At the 2026 AI-Defined Vehicle Forum, Meng Chao, senior director of SAIC's passenger car intelligent software center and CTO of Zeros Technology, challenged the industry to rethink what makes a car truly AI-driven. "What is a real AI car? I get asked this every time, but there is never a standard answer," he said, according to Chinese outlet First Electric Vehicle (D1EV).
Meng drew a clear line between software-defined vehicles (SDV) and AI-defined vehicles (AIDV). In the SDV era, software still follows deterministic rules written by humans, with AI bolted on as an extra feature whose main value is cost reduction and efficiency. In an AIDV, model inference and learning generate behavior within safety boundaries, making AI the core of the vehicle's native architecture, with the goal shifting to user experience and generational product leaps.
From voice commands to task agents
Meng illustrated the difference with a scenario: at night in the rain, a child sleeps in the back seat, and the user says, "Be quiet, find a pharmacy that is still open on the way, and remind me when we get close." In a traditional software system, the user would have to search for the pharmacy, modify the route, adjust the cabin environment, and set a reminder separately. In a native AI system, the vehicle understands the goal of buying medicine and the constraints—don't disturb the child, minimize detours, remind before arrival—then coordinates navigation, cabin, vehicle control, and ecosystem services, re-planning as the environment changes.
"Judging an in-car AI should shift from 'what model it uses' to 'what complete task it can accomplish for the user,'" Meng said. He proposed a five-stage maturity ladder for vehicle AI, from A1 (large-model-enhanced voice) through executable task agents and whole-vehicle physical agents, eventually reaching an "AI-native mobility companion."
On the architecture side, Meng said an AI-defined vehicle requires native AI products, a cloud-edge collaborative model system, AI software infrastructure, and supporting chips and hardware. Zeros Technology is exploring a "1+N" agent architecture: one System Agent handles intent recognition, task orchestration, and conflict arbitration, while domain agents for cabin, driving, vehicle control, and energy execute in coordination.
Meng stressed that "1+N" does not mean letting a large model directly control the vehicle. The System Agent manages global tasks, while domain agents and SOA services guard capability and safety boundaries.
Beyond TOPS and memory walls
As models enter vehicles, underlying hardware must evolve. Meng argued that evaluating chips cannot rely on TOPS alone; model compilation efficiency, memory bandwidth, scheduling utilization, thermal stability, and safety availability together determine true "effective computing power." For the "memory wall" posed by large models, he suggested HBM and 3D stacking to increase bandwidth in the short term, computing-in-memory in the medium term, and optical computing as a longer-term direction.
Meng also noted that AI-native products demand organizational change. Following Conway's law, if the product architecture has moved to the next generation while the internal organization remains stuck in the previous generation's software and function divisions, the new AI system cannot be truly established. In his view, whether in cars or embodied robots, AI is just landing in different terminal forms—what matters is whether a sustainable evolution of models, software infrastructure, data loops, and computing systems exists behind it.
