As L3-level high-level autonomous driving accelerates toward mass production, the industry has officially entered a new dawn of commercialization. The intelligent upgrade of complete vehicles drives full-chain hardware iteration, increased usage per vehicle, and value reassessment, ushering in adeterministic high-growth cycle for the intelligent driving hardware sector.

According to estimates, the domestic market size for L2 and above intelligent driving hardware will reach RMB 460.8 billion by 2029, with a net increase of over RMB 150 billion compared to 2026. By sub‑segment, the top five hardware growth tracks over the next three years are: chips, cameras, brake‑by‑wire, active suspension, and LiDAR. Ranked by compound annual growth rate (CAGR), the order is: intelligent driving chips, LiDAR, cameras, brake‑by‑wire, and steer‑by‑wire. Each sub‑segment faces structural growth opportunities, with intelligent driving chips showing the most prominent growth elasticity.

From an industry chain perspective, the intelligent driving sector has formed a mature "perception‑decision‑execution" hardware supply chain system, with each link presenting structural opportunities: On the perception side, cameras, LiDAR, and 4D millimeter‑wave radar continue to penetrate thanks to multi‑sensor fusion solutions, while the pace of domestic substitution accelerates across the board. On the decision side, the demand for high‑level autonomous driving computing power continues to surge, with SOC chip installations rising rapidly, leaving ample room for domestic chip manufacturers to break through. On the execution side, with the implementation of new national standards, chassis brake‑by‑wire, steer‑by‑wire, and active suspension systems are iterating faster, and the penetration rate of high‑end configurations continues to rise, opening up entirely new incremental space.

At the same time, downstream Robotaxi unmanned mobility formats are reaching a commercial turning point, with large‑scale operation in multiple scenarios driving upstream hardware upgrades and cost reductions, further consolidating the long‑term growth logic of intelligent driving hardware. Overall, the large‑scale deployment of L3 high‑level autonomous driving is pushing the industry from functional iteration toward hardware value re‑evaluation, and the entire industry chain is entering a golden growth window.

01 Key Intelligent Driving Directions to Watch: Smart Driving Chips, Cameras, Brake‑by‑Wire, etc.

Our calculations show that over the next three years, the incremental market size for intelligent driving hardware will reach RMB 5 trillion. The top growth areas include cameras, brake‑by‑wire, active suspension, and LiDAR; the CAGR ranking is: chips, LiDAR, cameras, brake‑by‑wire, and steer‑by‑wire.

02 Intelligent Driving Industry Overview: Continuous Technological Iteration, L3 Commercialization Around the Corner

2.1 Industry Classification and Development Status

The industry standard classifies intelligent driving into six levels from L0 to L5, representing a complete evolution from manual emergency assistance and basic driver assistance to fully high‑level autonomous driving.

02 Intelligent Driving Industry Overview: Continuous Technological Iteration, L3 Commercialization Around the Corner

2.1 Industry Classification and Development Status

The industry standard classifies intelligent driving into six levels from L0 to L5, representing a complete evolution from manual emergency assistance and basic driver assistance to fully high‑level autonomous driving.

In terms of core architecture, all intelligent driving systems consist of three essential layers—perception, decision, and execution—with clear division of labor and close linkage: the perception layer relies on cameras, various radar types, and other sensors to collect environmental information around the vehicle and convert it into electronic signals; the decision layer uses intelligent driving chips to process and analyze perception data and output corresponding driving decisions; the execution layer relies on braking, steering, and other chassis components to implement decision commands and complete vehicle maneuvering. Currently, the penetration rate of intelligent driving in China is rising rapidly, and intelligence has become a standard feature for new vehicles. In 2025, the penetration rate of L2 and above intelligent driving functions in new domestic vehicles reached 71%, with 75% of new energy vehicles listing intelligent assisted driving as standard equipment.

According to Zoss Automotive Research data, the share of domestic models equipped with L2 and above ADAS functions increased from 34.8% in 2022 to 64.9% in 2025, while the share of high‑level L2++ models rose from 3.1% to 14.3%. The continued increase in penetration is driven by three main factors: first, the maturity of the upstream industrial chain and the continuous decline in hardware costs, significantly lowering the threshold for vehicle intelligence; second, ongoing algorithm iteration and optimization, markedly improving the practicality and ride comfort of intelligent driving functions; third, the implementation of new national standards that incorporate high‑level assisted driving functions into new‑car safety rating systems, forcing OEMs to equip L2 and above intelligent driving functions to obtain high safety ratings.

2.2 Technological Evolution: From Rule‑Based Modular Architecture to AI‑Driven End‑to‑End Integration

The overall technological trajectory of intelligent driving follows the path of "manual rules → deep learning → end‑to‑end large models." The core shift is from a layered modular architecture to an unbounded, unified neural network architecture that completely breaks down the barriers among perception, decision, and execution.

2.2.1 Traditional Technology: Three‑Layer Modular Architecture with Obvious Limitations

Before the popularization of deep learning in 2012, autonomous driving algorithms relied entirely on manually designed rules. Engineers manually wrote feature recognition equations and massive "if‑else" decision logic, which could only recognize preset objects and handle fixed scenarios, failing to cover complex road conditions and long‑tail events. System stability was extremely poor, making large‑scale commercial deployment impractical.

In 2012, the advent of CNN (convolutional neural networks) fundamentally changed the path of intelligent driving technology. This algorithm could autonomously learn object features through training and automatically optimize extraction weights, eliminating the need for manual rule presets and significantly improving object recognition accuracy and system robustness, laying the technical foundation for large‑scale application. However, the CNN approach still had three major shortcomings: it only supported 2D image capture and relied on unstable algorithms to estimate 3D depth information; multi‑camera fusion was difficult and depended on manual calibration and stitching, leading to overlapping area conflicts and missed detections; and it could not accurately determine the relational associations among objects within an image.

In 2017, the Transformer algorithm emerged. Its core self‑attention mechanism can globally correlate information across all regions of an image, perfectly addressing the issue of local information isolation inherent in CNNs. Based on this technology, Tesla introduced the BEV (Bird's Eye View) solution in 2021, which fuses features from multiple cameras to convert 2D images into a top‑down 3D perspective of the vehicle's surroundings, accurately identifying object distance and speed, thus overcoming the fusion difficulties of traditional 2D vision solutions.

Nevertheless, all the aforementioned layered modular architectures ultimately share unavoidable defects: inherent errors in information transmission between layers are amplified as they are passed along, leading to cumulative decision biases; the different objective priorities of perception, decision, and execution modules make it difficult to balance safety, comfort, and handling requirements; and core decision‑making still relies on manual rules, which cannot cover the infinitely complex real‑world driving scenarios.

2.2.2 Future Technology: End‑to‑End Large Models as the Ultimate Direction

In 2023, industry research papers formally established the end‑to‑end autonomous driving paradigm, which has become the core technology roadmap for the next generation. This approach abandons manual layered design and relies on neural networks trained on massive amounts of human driving data, combined with reinforcement learning, to achieve driving decisions comparable to or even surpassing human drivers. The entire system is managed by a unified neural network that directly receives raw sensor data and outputs vehicle control commands, completely eliminating the information loss, objective mismatch, and insufficient scenario coverage of modular architectures.

On this basis, the industry has further developed two advanced paths—VLM/VLA large models and World Models—to continuously optimize end‑to‑end capabilities:

VLM (Visual Language Model) focuses on enhancing visual semantic understanding, while VLA (Vision‑Language‑Action) models add action awareness capability on top. These can be deployed in two ways: first, on‑vehicle direct deployment, which accurately interprets complex traffic conditions and efficiently handles long‑tail scenarios, but demands extremely high computing power; second, lightweight auxiliary applications, used only for handling special scenarios or assisting model training, striking a balance between performance and computational cost.

World Models are a more advanced end‑to‑end architecture, whose core is to predict future environmental changes based on historical data and real‑time observations. Currently, they are mainly used to generate various extreme and long‑tail scenario data for iterative training of autonomous driving models; in the future, they will be deployed directly on vehicles to make proactive decisions by anticipating dynamic changes in road conditions. Their advantages include extremely low information loss, better adherence to physical laws, and higher prediction accuracy; the only drawback is the need for ultra‑high computing power on the order of thousands of TOPS.

2.3 Comparison of Autonomous Driving Policies: China Prioritizes Safety, the U.S. Embraces Innovation

The regulatory systems for autonomous driving in China and the United States exhibit distinct differences, each tailored to the development pace of its respective industry.

The U.S. has established a unified federal regulatory framework, with policies focused on encouraging innovation and relaxing deployment restrictions. Early regulations aimed to simplify hardware requirements; in 2025, the U.S. further relaxed autonomous driving safety standards, allowing R&D and demonstration autonomous vehicles to be exempt from certain human‑driving‑related regulations. The proposed SELF DRIVE Act (expected in 2026) establishes the principle of federal regulatory primacy to eliminate state‑level fragmentation, while also substantially raising the annual safety exemption cap for fully autonomous vehicles per automaker from 2,500 to 90,000 vehicles, removing policy obstacles for large‑scale commercial operations by companies like Tesla and Waymo.

China's policies adhere to the principles of safety first and steady implementation, forming a complete framework of "central top‑level design + local pilot deployments + mandatory national standards as a safety net." In 2024, multiple ministries jointly issued a notice allowing L3 and L4 high‑level autonomous vehicles to conduct road tests in designated areas; pilot cities such as Shanghai, Guangzhou, and Chongqing simultaneously released detailed implementation rules, supported by five mandatory national standards and two recommended standards, specifying uniform requirements for system safety, data recording, risk response, and more, to reinforce the safety baseline. In December 2025, China's first batch of L3 autonomous driving production models received market approval, officially ushering in the era of large‑scale pilot deployment of high‑level autonomous driving.

By comparison, China's regulatory approach places greater emphasis on safety redundancy and a more rigorous deployment process: L3 and above vehicles must undergo multi‑stage testing, evaluation, and approval, and are required to be equipped with safety drivers and manual control devices. The U.S., on the other hand, may waive certain safety procedures and supports the deployment of fully driverless vehicles without safety drivers or steering wheels.

03 Full Industry Chain Analysis of Intelligent Driving

The intelligent driving industry chain is complete and clearly structured: the upstream consists of various core hardware components for perception, decision, and execution, forming the foundation for the deployment of intelligent driving functions; the midstream comprises intelligent driving system and algorithm suppliers, responsible for hardware integration and functional tuning; the downstream covers diverse commercial scenarios including vehicle sales, mobility services, and platform partnerships. The industry participants are primarily divided into three core groups: OEMs, intelligent driving system suppliers, and hardware suppliers.

3.1 Perception Layer: Multi‑Sensor Iteration and Upgrade, Accelerated Domestic Substitution

The perception layer serves as the "eyes" of intelligent driving. Its core sensors include cameras, ultrasonic radars, millimeter‑wave radars, and LiDAR, each with complementary performance, cost, and application scenarios, together forming a complete environmental perception system.

Ultrasonic radar has the lowest cost—averaging only RMB 28 per unit in 2025—but its detection range is only a few meters and its function is limited; it is currently used solely for reversing assistance.

Cameras offer a balanced cost‑performance ratio and can recognize semantic information such as colors and traffic signs. They are standard hardware for all intelligent driving systems, with an average unit price of about RMB 266 in 2025. Their only drawback is relatively weak perception stability in harsh environments.

Millimeter‑wave radar excels at anti‑interference, unaffected by rain, snow, or fog; it provides high speed measurement accuracy and a detection range of 200‑400 meters. In 2025, the average unit price dropped to RMB 111. Its accuracy is slightly inferior to LiDAR, but high‑end 4D millimeter‑wave radar performs comparably to a 64‑line LiDAR, making it a mainstream blind‑spot supplement sensor.

LiDAR leads the industry in perception accuracy, point cloud density, and detection range. For example, the Hesai ATX LiDAR achieves a point cloud density of 3.84 million points per second, far surpassing the 30,000‑100,000 points per second of 4D millimeter‑wave radar. It can precisely identify small targets and irregular obstacles, meeting the safety redundancy requirements of high‑level autonomous driving. Its drawbacks are relatively high cost and weaker penetration in rain and snow; the average unit price in 2025 was about RMB 1,835.

Currently, the core debate in the perception layer revolves around LiDAR adoption: on one hand, LiDAR costs remain high, prompting mid‑ and low‑end vehicles to adopt "millimeter‑wave radar + camera" solutions without LiDAR for cost control; on the other hand, multi‑sensor fusion algorithms are complex, making it difficult to define priority levels among different hardware data, leading to information conflicts and system misjudgments that require ongoing technical optimization.

3.1.1 Cameras: Volume and Price Rise Together, Domestic Substitution Accelerates Across the Board

Cameras are the only sensors that can capture visual semantic information and cannot be replaced by radar. They are widely used for external driving, reversing, side‑view monitoring, and in‑cabin occupant monitoring.

As high‑level autonomous driving penetrates lower‑tier markets, higher pixel counts, miniaturization, and improved stability have become the key iterative directions for automotive cameras. Among them, forward‑view cameras have higher resolution and significantly greater value than surround‑view or rear‑view cameras.

Policy support continues to expand industry demand. The new national standard explicitly requires that, starting January 1, 2027, all L2 driver assistance systems must be equipped with driver gaze‑off detection and return‑to‑attention alerts; prolonged disengagement must trigger a forced system disablement. This directly boosts both the penetration rate and the number of cameras installed per vehicle for in‑cabin monitoring.

Currently, the automotive camera industry is experiencing simultaneous growth in volume and price: since 2022, annual sales volume growth has averaged 26%; in the first 11 months of 2025, sales reached 93 million units, and the full year is expected to exceed 100 million. The average camera price per vehicle rose from RMB 223 in 2022 to RMB 266 in 2025, largely driven by the increasing share of high‑pixel, high‑end cameras.

In terms of market structure, the trend of domestic substitution in forward‑view cameras is clear. The current industry CR5 is only 46%, with a relatively fragmented landscape—the era of foreign dominance has ended. The market share of leading foreign brand Bosch fell from 20.1% in 2024 to 13.2% in 2025, and other foreign players like Denso have also seen declines. Domestic manufacturers such as Sunny Intelligent, BYD, Baolong, and O‑Film are steadily capturing market share. The top five optical lens suppliers in the industry chain are all domestic companies; Sunny Optical holds over 30% of the global market share and has been the world's No. 1 in shipments for many consecutive years. Local companies now possess full‑stack camera module solution capabilities.

3.1.2 Millimeter‑Wave Radar: 4D Technology Iteration, Domestic Players Overtaking on the New Curve

Millimeter‑wave radars are divided into forward‑looking radars and corner radars: forward radars are mounted at the front of the vehicle for medium‑ to long‑range detection, supporting adaptive cruise control, emergency braking, and other core functions; corner radars are installed on both sides of the vehicle to cover lateral and rear short‑range areas, suitable for lane‑change assistance and blind‑spot monitoring.

The iterative core direction for the industry is 4D millimeter‑wave radar, which adds vertical height detection capability, solving the traditional radar's inability to identify low‑height obstacles. Its angular resolution can reach 0.5 degrees, approaching the performance of a 64‑line LiDAR, while maintaining a significant cost‑performance advantage. In 2025, the unit price of 4D millimeter‑wave radar dropped to RMB 200‑400. Combined with the new national standard taking effect in 2028, which mandates AEB function upgrades, 4D millimeter‑wave radar is expected to become standard on intelligent driving models. In 2025, domestic deliveries of 4D millimeter‑wave radar reached 15.17 million units, a staggering 862% year‑on‑year increase, and its share of total millimeter‑wave radar installations reached 37%.

Leveraging high stability and low cost, overall millimeter‑wave radar installations have continued to grow rapidly: deliveries increased from 13.68 million units in 2021 to 41.01 million units in 2025, a four‑year CAGR of 32%. Economies of scale and technological iteration have driven prices down, with average mainstream product prices falling from RMB 249 per unit in 2022 to RMB 111 in 2025—a drop of more than half.

The market landscape features foreign dominance in the traditional market, while domestic players lead in the new 4D products: in the overall millimeter‑wave radar market, Continental, Bosch, and Denso hold a combined CR3 of over 71%, with foreign companies still dominant; however, in the high‑end forward 4D millimeter‑wave segment, domestic manufacturers captured over 95% market share in 2025, successfully overtaking foreign incumbents through the new technology track. As intelligent driving functions become popular in mass‑market models priced around RMB 150,000, the high‑end perception suite of "LiDAR + 4D millimeter‑wave" is accelerating its penetration, and the future penetration rate is expected to rise further.

3.1.3 LiDAR: Rapid Technological Iteration, Sharp Cost Reduction, and Complete Domestic Substitution

Over the past seven years, LiDAR has undergone a leapfrog technological evolution, transitioning from traditional analog architectures to digital architectures. By integrating semiconductor chips such as CMOS, SPAD, and TDC, and eliminating complex mechanical rotating structures, the technology has significantly reduced size, power consumption, and cost, while also breaking through line‑count limitations—analog LiDAR took 15 years to upgrade from 64 to 128 lines, whereas digital LiDAR developed 1440‑line and 2160‑line high‑end products in just 15 months, achieving qualitative leaps in point cloud density and obstacle recognition accuracy.

Mass production and technological upgrades have driven LiDAR prices down precipitously, ushering in the "thousand‑yuan era." Taking Hesai Technology as an example, its LiDAR average selling price (ASP) dropped from RMB 14,000 per unit in 2022 to RMB 1,800 in 2025, an annual decline of about 50%. Lower costs have driven rapid penetration growth: the penetration rate of LiDAR in domestic passenger vehicles rose from 0.04% in 2021 to 13.68% in 2025, with an average annual growth rate of 321%. It has now fully penetrated mainstream family models in the RMB 100,000‑150,000 price range.

At the same time, the proliferation of end‑to‑end autonomous driving technology has further expanded the demand for LiDAR. Traditional modular architectures were prone to information conflicts among multiple sensors, but end‑to‑end models can directly fuse data from cameras, LiDAR, and other sources without manual rule‑based trade‑offs, completely solving the sensor fusion challenge and fully realizing the high‑precision perception value of LiDAR.

The domestic LiDAR market has undergone a thorough restructuring, achieving a high degree of domestic substitution with a domestic market share exceeding 90%. The industry landscape has shifted from the 2022 triopoly of Innovusion, Hesai, and Livox (with a combined 97% share) to the new 2025 leading trio of Huawei, Hesai, and RoboSense, which together hold a 92% market share. Among them, Huawei's share leaped from 1.1% in 2022 to 38.7%, ranking first in the industry; the former leader Innovusion saw its share drop sharply to 7.8%, with a narrow customer base and slow iteration speed being its core weaknesses. Currently, Hesai Technology has secured binding contracts with major OEMs such as BYD, Li Auto, and Geely, with large‑scale orders consistently being fulfilled.

Source: Tiqi Auto