ZoS Automotive Research and Shuiqingmuhua Research Center jointly released the "2026 Research Report on SoC Chips for Intelligent Driving".
In terms of autonomous driving system installation, mid-level intelligent driving systems (L2+L2.5) have gradually become standard equipment, with the penetration rate rising from 42.42% in 2023 to 50.89% in 2025. Driven by automakers such as BYD and Geely, the incremental growth of L2.5 (highway NOA) is significant. The penetration rate of high-level intelligent driving systems (L2.9+L3, including hardware pre-installation) increased from 2.18% in 2023 to 14.94% in 2025. From January to May 2026, this penetration rate further grew to 21.3%, as Chinese OEMs are rapidly promoting the mass production and installation of L3 autonomous driving systems in the form of hardware pre-installation.
Forecast of Penetration Rates for Different Levels of Intelligent Driving Systems in China's Passenger Cars from 2022 to 2030
Source: "2026 Research Report on SoC Chips for Intelligent Driving"
National Standards for L3/L4 Autonomous Driving Systems Released, Driving OEMs to Accelerate the Deployment of High-Level Autonomous Driving
On July 30, 2026, as China's first mandatory national standard targeting L3/L4 autonomous driving systems, GB/T 44721-2026 "Intelligent Connected Vehicles - Safety Requirements for Autonomous Driving Systems" was officially released and is scheduled to take effect on July 1, 2027. This standard not only requires vehicle manufacturers to establish assurance capabilities covering the entire process of product design and development, production and manufacturing, and post-deployment, but also puts forward various safety requirements for the execution capability of dynamic driving tasks by the autonomous driving systems.
Meanwhile, the UN Global Technical Regulation for Autonomous Driving Systems (ADS GTR), jointly led by China, the EU, the UK, the US, Canada, and Japan, was also officially approved and released. This regulation clarifies the core technical indicators of autonomous driving system products, simultaneously proposes manufacturer-level requirements along with supporting audit and evaluation methods, and builds a regulatory framework covering the entire life cycle of the product.
Whether it is the world's first global technical regulation for autonomous driving or the domestic mandatory national standard targeting L3/L4 autonomous driving systems, both mark the entry of high-level autonomous driving into the institutionalized stage of commercial implementation.
Source: "2026 Research Report on SoC Chips for Intelligent Driving"
Currently, many automakers have successively launched L3/L4 autonomous driving systems, such as Huawei Qiankun ADS 5.0, Geely Qianli Haohan H9, Chery Falcon 900, and BYD Eye of God Autonomous Driving Edition, among others. At the same time, they are conducting hardware pre-installation for high-level autonomous driving in new models, equipped with high-line-count LiDARs and high-computing-power chips to meet the higher requirements of L3/L4 high-level autonomous driving. Among them, most solutions tend to use high-computing-power chips of 700-2000 TOPS for hardware pre-installation, aiming to reserve space for software upgrades in the next 3-5 years.
Horizon Robotics' next-generation autonomous driving chip Journey 7 (J7) series may adopt the 3nm or 4nm process and is expected to be mass-produced in 2027. Among them, the computing power target of the top-spec version J7P significantly surpasses that of NVIDIA Thor-X.
The Geely Qianli Haohan H9 is equipped with dual NVIDIA DRIVE Thor-U chips. The two chips are mutually redundant at the underlying level, with a total computing power exceeding 1,400 TOPS. It has been installed in flagship models priced above 350,000 RMB, such as Zeekr 9X, Zeekr 8X, and Zeekr 009. It can support a fully redundant L3 architecture, full-scenario D2D, map-free campus roaming, and one-click valet parking summon.
In addition, the team of OEMs developing self-designed chips is also continuously expanding. Li Auto Mach 100 and XPeng Turing AI chips have been installed and mass-produced. Xiaomi Xuanjie D100 and BYD Xuanji A3 made their brand-new debuts, while Volkswagen jointly developed the C7H chip with Horizon Robotics.
Li Auto Mach 100 is an SoC specifically designed to support automotive AD software stacks. In the automotive field, Mach 100 is oriented towards high-level NOA and future L3/L4 capability reserves. The effective computing power of a single Mach 100 reaches 1,280 TOPS, and the effective computing power of a dual-chip combination can reach 5-6 times that of NVIDIA Thor-U. Mach 100 is equipped with 8 LPDDR5X subsystems, providing 64GB of memory and a peak bandwidth of 273GB/s. The MIPI-CSI system supports input from up to 11 cameras, and its Image Signal Processor (ISP) subsystem is responsible for processing raw images for use by the perception models of the NPU.
BYD's self-developed 4nm autonomous driving chip Xuanji A3 has initiated scaled mass production upon its launch. It supports L3/L4 autonomous driving. Through the efficient collaboration of three chips, it achieves a total computing power of over 2,100 TOPS, while taking into account both power consumption control and computing power utilization rate. It is planned to be installed in new Denza models.
A single XPeng Turing AI chip can deliver up to 750 TOPS of computing power. The brand-new mass-produced L4 model XPeng GX can be equipped with up to 4 chips, achieving a total in-vehicle computing power of 3,000 TOPS, which is the highest level among current mass-produced models globally. The technical route adopts a vision-centric scheme supplemented by millimeter-wave radar, and the decision-making latency of the VLA2.0 end-to-end large model is less than 80 milliseconds.
Xiaomi's self-developed high-computing-power AI chip for autonomous driving, Xuanjie D100, made its debut. As the country's first 3nm autonomous driving chip, it features a 20-core high-performance CPU and a 16-core high-computing-power NPU. It supports up to 160GB of memory and up to local deployment of 200B large models, benchmarking against NVIDIA Thor, NIO Shenji, and XPeng Turing.
CARIZON (Volkswagen China's autonomous driving subsidiary) has successfully developed an autonomous driving SoC chip, internally codenamed C7H. Adopting the 3-5nm process technology and designed based on Horizon Robotics' fourth-generation BPU Riemann architecture, a single chip delivers 500-700 TOPS of AI computing power. It can support full-scenario / L3-L4 level autonomous driving under complex urban road conditions and is expected to begin large-scale vehicle installation in 2028.
Comparison of Some High-Level Autonomous Driving SoC Products
Source: "2026 Research Report on SoC Chips for Intelligent Driving"
Fierce Competition in Urban NOA: NVIDIA Thor-U Rapidly Scales Up to Replace Traditional Orin-X Solutions, and Numerous New Products Will Be Installed Successively
The selection of autonomous driving SoC chips installed in L2.9 intelligent driving systems (including hardware pre-installation) is basically high-computing-power chips. In 2025, NVIDIA Orin-X was installed in the most models, followed by Tesla AI 4.0 (self-developed models) and Huawei HiSilicon Ascend series ("Five Realms" models). Since its mass production in 2024, NVIDIA Thor-U has shown very strong growth. From January to May 2026, its market share has risen to the top 1, accounting for 27%. In addition, in 2026, autonomous driving SoC chips such as Qualcomm Snapdragon 8797 (320 TOPS), XHEART BMC X7 (272 TOPS), SiEngine Xingchen No.1 (512 TOPS), and Horizon Robotics J6P have also begun to be installed in vehicles one after another.
Source: "2026 Research Report on SoC Chips for Intelligent Driving"
SiEngine's full-scenario high-level autonomous driving chips Xingchen No.1 (AD1000) and Xingchen No.1 Lite (AD800) achieved mass production in December 2025 and will be applied in vehicles on a large scale in 2026. Meanwhile, Xingchen No.1 obtained the automotive chip information security product certification certificate issued by CATARC Huacheng Certification in March 2026. Xingchen No.1 comprehensively benchmarks against the most advanced international autonomous driving products currently available, and comprehensively surpasses international advanced mainstream products in key indicators such as CPU performance, AI computing power, ISP processing capability, and NPU local storage capacity. A single AD1000 chip can realize L2++ level advanced driver assistance functions, two chips can realize L3 level conditional autonomous driving functions, and four chips can realize L4 level autonomous driving functions.
The chip adopts the 7nm automotive-grade process, complies with the AEC-Q100 standard, and features a multi-core heterogeneous architecture that makes the intelligent driving computing power more powerful: the CPU computing power reaches 250 KDMIPS, and the NPU computing power is up to 512 TOPS. Through multi-chip collaboration, a maximum computing power of 2,048 TOPS can be achieved. In terms of hardware configuration, the AD1000 integrates high-performance VACC and ISP, has a built-in ASIL-D functional safety island, and possesses rich interfaces, which can fully meet the intelligent driving needs from L2 to L4.
In addition, the high-performance NPU architecture of Xingchen No.1 natively supports Transformer large models, fully adapting to the trend of intelligent driving developing towards end-to-end large models. At the same time, the high-computing-power DSP unit provides programmable capabilities for the iteration of customized custom operators.
Momenta, on the other hand, has joined hands with chip partner XHEART to build a synergistic advantage of integrated hardware and software. In December 2023, XHEART was officially established; in 2024, it set up a research and development team of nearly 100 people; in 2026, XHEART officially launched its first autonomous driving SoC -- X7. XHEART has tailor-made a dedicated NPU architecture for X7, specifically designed around BEV+Transformer. It deeply customizes the hardware for core operators such as the Attention mechanism and BEV spatial transformation. The supported large model parameter size can reach about 10 times that of general-purpose chips; a single chip can meet the full-scenario needs of urban NOA functions, directly benchmarking against industry benchmarks.
"Contact Information" Mobile number is the same as WeChat number
Industry Research Department | Mr. Fu 15810027571
Mr. Zhao 18702148304
Data Service Department | Ms. Zhang 13716037793
Strategic Consulting Department | Ms. Han 15810133447
Promotion and Communication Department | Ms. Liao 13718845418
Mr. Du 13910162318