The year 2026 is witnessing AI Agents becoming the most crowded track for companies from chips to vehicles to software suppliers.
On June 5th, Qualcomm jointly launched the "Automotive AI Claw Ecosystem Initiative" with six other companies. Not long before, Horizon Robotics unveiled its cockpit-drive integrated chip "Starry Sky" and the KaKaClaw cockpit system.
It's not just chip companies vying for position. The Li Auto L9 Livis integrates the intelligent agent into the core of cockpit interaction, the AITO M9 also emphasizes its own interactive capabilities, Banma Zhixing introduced the AutoClaw vehicle Agent framework, and ThunderSoft achieved cross-domain invocation of ADAS sensor data by intelligent agents on the Qualcomm platform.
A fundamental technical necessity drives this crowding: for an intelligent agent to function effectively, it must simultaneously access data from both the cockpit and intelligent driving domains. While it sounds simple for an ADAS camera to identify a pedestrian ahead and for the cockpit AI to receive this information to alert the driver, achieving a satisfactory experience has been challenging under the traditional architecture where cockpit and driving functions reside on separate chips. Cockpit-drive integration has evolved from a technical direction to a mass-production necessity.
Once cockpit-drive integration becomes widespread, its impact extends far beyond chip selection. The competitive boundaries among suppliers are blurring, internal organizational structures of automakers are under pressure, and profit distribution within the industry chain is being renegotiated.
Accelerating Volume
Over the past two to three years, the industry has frequently discussed "cockpit-drive integration," but year-end summaries often showed limited progress.
This is partly due to the time lag between chip R&D and product implementation. It typically takes three to five years for a chip generation to materialize into specific supplier solutions. A "generational gap" exists between the surging computing power demands, as city NOA transitions from a premium feature to a standard one—even entering sub-100,000 RMB segments—and the solutions already deployed in the market.
Approximately three years ago, Qualcomm internally concluded that, given the pace of AI expansion, maintaining separate systems for cockpit and ADAS was unreasonable.
Why has this three-year-old judgment only been widely validated this year? Because 2026 is the year several forces converge.
Regarding intelligent agents, Agents impose rigid requirements on the underlying architecture, necessitating real-time, cross-domain access to sensor data. This cross-domain operation is most direct in specific scenarios: an ADAS camera identifies a pedestrian around a corner, and the cockpit AI agent simultaneously accesses this information to alert the driver. This cross-domain invocation is either impossible or poorly executed under a separated architecture.
On June 5th, Qualcomm launched the "Automotive Artificial Intelligence Claw Ecosystem Plan" with partners including ThunderSoft, Banma Zhixing, Desay SV, and others, establishing a framework for the cross-domain operation of intelligent agents. Whoever controls this framework gains the entry point for in-vehicle service distribution.
Another critical factor is cost. When asked about challenges from rising memory prices on June 5th, Nakul Duggal, Senior Vice President and General Manager of Automotive, Industrial & Embedded IoT and Robotics at Qualcomm Technologies, stated that the global memory premium cycle might persist for another 12 to 18 months. The separated architecture, requiring two SoCs each with its own memory set, amplifies BOM costs during this涨价 cycle. He noted that cockpit-drive integration saves the supporting memory for one chip, which was one reason for Qualcomm's initial launch of the Ride Flex solution.
Had memory prices not increased, many automakers might have remained on the sidelines for another year or two. The price surge has compressed the decision-making window.
The叠加 of these two forces is turning cockpit-drive integration from a technical direction into a mass-production reality.
It is understood that since last year, BAIC's Arcfox has implemented single-chip cockpit-drive integration solutions in three models: the Alpha T5, Alpha S5, and Wenda V9. Leapmotor's D19 also debuted this year with a dual Snapdragon 8797 central domain controller. Currently, the Snapdragon Ride Flex platform has secured design wins for 9 vehicle models, while the Snapdragon Automotive Platform Premium Edition has 18 model design wins, with 10 already in or entering mass production.
A significant number of automakers are still观望. The industry has experienced a shift away from in-house development over the past two years, with events like the dissolution of Haomo.ai and the closure of Ji Yue. Automakers are outsourcing both intelligent driving and cockpit functions to external suppliers, but many have yet to determine who will coordinate these two external supply chains. Multiple industry insiders indicate the industry still lacks a convincing "showcase model"—a blockbuster vehicle that allows end-users to clearly perceive that "one chip is better than two."
The period without this showcase will not be long. Whoever achieves it first will define user expectations for this category.
Chain Restructuring
The competition surrounding cockpit-drive integration appears on the surface as a product battle among chip companies, but at its core, it's about how supply chain profits are divided.
Qualcomm adopts an open platform strategy, avoiding software and algorithms itself and leaving the upper layers entirely to partners. Qualcomm does not collect chip-related data; data ownership resides with software stack partners. This positioning bets on the platform being the critical control point. If enough algorithm companies develop solutions on its chips and enough automakers mass-produce using its architecture, that solution becomes the de facto standard.
Nearly all major intelligent driving algorithm companies, including WeRide, Momenta, DeepRoute.ai, QCraft, and ZOYI Technology, have mass-production solutions on the Qualcomm platform.
Notably, Qualcomm introduced the Snapdragon 8787, a new member of the Premium Edition family. This chip directly competes with the two-chip solution of "Snapdragon 8295 plus a competitor's ADAS chip." This dual-chip组合 is currently very common in the 150,000 to 250,000 RMB vehicle segment, using the 8295 for the cockpit and Horizon Robotics' Journey series or other ADAS chips for driving. The 8787 aims to replace these two chips with one, while also supporting the Flex cockpit-drive fusion architecture.
The 150,000-250,000 RMB range is precisely the most concentrated price band for Chinese passenger vehicle sales. If the 8787 gains a foothold here, it will directly impact the market share of intelligent driving chip suppliers like Horizon Robotics in the mainstream price segment. Furthermore, the 8787 and 8797 belong to the same product family with software binary compatibility. Development work automakers do on high-end models based on the 8797 can be directly移植 to the 8787 for mid-range models, enabling one R&D investment to cover the entire price spectrum.
Horizon Robotics bets on integrated hardware and software. Its "Starry Sky" chip is natively designed for cockpit-drive integration, with hardware isolation achieving ASIL-D level. Founder Yu Kai calculates the cost savings: 4,000 RMB saved per vehicle in BOM costs translates to billions over a million vehicles. The deeper logic is that when the chip, algorithm, and operating system come from the same company, performance释放 is most complete. Automakers seeking the ultimate experience would need to buy the entire package. The showcase project chosen is the Chery iCAR, whose success or failure will directly determine this path's persuasiveness within the industry.
NVIDIA chooses alliance. Its Thor platform targets the high-end market, with the Zeekr 9X and 8X already in mass production and搭载. The cockpit side is handed to other partners; for instance, the Zeekr 8X relies on MediaTek's C-X1 to catch up. NVIDIA bets on the computing power ceiling: as long as high-level intelligent driving持续 requires the strongest computing power, Thor maintains an irreplaceable position.
Behind these three choices lie three different judgments about the industry. Who is right ultimately depends on what automakers are truly willing to pay for.
However, the most剧烈 change at present is not among chip companies but at the supplier layer.
ZOYI Technology started with intelligent driving and has now developed mass-production solutions for cockpit-drive融合 domain controllers on both Snapdragon 8775 and 8797, upgrading its strategic cooperation with ChinaTSP during the summit to advance 8797 implementation. ChinaTSP itself切入 from the cockpit side, doing the same thing. Desay SV, Momenta, and DeepRoute.ai are also making similar布局. Currently, over 25 suppliers have developed cockpit-drive integration-related solutions based on the Qualcomm platform.
The integrator role is shifting from the cockpit side towards the intelligent driving side. The logic is simple: intelligent driving has higher functional safety thresholds. Whoever can handle ASIL-D is more qualified to integrate and commands higher profit margins. Traditional cockpit Tier-1 suppliers risk退化 from solution providers to execution layers if they don't upgrade their capabilities, facing大幅压缩 profit空间.
A batch of companies with L4 origins are also加速切入 this track. WeRide won five national intelligent driving competition championships in the past five months and jointly mass-produced the sub-100,000 RMB Aion N60 with GAC, featuring standard city NOA across the lineup. QCraft started with RoboTaxi; its ADAS solutions have累计 over 3.5 billion kilometers of user travel, with an AEB false trigger rate below once per 500,000 kilometers. It is now jointly developing next-gen solutions with Qualcomm based on the higher-compute QAM8797P cockpit-drive fusion platform. DeepRoute.ai's首发 project on the Snapdragon 8797 took仅 six months from development to mass production. These companies first establish a foothold on the intelligent driving side before upgrading to cockpit-drive fusion platforms, making their paths increasingly clear.
Zhao Gang, Vice President of QCraft, stated that from BEV to end-to-end and VLA, the fundamental driver of technology始终 is user demand. Users increasingly recognize higher接管里程 and better safety, and this demand is bringing about an explosion in volume.
However, the crowded track will not last long. Industry consensus is that the landscape of intelligent driving suppliers is加速收敛 to five to seven players. The engineering投入 for融合 solutions is larger; automakers cannot simultaneously engage with over a dozen algorithm suppliers. Competition on fusion platforms is, in fact, accelerating supplier淘汰.
The Future Endpoint
In the immediate term, cockpit-drive integration is a cost-reduction story and a technical prerequisite for intelligent agents entering vehicles. But taking a step back, what is happening is larger than most anticipate.
Currently, an increasing number of suppliers are placing automotive and robotics within the same business unit.
A senior executive from a leading intelligent driving solution supplier indicated that technically, building a car with驾驶辅助 capabilities and building a robot are highly similar. Core capabilities like deploying VLA models, data collection and annotation, and the AI飞轮 are共用. The difference lies in robots having far higher degrees of freedom and operating in more complex environments. However, the automobile is the first载体 for physical AI to achieve scale, and the computing architecture of cockpit-drive integration is becoming the starting point for a much larger landscape.
Consider this from another angle: the market size for cockpit-drive integration cannot be measured solely by the automotive sector's scale. If this architecture can extend from cars to robots, drones, and industrial automation, the total addressable market could be several times larger than that of纯 automotive chips. This also explains why Qualcomm, NVIDIA, and Horizon Robotics are fiercely competing for the right to define this architecture—they are vying not for orders from a single car model but for the entry point to下一代 computing.
Li Dong, CTO of QCraft, pointed out that the industry is at an inflection point, moving from single-scenario无人驾驶 towards general-purpose physical AI, with world models and reinforcement learning serving as the bridge connecting the two. He described QCraft's cloud-based world model, capable of generating highly controllable videos based on motion simulation and BEV layouts, using natural language as a "world editor" to一键合成 long-tail scenarios and extreme weather, supporting持续强化学习 with low-cost closed-loop simulation. He特别强调 this approach can endow AI with defensive driving instincts, "truly achieving safety by preventing issues before they occur."
QCraft's持续投入 in the L4 unmanned logistics field印证 this point. The same "world model + reinforcement learning" architecture can be used to generate adversarial scenarios and持续优化 strategies, whether for complex博弈 in city NOA or高危场景 for unmanned logistics vehicles. The greater the technical通用性, the higher the value of the architecture.
If physical AI is the direction, current chip solutions are already advancing towards it. The dual Snapdragon 8797 cross-domain fusion central computing solution currently offered by Qualcomm runs encoders跨双芯, employs a MoE architecture for VLA, has over 30 billion parameters, and outputs三层冗余 trajectories. This solution doesn't just serve automotive intelligent driving; it simultaneously serves as an infrastructure prototype for deploying physical AI on the edge. Nakul also stated that significant demand will emerge in the edge AI field because any task that can be automated will ultimately be carried by AI models capable of推理 on the edge. The automobile is merely the starting point.
Even this solution might only be a transitional state. The aforementioned senior executive from a leading intelligent driving solution supplier indicated that on the path towards higher-level autonomous driving, the core design constraint for new solutions is no longer TOPS but DDR bandwidth.
When the bottleneck shifts from computing power to data throughput, it signifies that the models to be run are no longer just for inference but involve data-intensive world models and multi-agent scheduling requiring大量端侧运算. The demand for computing architectures is far from settled.
The finish line of this long race is not a single vehicle but the intelligentization of the entire physical world. The course will also be longer than anyone expects.
Comments