ADAS and Autonomous Driving Tier 1 Suppliers Research Report, 2026 – Chinese Companies
Intelligent Driving Industry Leap: From Widespread L2.9 Adoption toward Physical AI and New Full?Domain Mobility Ecosystem
As vehicle intelligence advances, the installations and penetration rate of ADAS (L1?L2.9) in passenger cars in China have sustained stable growth. ADAS was installed in 9.842 million passenger cars in China in 2022, compared with 6.751 million units in 2021, representing a compound annual growth rate of 20.7%. From 2024 to 2025, the installations increased from 14.559 million cars to 16.231 million cars in 2025, marking a year?on?year growth of 11.5%. In H1 2026, the structure of intelligent driving underwent further reshaping. The penetration rate of L2 and higher?level assisted driving exceeded 72.0%, of which the combined penetration rate of NOA (L2.5+L2.9) reached 37.4% and the aggregate penetration rate of basic assisted driving (L2?L2+) surpassed 34.6%.
Insight 1: New vehicle product definitions have made a leap to L2.9 in advance. Overall structure will rapidly tilt toward L2.9 following model replacements of existing stock vehicles over the next 1?2 years.
According to penetration rates by level of newly launched models across years, L0/L1 low?level intelligent driving solutions have become nearly marginalized in new vehicle model selection. Automakers no longer separately develop discrete low?level solutions in new vehicle project planning. The share of L0 has long hovered close to zero, and the penetration rate of L1 in new vehicles stood below 5% in H1 2026. Basic ADAS covering L2?L2+ and L2.5 solutions remain mainstream configurations for mid?range mass?market models. 2025 saw intensive launches of new L2.5 products, with numerous mid?range models equipped with L2.5 hardware pre?embedded solutions. In H1 2026, the share of new L2.5 products declined moderately as automakers shifted product resources toward L2.9. At the level of new vehicle product definition, the leap to L2.9 has already been completed ahead of schedule. It is projected that over the next 1?2 years, as existing vehicle models undergo iteration and replacement, the overall market structure will keep shifting rapidly toward L2.9.
Insight 2: NOA growth logic undergoes shift: Growth drivers shift from high?end vehicles to entry?level market segments. Highway NOA completed market validation in vehicles in the price range of RMB0?100,000 in 2025, and urban NOA achieved deployment in the same price range in H1 2026.
The core growth logic for the NOA industry has experienced a fundamental shift. The incremental growth pattern once dominated by high?end vehicles has come to an end. The entry?level market segment for vehicles priced below RMB100,000 has emerged as the new core growth engine for intelligent driving. In terms of deployment timeline, highway NOA achieved adaptation and deployment in vehicle models in the price range of RMB0?100,000 in 2025, completing commercial validation in the entry?level market, becoming the third?largest segment by installations of highway NOA, with a penetration rate rising to 9.8% and installations in 162,000 vehicles. Following the penetration of highway NOA into the RMB0?100,000 segment in 2025, urban NOA was deployed in this same price range in H1 2026. Supported by lightweight algorithms for urban NOA and low?cost domain controller solutions with declining hardware costs, urban NOA is gradually becoming a standard capability for entry?level and mainstream vehicle models.
Future competitive focus will shift from “whether high?level intelligent driving can be realized” to “whether low?cost mass?production and deployment in lower price ranges can be achieved”. The adoption of NOA in vehicles priced below RMB100,000 compels Tier1 suppliers and chip suppliers to cut BOM costs. Solution providers with capabilities of low?cost mass production and end?to?end lightweight implementation will take larger market shares in lower-tier segments.
Insight 3: 2025?2026 marks a critical turning point for mass?production shift to L2.9 chips, with new?generation chips commencing deployment in vehicles collectively.
Prior to and in 2024, chip supply for urban NOA was highly concentrated, dominated by NVIDIA Orin?X, Tesla’s self-developed FSD, and Hisilicon Ascend 1951. 2025?2026 opens a core window period for shift between old and new hardware platforms. New?generation chips including Horizon J6M/J6P, Qualcomm SA8650/8775/8797, Hisilicon Ascend 1951 Lite, NIO Shenji NX9031, and Xpeng Turing achieve volume deployment in vehicles. Installed in a total of 2.3308 million vehicles from 2024 through H1 2026, NVIDIA Orin?X remains the core foundation for high?level intelligent driving, yet its product lifecycle has entered the second half. Next?generation platforms such as NVIDIA Thor?U and Orin?Y will address replacement demands, and the industry overall accelerates migration toward new?generation chips.
From the perspective of supplier landscape, domestic intelligent driving chips have formally entered a cycle of large?scale commercial delivery. With installations in a total of 1.126 million vehicles, Hisilicon Ascend 1951 ranks first among domestic urban NOA chip suppliers and third across the whole industry. Horizon J6M follows a differentiated approach, leveraging its 128 TOPS medium?compute solution to support urban NOA deployment. It empowers third?party intelligent driving suppliers to deliver cost?effective urban NOA solutions, and facilitates intelligent driving penetration into lower?tier markets. QCraft is the first third?party intelligent driving algorithm providers in China to realize mass production of urban NOA based on a single Horizon J6M chip. The chip self-development strategies adopted by leading domestic automakers have yielded tangible results: NIO Shenji and Xpeng Turing have both achieved large?scale installation in vehicles. Through independent chip development, automakers bind their vehicle intelligent driving systems and build differentiated product barriers; XHEART BMC X7 completed mass?production validation from Zero to One, further enriching the local chip supply echelons.
In terms of computing power, the 100?260 TOPS medium computing power range constitutes the absolute mainstream for L2.9. From January to May 2026, this computing power range saw installations in 507,700 vehicles, mostly adopting single?SoC architectures. Chips including Horizon J6M (128 TOPS), Orin?Y (200 TOPS), Orin?X (254 TOPS), and Qualcomm SA8650 represent core mass?market offerings. Domain controllers for mid?range vehicles are predominantly configured with computing power within this range. This confirms that the industry no longer blindly pursues ultra?high computing power. Medium?compute, cost?effective single?chip solutions are better suited for volume production and deployment of urban NOA in lower-tier markets.
The domestic passenger car intelligent driving industry has completed phased transformations including large?scale popularization, technological penetration into lower?tier markets, and hardware iteration. L2.9 has become a mainstream standard vehicle feature, lower-tier market segments represent core growth drivers, and domestic chips have achieved large?scale substitution. A competitive landscape in the stock market has taken shape within the industry.
Nevertheless, the intelligent driving industry does not stop at the existing L2.9 mass?production system. Based on research covering strategic plans, product layouts, technical routes, and ecosystem cooperation trends of 24 ADAS Tier1 suppliers, ResearchInChina observes that ADAS Tier1 suppliers are entering a new development cycle spanning from 2025 to 2027. This cycle features technical advancement, scenario extension (L3/L4), segment expansion (commercial vehicles, two?wheelers, embodied artificial intelligence (EAI)), and global overseas layout.
Trend 1: ADAS Tier1 suppliers accelerate deployment of L3 autonomous driving. 2026 serves as a solution verification window, and mass?production and deployment of intelligent driving products will arrive in 2027.
Several core domestic ADAS Tier1 suppliers have completed strategic layouts of L3 high?level intelligent driving. Seven of these suppliers secured L3 vehicle project designations from 2025 to 2026 and have launched full?scale project development. Five enterprises including Jingwei Hirain, Huawei, Momenta, Horizon Robotics, and Joyson Electronics designate 2027 as the core milestone for mass?production, launch and large?scale commercial application of L3 products.
From an industry rhythm perspective, 2026 is a window period of intensive solution verifications the L3 industry. Core industry tasks focus on qualification application, MIIT access review, real?vehicle acceptance testing, and road pilot implementation. 2027 will mark the first year for large?scale commercial adoption of L3 intelligent driving in China. Leading suppliers maintain highly aligned mass?production timelines, delivering markedly improved industry certainty.
In terms of deployment progress, the first echelon is represented by Huawei. Its supported vehicle models have obtained MIIT L3 access permit and entered road pilot trials across multiple cities, making Huawei the solution supplier with the fastest deployment progress in China. The second echelon include Jingwei Hirain, Momenta, Horizon Robotics, and DeepRoute.ai. These suppliers have secured OEM designations and are in development or real?vehicle testing phases without yet obtaining access permits. They will submit access applications primarily during 2026?2027.
Regarding L3 intelligent driving product layouts, traditional Tier1 suppliers such as Jingwei Hirain primarily deliver domain controller and sensor hardware for L3 without developing full?stack algorithms, relying on external algorithm partners. Native intelligent driving solution suppliers including Huawei, Momenta, Horizon Robotics, DeepRoute.ai, and Afari Technology deliver hardware?software integrated offerings with self?developed algorithms. Their L3 and 2+/L4 systems share the same technical foundation and reuse a data closed loop.
Consumer?oriented commercial deployment of L3 intelligent driving needs to pass through three successive phases. Phase 1 covers local L3 road testing limited to internal technical validation, requiring safety officials on?board and prohibiting commercial sales to end users. Phase 2 covers road?traffic pilot operations. Enterprises must obtain conditional product access permits from MIIT and dedicated L3 license plates issued by local governments. This phase opens only for business operating entities and excludes retail sales to individual consumers. Phase 3 represents full national access. Formal MIIT L3 product access permits are required before sales to private individuals. No such permits have been issued to date. Gradual liberalization will follow enforcement of the mandatory national standard GB 44721?2026 in 2027. Possession of access qualifications does not authorize unrestricted nationwide operation. L3 functions may only activate within designated road segments and speed limits defined by pilot cities.
Up to now, Deepal SL03 and Arcfox αS are the only two domestic vehicle models that have obtained MIIT L3 conditionally autonomous driving product access (Announcement Batch 401, December 15, 2025). The vehicles are undergoing “product access + on-road use pilot”. Neither model is retailed to ordinary private consumers. They are operated by mobility service affiliates of automakers, with strict operational limits on designated road sections, ODD ranges, and vehicle speeds.
Trend 2: Following large?scale deployment of urban NOA, intelligent driving suppliers target Physical AI as the next key breakthrough area. L4 autonomous driving represents the optimal entrance to Physical AI and is also the first half of Physical AI industrialization.
Physical AI refers to enabling AI to step out of the screen and be embedded in physical entities (such as cars, robots, and drones) to perceive, understand, and change the real physical world. It requires AI to be embodied: relying on carriers including autonomous vehicles and general?purpose robots to directly perceive real?world environments, develop physical intuition, and make autonomous decisions to execute physical actions, completing a full closed?loop from perception to action.
Autonomous vehicles will emerge as benchmark applications for this paradigm. They will evolve into super onboard agents capable of understanding voice commands and completing multiple tasks end?to?end. Leveraging multimodal perception together with linkage between vehicle status and environmental data, they anticipate user intentions prior to explicit requests and achieve full integration of cockpit and driving functions. At this stage, automobiles will break away from their definition purely as transportation tools and transform into full?scenario intelligent hubs connecting mobility, office work, home life, and third?party ecosystems.
Underlying logic positioning L4 as the optimal gateway to Physical AI stems from the fact that the automotive industry has already completed partial engineering pathways for Physical AI.
First, the full intelligent vehicle industry chain is mature. Any physically capable autonomous agent requires complete hardware links covering perception, computation, and chassis control. Automotive chassis, sensors (like lidar and radar), automotive?grade high?compute chips, and intelligent driving domain controllers are highly standardized and support rapid deployment of tens of thousands of physical agents. Humanoid robots, by contrast, remain in a critical stage of development of robots, reducers, and dexterous hands, with high hardware costs and reliability insufficient for volume production.
Second, autonomous driving represents the sole Physical AI segment that enables “data closed?loop + commercial closed?loop”. Two major sore points hinder Physical AI implementation. One is the shortage of massive diverse real?world physical interaction data essential for iterative model training. The other is a viable commercial closed?loop. Large?scale regular deployment cannot be sustained through perpetual costly testing alone. Compared with humanoid robots, autonomous driving offers natural data closed?loop advantages. Intelligent vehicles themselves serve as mobile data collection terminals. With penetration rates of L2 and higher?level intelligent driving vehicles exceeding 70% in China, vehicle travel continuously generates massive volumes of real?world scenario data. Around-the-clock operation of Robotaxi, Robobus, and unmanned logistics delivery fleets feeds back diverse road condition samples, refining models for physical causal reasoning. This forms a complete link: real?road operation → collection of physical interaction data → cloud simulation training → model deployment onto vehicles → renewed on?road data collection.
Additionally, autonomous driving features controllable task complexity and structured scenarios, focusing solely on planar movement to address the problem of mobile agent subsets. Its core capabilities include travel, obstacle avoidance, and traffic participant prediction, with outputs of steering, braking and throttle control commands. Fine-grained grasping and manipulation of flexible objects are not required, and the mechanical dynamics of the vehicle remain limited in complexity. By contrast, humanoid robots represent the second half of EAI. Equipped with dozens to hundreds of joints, they need to perform actions such as grasping, twisting and pressing, which involve full-body three-dimensional dynamics and tactile interaction. Their task complexity rises exponentially. Furthermore, road environments are relatively structured and constrained by traffic rules, lane markings and traffic lights, unlike highly unstructured scenarios such as homes and wild fields.
Compounded by large?scale deployment of intelligent driving large models in vehicles, if massive datasets fuel autonomous driving as the first half of Physical AI, end?to?end intelligent driving AI large models serve as core engines converting data into practical value. Early autonomous driving implementations depended on hand?written rules and modular development, with perception, prediction, and planning & control operating as independent modules. This block?style approach delivers weak generalization capabilities and struggles to migrate onto general?purpose robot carriers. Intelligent driving large models such as End?to?end, VLM, VLA and world model began to be implemented in 2025. Multiple technical paradigms converge and reshape industry logic in 2026. Each approach carries distinct advantages and disadvantages bounded by the data?computing?power?safety triangle constraint. Players explore differentiated solutions based on their available resources.
Finally, autonomous driving has a well-defined regulatory and safety verification pathway. Mature global grading regulations for autonomous driving have been established, enabling phased deployment starting from closed campuses and urban secondary trunk roads, with safety validation completed step by step. In contrast, safety standards, liability definitions and relevant regulations for humanoid robots in open environments are largely blank, and they do not yet possess the fundamental prerequisites for large-scale public deployment.
Trend 3: ADAS Tier1 suppliers make cross?border inroads into EAI, and pursue breakout growth via technical homology and tap into second?curve business opportunities.
Automotive ADAS Tier1 suppliers are undergoing a second industrial migration and capability reconstruction. Amid stock?market saturation, intensifying price competition, and pressure from self-development by automakers squeezing supply chain profit margins, leading local Tier1 suppliers collectively initiate horizontal migration from “vehicle intelligent assets to general?purpose physical intelligence assets”.
Research examining strategies and product layouts of 13 domestic ADAS Tier1 suppliers within EAI sectors identifies five primary directions for product development: hardware products (domain controller architectures), unified AI foundation models (one underlying algorithm model empowering multiple vertical scenarios), Physical AI infrastructure providers (data solution providers/cloud brain + simulation training + AI models/computing power (robot chips) + cerebrum and cerebellum models + RDK?series development platforms), robot components (sensors, joint modules, dexterous hands, robot head assemblies, energy), and robots (charging robots, data collection robots, business-oriented humanoid/wheeled robots for flexible operation).
First, robot domain controllers have become a core strategic focus for ADAS Tier1 suppliers entering the EAI sector. With mature expertise in automotive-grade technology, multiple players have launched standardized, mass-producible robot hardware products.
Desay SV launched AI Cube, its robot intelligent base, in November 2025. It delivers three differentiated computing power configurations adaptable to industrial, household, and automotive scenarios. The product has completed iteration and formally entered the mass?production and delivery phase.
Jingwei Hirain completed strategic brand upgrading, repositioning itself from “electronic system & technology service company” to “intelligent technology company”. It expands core businesses comprehensively around Physical AI and leverages mature automotive?grade domain controller technical capabilities to enter EAI sectors. It focuses on robot electronic and electrical hardware bases and system integration businesses, and does not prioritize in-house development of large models and world models. It plans to launch a robot cerebrum domain controller product based on Huixi Guangzhi R1 in 2026, with maximum computing power reaching 500 TOPS.
IMotion iteratively released its new?generation robot main controller iRC600 supporting rapid end?side AI?model deployment. Its previous?generation product iRC100P has achieved stable mass?production and delivery.
TZTEK established wholly?owned subsidiary Stellarmind to deepen engagement in the Physical AI sector. Focusing on the R&D of end-side Physical AI infrastructure, it has launched the Star Series edge intelligent computing units, forming a matrix covering all computing power ranges: Star L is a 100 TOPS low-compute version adapted for quadruped robot scenarios; Star M is a 200–500 TOPS medium-compute version targeting wheeled robots; Star H is a 500–1000 TOPS high-compute version precisely tailored to advanced requirements of humanoid robots.
Beyond hardware domain controllers, Physical AI foundation (data + algorithms + computing power) is the second core layout direction for the industry. Data, algorithms, and computing power jointly form three cornerstones enabling Physical AI implementation. Leveraging long-term ecosystem expertise and engineering deployment experience in the automotive sector, Huawei and Horizon Robotics are building comprehensive Physical AI underlying infrastructure covering robot computing power, scenario data, cloud intelligent brains, and robot cerebrum-cerebellum layered architecture. They open up standardized underlying platforms to ADAS Tier1 suppliers and robot manufacturers across the industry, assuming the role of infrastructure providers within the industrial chain. This creates an industrial landscape featuring "collaboration and complementarity with partial competition" alongside the hardware foundation businesses of Tier1 suppliers.
Specifically, based on the computing power of its Sunrise series chips, Horizon Robotics’ supporting robot development stack extending upwards, data toolchain and cerebrum-cerebellum model system provide standardized edge computing power and software development services for humanoid robot manufacturers, substantially cutting customers’ costs of hardware adaptation and model migration.
Huawei has built a complete layered technical architecture of “cloud brain (simulation + model training) + robot cerebrum (Pangu Models) + robot cerebellum”, covering the entire industrial chain of computing power, data and algorithms. NavInfo focuses on agent data services and develops an integrated data solution spanning from data collection to full-chain compliance control, filling gaps in Physical AI data deployment.
The third major layout direction is core components for EAI. Manufacturers continuously expand their product portfolios by extending technologies from automotive EEA to achieve full coverage of core robot components. In the perception component field, Baolong Automotive, as a representative player, develops multiple types of critical robot sensors for perception, force sensing and position detection to meet the fundamental requirements of robots for environmental perception, posture recognition and precise localization. In the actuation component segment, Joyson Electronics leverages mature R&D and mass-production experience in vehicle body actuation systems to expand into core actuation parts such as robot head assemblies and TeleHand dexterous hands. These products are designed to address key problems of robots including bionic interaction, fine manipulation and anthropomorphic motion, and complete the hardware system of the robot actuation layer.
Trend 4: The pace of deploying commercial vehicle NOA quickens, and Tier1 suppliers capture incremental opportunities in trunk?line intelligent driving leveraging technical homology between passenger cars and commercial vehicles.
Starting from January 14, 2026, the Announcement No.21 issued by the Ministry of Transport revised four operational vehicle standards including the Safety Specifications for Commercial Bus. Newly qualified models of commercial buses, trucks and hazardous goods transport vehicles applying to enter the road transport market must be equipped with active safety devices such as AEBS and LDWS. New models of operating commercial vehicles in China have officially entered an era where ADAS is mandatory standard configuration.
Meanwhile, commercial vehicle NOA is advancing from trials to mass production. The commercial rollout of commercial vehicle NOA hinges on whether transportation is safer, drivers feel more relaxed, and operating costs are lower. Mandatory ADAS configuration directly addresses the requirement for "enhanced safety", while NOA further enables "reduced driver burden" and "lower operating costs" on this basis. Mandatory ADAS installation acts as a compliance threshold, whereas NOA serves as an efficiency tool.
Case Study of Zhuoyu Technology’s Commercial Vehicle NOA Mass Production:
From a strategic perspective, Zhuoyu Technology’s core strategy for 2026 is a comprehensive transformation from an "intelligent driving Tier1" to a "Mobile Physical AI company". Its self-developed native multimodal large model technology will constitute a core component of Zhuoyu’s mobility foundation. That is, the multimodal foundation model enables autonomous mobility for diverse mobile platforms across multiple vertical sectors, scenarios and countries. It not only supports L2–L4 passenger cars, commercial vehicles, logistics vehicles and Robotaxis, but also adapts to various mobility scenarios for embodied robots.
From a manufacturing perspective, the Changzhou factory represents a critical step for Zhuoyu’s "Intelligentize All Mobility" strategy to accelerate the intelligent transformation of its commercial vehicle business. Officially commissioned in July 2026, the Changzhou factory establishes a dual manufacturing base layout alongside the Shenzhen facility. The new factory mainly undertakes the manufacturing and inspection of intelligent driving components, including visual perception systems, domain controllers and LiDAR. Phase 1 has a planned capacity of 100,000 sets of intelligent driving domain controllers, and Phase 2 targets 500,000 sets. The long-term employee headcount will exceed 100. Zhuoyu expanded the Changzhou factory to broaden the boundaries of its intelligent driving business, migrating intelligent driving capabilities accumulated in the passenger car sector to diversified scenarios such as heavy-duty trucks, buses, unmanned logistics vehicles and Robotaxis. As business expands across multiple scenarios, manufacturing capacity needs to be expanded in tandem. At present, passenger cars remain Zhuoyu’s core mass-production business, and the new capacity mainly targets incremental business opportunities in commercial vehicles and autonomous vehicles.
From a business perspective, Zhuoyu Technology has formed partnerships with all six top heavy-duty truck brands in China, including FAW Jiefang, Sinotruk, Shaanxi Automobile, XCMG and Foton Motor. Mass production of its first heavy-duty truck has started, with dozens of vehicle models which are to be produced from June 2026 to the first half of 2027. The company is also accelerating overseas projects. It collaborates with Sinotruk on a commercial vehicle intelligent driving project for the European market, which was showcased at the IAA Transportation Hannover.
Looking at the overall development landscape of China’s automotive intelligent driving industry, 2024–2026 marks a pivotal transition period, during which the industry shifts from large-scale popularization toward higher-level functionality, universal accessibility and diversification. In the short term, L2.9 will become a mainstream standard intelligent driving configuration for passenger cars. Lower-tier markets serve as the core growth engine for NOA. Domestic intelligent driving chips achieve large-scale substitution, and low-cost, highly adaptable medium-compute solutions become the market mainstream. Industry competition has completely moved away from the high-end premium pricing model and turned into a contest over mass-production capability and cost control.
In the medium and long term, industrial boundaries keep expanding. Technologically, the industry advances from L2 driver assistance to L3 high-level autonomous driving and L4 full-domain autonomous driving. Relying on the underlying autonomous driving technology foundation, it enters new segments of Physical AI and EAI. In terms of market coverage, the business scope extends beyond the single passenger car scenario to diversified fields covering trunk logistics for commercial vehicles, unmanned delivery and intelligent robots, unlocking the second growth curve for the industry. With continuous improvement of L3/L4 policies and standards, constant breakthroughs in domestic technologies and deeper industrial chain collaboration, China’s intelligent driving industry will break free from the single attribute of traditional automotive hardware, gradually evolve into a full-domain intelligent mobility ecosystem, and complete the core transformation from vehicle intelligence to industrialized Physical AI, gaining competitive edges in the global intelligent transportation and EAI sectors.