At the fourth AI-Defined Auto Forum, hosted by Gasgoo in 2026, the industry's buzzword was unmistakable: 'unified agent.' After 'large models on board' in 2024 and 'end-to-end' in 2025, almost every technical presentation this year repeats the same phrase. But behind the marketing, the reality for production cars is that cockpit AI and ADAS (advanced driver assistance systems) still operate as two separate systems. Executives from Nio, Joynext and Zhongke Yaoluchuan gathered for a roundtable to ask whether the industry is truly moving toward a single intelligence, or just talking about it.
According to First Electric, the debate revealed that 'unification' means different things to different companies. Geely's '1+2+N' framework, built around a vehicle-level agent called Eva, lets domain agents for driving, cockpit, chassis and energy talk to each other rather than merging into one model. Volcanic Engine, the cloud and AI arm of ByteDance, leans toward a 'one brain' approach, but its vice president Yang Liwei clarified that the brain handles user intent while driving remains the 'cerebellum' for motion control. Nio takes a more radical path: its self-developed SkyOS·Tianshu operating system unifies more than 1,600 atomic capabilities across six domains, from assisted driving to the cloud, at the operating-system level.
Different roads to the same destination
The technical paths diverge sharply. Horizon Robotics, a chip designer, pushes a hardware-centric route: its Star 6P chip, built on a 5nm automotive process, delivers 650 TOPS of BPU compute and 273 GB/s of memory bandwidth, shrinking the space needed for computing by 50% and cutting per-vehicle cost by 1,500 to 4,000 yuan (about $211 to $563). Dongfeng, working with Black Sesame Technologies, has put the Tianyuan cockpit-plus-driving platform into its Yipai 007, using a single chip to handle cockpit, L2+ assisted driving and parking. Volcanic Engine argues that merging models is inefficient because driving models need 10 to 48 Hz inference while cockpit models need only 1 to 2 Hz; the cost of combining them would be prohibitive. Joynext integrates chips and software in its nCCU central computing platform, but its global ADAS head Wang Haowei admits the real challenge is that automakers have different priorities for cockpit and driving, so the company adopts a 'seek common ground while reserving differences' approach.
Nio's Gao Jie, head of cockpit AI, is blunt about what he sees as a wrong turn: 'In the past few years, there were attempts to do VLA (vision-language-action) from the driving domain for human-machine interaction. I think that direction is completely wrong.' His reasoning: driving decisions need fast, error-free loops, while cockpit interaction is slow and high-level. The two are fundamentally different modes.
Safety demands isolation
The push for unification collides with safety concerns. In March 2026, a Lynk & Co Z20 crashed into a highway barrier after its voice system misinterpreted a command and turned off the exterior headlights. The incident shows what happens when the interaction layer reaches safety-critical actuators without a barrier. Waymo's robotaxi, the Ojai, integrates Google Gemini as a passenger assistant, but Gemini cannot change the route or control windows and seats, and it is explicitly barred from claiming driving capability. Horizon's Star chip includes a 'castle' physical isolation architecture, keeping cockpit and driving domains separate and ensuring that a cockpit reboot does not affect driving functions, which meet the ASIL-D safety standard.
Gao Jie's advice for near-term deployment: 'Models stay models, product code stays product code. All hard safety rules must be written in code, not handed to the model.' Wang Haowei agrees: 'No matter how intelligent autonomous driving becomes, safely moving people from A to B remains the first principle.' The industry's current answer is to unify at the interaction layer while isolating at the safety layer.
Underneath it all, executives argue, the real foundation is not chips but the engineering system. Gao Jie says the core of full-domain AI is 'full perception, full execution, central control architecture, SDV software stack and data loop.' Without learning, he says, the agent cannot survive. Tesla's Shanghai Lingang AI training center has built a complete loop from data collection to local training to over-the-air deployment, accumulating more than 3 billion kilometres of Chinese road data. Nio has been building its data-loop framework since 2021, upgrading to a three-layer training architecture of world model, supervised fine-tuning and closed-loop reinforcement learning. As Wang Haowei puts it, 'Models define the ceiling, but infrastructure decides the future of AI.'
