longer on the smoothness of in-car screens or the wit of voice assistants, but on how to implant a "vehicle intelligence" brain into this steel body—one that can think, perceive, and more accurately "act."

1. Understanding and Control: AI Learning to Communicate with the Physical World

The core demand of Physical AI is to master physical knowledge and make safe decisions, which has been the biggest weakness of traditional smart vehicles. In the past, algorithms relied on simulation training, excelling in standardized scenarios, but struggling to handle complex real-world conditions such as non-standard road situations, unexpected obstacles, and extreme weather. The execution capability on the ground has been severely lacking.

In 2026, the industry is fully committed to breaking down the barriers between virtual algorithms and the real world, abandoning homogenized conversational large models and focusing on developing native vehicle intelligence systems. Great Wall Motors' Guiyuan platform was the first to implement a dual VLA large model architecture, relying on human-like vehicle design and opening up full system control interfaces such as steering, braking, and thermal management to achieve AI-driven global coordination. Leveraging multimodal fusion capabilities, AI can accurately interpret natural user commands, coordinate multiple systems to work together, automatically adjust temperature control, plan the optimal route, and adapt to driving styles, ensuring a seamless transition from semantic understanding to physical execution.

2. Architectural Reshaping: From "Functional Silos" to "Unified Command Hub"

The implementation of Physical AI forces a fundamental restructuring of automotive electronic and electrical architectures. In the past, areas such as intelligent driving, cabins, chassis, and body were independent, with fragmented hardware and software and disconnected data, creating functional silos that severely limited AI's cross-domain coordination capabilities. The current industry solution is to use in-vehicle large models as a unified command hub, deeply integrating cabin and driving functions, and restructuring the entire electronic and electrical architecture, elevating AI to the role of the vehicle's commander. The IM Ultra Agent from Zhiji, equipped with the IM Fusion Nova fusion architecture, connects the steer-by-wire chassis, advanced intelligent driving, and smart cabin. With the vehicle-grade Qianwen large model, it can recognize complex, everyday commands like "find a shaded parking spot" and autonomously complete the entire process of environment perception, path planning, and automatic parking.

Geely's "Super Eva" system, developed in collaboration with JY Star, follows a similar logic. Its core large model can memorize user habits and, based on real-time road conditions, battery levels, and even passenger status, dynamically plan routes. It also automatically calls services such as charging and dining along the way.

The mass production of Super Eva marks the first time that a car has embodied the form of a "super intelligence entity.

This marks the first time that AI has gained deep, global control over a vehicle. The car transforms from a collection of functions that require multiple layers of commands into an organic intelligent entity capable of autonomous planning and collaborative action.

3. Facing Challenges: Compliance, Costs, and Data Barriers

With AI deeply integrated into vehicle control, industry challenges have become more pronounced, forming key constraints for large-scale implementation. Compliance risks have emerged as the number one concern: in March 2026, the National Market Supervision Administration's Defective Product Recall Technical Center began public consultations on the "Automotive Software Quality and Defect Management Guidelines (Draft for Comments)." As the world’s first automotive-grade standard combining functional safety, expected functional safety, information security, and AI safety, its self-assessment for safety covers over 160 indicators, with extremely strict auditing standards.

At the same time, the core provisions of the EU's Artificial Intelligence Act (AI Act) will come into full effect on August 2, 2026. Although compliance obligations for embedded high-risk AI systems have been postponed until August 2, 2028, the Act’s definitions of "high-risk" categories, data governance compliance, and mandatory human oversight principles have already set clear compliance boundaries for global automotive companies. The sharp rise in compliance costs is forcing some companies to reassess their technological approaches.

• The Cost Paradox of the "Full Inference Era":Baidu Vice President Shi Qinghua pointed out that the automotive industry is rapidly entering the "Full Inference Era," where the incremental computing power brought by AI inference will account for more than 80% of total computing needs. This means that as intelligent cabin features become more popular and user interactions increase, the cloud computing costs required to support real-time inference will continue to rise. This creates a sharp contrast with the traditional automotive business model, which calculates costs based on a one-time BOM (Bill of Materials). Balancing optimal user experience with a sustainable business model has become a key challenge for large-scale adoption.

• Data Closed Loop Determines Experience Depth:The quality of algorithms increasingly depends on the accumulation and iteration efficiency of large-scale, real-world road data. Dr. Yu Qian, Chairman of Lightrun Zhihang, emphasized that world models and reinforcement learning are essential pathways to achieving general-purpose Physical AI, and closed-loop simulation and real data are the core supports. The era of single sensors is over. The trend is toward the integration of multimodal data, including vision, voice, and biosensors, to construct a complete "perception-decision-execution" loop. Whoever’s data feedback loop accelerates the fastest will set the standards for "reliability" and "usability."

4. Collaborative Breakthrough: Accelerating Industry Chain Integration

Physical AI spans multiple fields, including chips, actuators, algorithms, architectures, and regulations, making it impossible for a single company to achieve full-stack in-house development breakthroughs. The industry chain is now collaborating around core elements such as computing power, algorithms, chips, and vehicle integration to address technological and cost challenges. Our association provides regulatory interpretations and technical matchmaking services to assist automotive companies in shaping next-generation intelligent architectures. At the same time, we are coordinating with upstream and downstream supply chains to promote the synchronized adaptation of sensors, steer-by-wire components, and in-vehicle AI systems, shortening development cycles and reducing integration risks.

2026 is a pivotal year for the implementation of automotive Physical AI, marking the shift from "brain-based interaction" to "hands-on control" in automotive intelligence. In the future, the vehicle intelligence system will redefine smart mobility. We invite industry partners to join hands in building the ecosystem, accelerating the mass production and implementation of Physical AI technology.