The mainstream automotive market is becoming the primary battleground for competition in cockpit chips, and the rules of this competition are changing.
According to the latest configuration data from Gasgoo Research Institute, in the first three quarters of 2026, SemiDrive's cockpit chips (covering central control hosts and cockpit domain controllers) achieved an installation volume of 353,000 units in the domestic passenger vehicle market priced between CNY 100,000 and CNY 300,000. This ranked first among mainland Chinese local suppliers, accounting for approximately 26% of the total installations by local enterprises under this statistical scope.
This achievement is built upon the mass production foundation of existing cockpit products. Today, large models are accelerating their entry into automotive cockpits, and more complex computing tasks are reshaping OEM requirements. From satisfying traditional cockpit functions to supporting on-device AI execution, chip companies need to solve not only performance upgrades but also how to control the incremental system costs.
For SemiDrive, which has already formed a significant installed base in the mainstream market, the next round of competition begins here. More evidently, SemiDrive's reserves have been positioned ahead of the starting gun.
What Does Scale Installation Behind 300,000+ Units Mean?
The number of vehicles a single chip can enter often depends on many factors beyond performance parameters.
Especially in the mainstream automotive market, cockpit chips must not only meet daily functional requirements but also adapt to vehicle development cycles, software compatibility demands, and long-term supply arrangements. For high-volume models, cost differences per chip become an undeniable factor as delivery scales increase.
These requirements are reflected in SemiDrive's specific supply chain integration.
In the central control host sector, SemiDrive has entered the flagship models of SAIC Volkswagen, Dongfeng Honda, Chery, and other automakers. According to the latest configuration data from Gasgoo Research Institute, in January–September 2026, the related chip supply volume for SAIC Volkswagen Lavida alone exceeded 110,000 units.
For such models, rather than solely pursuing higher computing performance, factors such as chip cost, operational stability, and long-term software maintenance capability equally influence OEM choices.
As cockpit architectures consolidate towards domain control, the tasks undertaken by chips become more complex.
In models such as Chery Jetour X70 PLUS, Jetour Dasheng, and BAIC BluePark ARCFOX, SemiDrive has achieved scaled supply of cockpit domain controller chips. Compared to central control hosts, cockpit domain controllers typically need to coordinate more display, interaction, and in-vehicle application tasks, placing different demands on computing resource allocation and hardware-software synergy.
From central control hosts to cockpit domain controllers, the expansion of application scope means that chip suppliers must adapt to more vehicle platforms. Different automakers vary in system architecture, operating systems, and feature configurations, requiring corresponding software adaptation and system verification work to be completed in specific projects.
According to public information from SemiDrive, its cumulative shipments of all series of automotive-grade chips have exceeded 15 million units, covering over 170 vehicle models. Continuous participation in vehicle development and mass production delivery has enabled SemiDrive to accumulate experience in cross-platform adaptation, customer collaboration, and engineering validation.
With large models entering automotive cockpits, existing computing resource configurations face new challenges: when traditional infotainment systems need to handle more complex AI tasks, is the original computing power still sufficient?
Large Models Onboard: How to Calculate the Computing Power Equation?
In the past, cockpit voice interactions mainly revolved around specific functions such as navigation, air conditioning, and entertainment. As large models enter the cockpit, cars begin to understand more complex natural language requests, such as planning itineraries including charging, dining, and rest, or completing multiple settings through continuous dialogue.
Such features are accelerating their entry into the mainstream market. Taking the new Deepal S07 AI Laser Edition priced at the CNY 150,000 level as an example, its equipped Doubao large model supports natural language itinerary planning, complex voice interaction, and other functions.
It should be noted that according to data from the China Passenger Car Association (CPCA), in the first three quarters of 2026, passenger vehicles priced between CNY 100,000 and CNY 200,000 accounted for 41.9% of the domestic retail market, representing the largest single price band by sales volume. As AI features enter this market, automakers must continuously upgrade intelligent experiences while controlling the resulting increases in hardware and development costs.
Users perceive changes in interaction methods, but automakers must decide where these functions are computed.
Tasks such as complex knowledge Q&A and content generation can leverage cloud resources; functions requiring rapid response, processing of local information, or reduced network dependency are better suited for on-device execution. The division of labor between edge and cloud directly impacts the configuration of local computing power.
Even if it is determined which models need to be deployed on the device side, the computing platform cannot be judged solely by NPU computing power.
Model execution requires memory to store parameters, inference processes rely on memory access bandwidth, and functions such as navigation, entertainment, and multi-screen display share resources with AI tasks. When automakers increase local computing power, storage, power consumption, heat dissipation, and system costs usually need to be re-evaluated.
At the current stage, one solution is to add an independent AI Box outside the existing cockpit SoC to specifically undertake model computation tasks. This approach is suitable for upgrading certain mature platforms, allowing retention of the existing cockpit architecture, reducing modifications to the original system, and facilitating relatively independent iteration of the AI computing module. However, the new device requires corresponding support for memory, power supply, heat dissipation, and software coordination.
For newly developed cockpit platforms, automakers can also choose SoCs with higher integration levels, consolidating traditional cockpit computing and AI inference onto a single chip. This offers the opportunity to reduce peripheral hardware but requires the chip to possess stronger concurrent processing, bandwidth allocation, and power control capabilities.
The two technical routes each have applicable conditions. For automakers, choosing a cockpit computing platform requires comprehensive consideration of investments in chip procurement, peripheral components, software development, and system verification, rather than merely comparing computing power parameters.
For chip suppliers, competition thus extends from providing higher peak computing power to how to support the deployment of more AI features in vehicles at a reasonable system cost.
From X9 to X10: What Is SemiDrive's Answer?
Facing changes in cockpit computing needs, SemiDrive has established a product layout covering different tiers.
From the X9 series, which introduced NPUs early in the product definition phase, to the X9SP, which boosted NPU computing power to 8 TOPS, and further to the X10 targeting on-device large models, its products do not simply represent generational replacements but correspond to different levels of computing demand.
Among them, the X10 needs to handle more complex tasks: ensuring normal operation of traditional cockpit functions such as multi-screen display and infotainment while running on-device large models. This requires the chip to not only possess higher AI computing power but also reasonably allocate computing resources needed for different tasks.
According to product parameters published on SemiDrive's official website, the X10 adopts a 4nm automotive process, achieving CPU performance of 300 KDMIPS, GPU performance of 3300 GFLOPS, and NPU computing power of 80 TOPS. Its memory bandwidth reaches 154 GB/s, supporting on-device deployment of large models with up to 9B parameters, and featuring architectural optimization for dynamic expert routing in MoE models.
Among these parameters, memory bandwidth is particularly noteworthy. Large model inference requires frequent reading of model parameters; if data transmission cannot keep up, computing units may fail to fully utilize their performance due to waiting for data. Especially when AI inference runs concurrently with tasks like multi-screen display and infotainment, memory bandwidth and resource scheduling capabilities directly impact system efficiency.
Therefore, the X10 does not simply increase NPU computing power but supports large model inference while balancing traditional cockpit applications through the synergy of CPU, GPU, NPU, and memory systems.
Beyond computing capability, system integration is also a key design direction for the X10.
As mentioned earlier, automakers can enhance on-device computing capabilities by adding external AI Boxes. The X10 integrates cockpit computing and large model inference into a single chip, replacing the traditional "cockpit SoC + external AI Box" combination scheme under conditions that meet corresponding functional requirements.
According to calculations based on SemiDrive's published schemes, compared to traditional combined architectures, the X10 can reduce DDR storage capacity by approximately 25%, potentially lowering the overall system BOM cost by CNY 1,500–3,000. Meanwhile, through process and architectural optimization to control power consumption, the X10 supports air cooling, reducing reliance on complex cooling systems.
However, hardware integration does not mean software development work can be omitted. Deploying large models onboard still involves model porting, operator adaptation, performance tuning, and synergy with existing cockpit software.
Around the X10, SemiDrive provides standardized model invocation interfaces, pre-adaptation for mainstream open-source models, and customized optimization support to reduce repetitive work during development. Across different platforms, SemiDrive also allows customers to conduct prototype development based on the X9 before migrating to the X10, reusing some existing development results.
From computing resource synergy and system integration to software adaptation, the product design of the X10 is no longer limited to improving the performance of a single chip but further considers the deployment efficiency and system cost of the entire cockpit computing solution.
To date, the X10 has entered the commercial implementation phase. According to previously disclosed information from SemiDrive, the X10 has reached mass production cooperation agreements with multiple automakers.
The Competitive Boundary of AI SoCs Is Expanding Beyond Automotive
Beyond the new generation of AI cockpit chips, SemiDrive's product layout has begun to extend to intelligent terminals outside the automotive sector.
Embodied intelligence and smart cars share certain technological commonalities in multimodal perception, model inference, and real-time control, providing space for automotive chip companies to expand into robotics business.
During the 2026 Beijing Auto Show, SemiDrive officially announced its Strategy 2.0 upgrade, expanding its business layout from automotive intelligence to general-purpose intelligence, and launched the R1, D9, and E3-R series chips, targeting large model inference, motion planning and real-time control, and torso/joint actuation control needs respectively.
Unlike automotive cockpits, robots must not only complete perception and model computation but also convert computational results into specific actions, placing new requirements on motion control, real-time response, and synergy between different computing units.
Chip manufacturers therefore need to provide more complete controller solutions and software development support.
The RoboX D9M, jointly launched by SemiDrive and Radxa Computer, represents concrete progress in this layout.
This robot cerebellum controller, based on the D9-Max, provides reference designs and an open-source software stack in addition to hardware, helping developers reduce underlying system construction work.
Notably, SemiDrive's official website discloses that its D9 series chips have already achieved mass production with leading embodied intelligence manufacturers.
In terms of collaborative ecosystem, SemiDrive has also conducted joint development with Apache Robotics regarding cerebellum controllers and joint modules, and cooperated with GALBOT in areas such as robot chips, control systems, and toolchains.
From scaled installation of existing cockpit chips, to X10's computing and system integration for on-device AI, to robot controllers and development ecosystems, SemiDrive's product applications are continuously expanding.
Competition among chip companies has consequently extended from the performance and cost of single products to computing architecture, software toolchains, and adaptation capabilities across different scenarios.
For SemiDrive, scaled installation accumulates not just shipment volumes, but also product development, software adaptation, and engineering delivery capabilities. These capabilities are extending from automotive cockpits to more intelligent terminals, providing a foundation for SemiDrive to expand its business from an automotive-grade chip supplier to a computing platform provider.