We believe that the current car companies have entered the stage of intelligent driving competition, and with the maturity of L2+ functions represented by urban NOA and the increase of L3 penetration, intelligent driving will have a profound impact on the market structure.
summary
The evolution of intelligent driving technology has flowed through the exploration of deep waters, and the development trend has gradually become clear. We comprehensively sort out the current situation and planning of the intelligent driving technology layout of independent car companies: 1) The E/E architecture upgrade is the foundation for the realization of intelligent driving, and some car companies take the lead in evolving to central computing + regional control. 2) The hardware side accelerates and reduces costs, and the configuration differentiation matches the needs of different market segments, with the market below 200,000 yuan mainly focusing on low-computing power chips + pure vision solutions, and more than 200,000 yuan matching high-computing power chips + lidar solutions. 3) The algorithms are basically switched to the BEV+Transformer+Occupancy scheme, and the end-to-end large model is the upgrade direction.
Intelligent driving capabilities are beginning to show a watershed, and it is expected to drive market concentration to increase in the medium and long term.
► The mass production pace of new forces and Huawei is leading, and engineering capabilities are the key to improving generalization. We believe that the urban NOA function is a direct embodiment of the current intelligent driving capabilities of car companies, Huawei and Xpeng are relatively ahead of the mass production rhythm, Li and NIO plan to accelerate the full push of urban NOA functions nationwide in 1H24, and other car companies will gradually achieve from 0 to 1. Looking ahead, before the large-scale mass production of end-to-end large models, we believe that engineering capabilities will become the key to the competition, and the generalization improvement speed of car companies with first-mover advantage has entered the stage of sharp increase in the S-shaped curve, but late-mover car companies with data advantages are expected to catch up after breaking through the key node.
► Intelligence is expected to bring about the second wave of pattern changes, and the concentration of high-end markets may increase. We predict that the urban NOA function is expected to enter a mature period in 2025, from "usable" to "easy to use", driving intelligent driving to become an important decision-making factor for car purchases, and if supported by laws and regulations, we predict that L3 intelligent driving will begin to penetrate and gradually improve by the end of 2020s. We judge that intelligence will guide the pattern change, but the degree of impact is differentiated in different price band markets, and the impact on the pattern of the mid-to-high-end market will be stronger. The reason is that low-end models require hardware and software to continue to reduce costs, most of them are mainly based on supplier solutions, and a few car companies with outstanding capabilities may conduct self-development, while mid-to-high-end brands need to highlight intelligent longboards through self-development, and there will be differentiation in capabilities. Referring to the evolution of the smartphone pattern, the rapid technological change brings about the knockout game, and the technology convergence period is more about the scale effect and cost reduction ability, we believe that intelligence is expected to drive the market pattern to concentrate again.
risk
Technology iteration and functional experience improvement are slower than expected, and regulations and policies are advancing slower than expected.
body
Introduction
The electrification and intelligent transformation of the automotive industry have driven vehicle structure and technological innovation. Electrification technology revolves around improving electric performance and solving the problem of energy replenishment, and the electrification of passenger cars in the Chinese market has gone through a bottleneck period, and the penetration rate of new energy has exceeded 40% since 2H23. With the introduction of electrification technology from the introduction to maturity, the market competition tends to be fierce, the elimination of car companies lacking core competitiveness, and the increase in the share of car companies with product and scale advantages, the market pattern has gradually moved from decentralization to concentration during the new energy introduction period.
Looking ahead, intelligence (including cockpit and driving intelligence) has gradually become the key to product differentiation, intelligent cockpit software and hardware configuration and experience upgrades, intelligent driving is ushering in a technological and policy inflection point, opening the second half of the competition. We believe that it is still in the early stage of the introduction of intelligent driving technology, and with the maturity of L2+ functions and the increase of L3 penetration, intelligence is expected to drive the market pattern to concentrate again.
In this report, we summarize the core technology stack of intelligent driving from the perspective of car companies, sort out the progress and layout of mainstream car companies, and analyze the current situation and influencing factors of their intelligent driving level. On this basis, we try to discuss several core issues that the market is concerned about: how to evaluate the current level of intelligent driving technology of car companies, as well as the convergence and differentiation of intelligent driving capabilities of car companies in the future? how to look forward to the trend of perceiving hardware technology and configuration? whether software charging is feasible in China, what is the premise of intelligent driving for model sales to begin to have a driving effect, and what impact will intelligent driving have on the competition pattern of the medium and long-term intelligent electric vehicle market?
Figure 1: Intelligent driving analysis framework for OEMs
Source: CICC Research
The evolution of intelligent driving technology has flowed through the deep water area, and the development trend is clear
In this section, we start from the core technical capabilities involved in autonomous driving, and comprehensively sort out the layout status and planning of various car companies, mainly including electronic and electrical (E/E) architecture, chips, perception hardware, algorithms, computing power and other software and hardware levels, so as to clarify the relationship between core technologies and intelligent driving capabilities. We believe that if the intelligent system of the whole vehicle is analogous to humans, the E/E architecture is the nervous system, the chip is the control brain, the perception hardware is the five senses that interact with the outside world, the algorithm is the blood flowing in it, and the computing power is the source of energy supply.
E/E architecture: laying the intelligent foundation of the whole vehicle, domain control is the current normal, and some car companies are taking the lead in evolving to central computing + regional control
The E/E architecture is the main framework for the development of vehicle electronics and electricity, laying the foundation for vehicle intelligence. Cross-domain integration is at the stage that most car companies are in, and they are gradually moving towards a central computing platform. At present, the E/E architecture of most car companies is in the transition stage from "domain centralized" to "central computing", and the vehicle is formally controlled by the central computing platform (OneBox structure), but the functions of each domain are still completed through different domain controllers. Most car companies have integrated the body domain, power domain and chassis domain, and some have realized the integration of cabin and parking. We believe that the integration of cabin and driver in the short term is an important development direction of cross-domain integration, and in the medium and long term, the central computing platform is the evolution direction of the E/E architecture, that is, the use of "one brain" for control. Of course, we can't simply assume that the more centralized the E/E architecture, the better, and there is also a game between efficiency and security, and car companies need to consider the whole picture to achieve comprehensive optimization.
From the perspective of the landing progress of car companies, mainstream car companies have generally realized the transition from distributed architecture to domain control (including central computing + functional domain control) from 2021 to 2022, and gradually evolved to central computing + regional control, and some car companies will start to get on the car in 2023. Overall, we believe that 2024-2025 is the time point for car companies to iterate from domain control to central computing platforms, but they still need to pay attention to the pace of mass production of actual models.
Figure 2: E/E architecture evolution roadmap
Source: Bosch official website, CICC Research
Chips: Suppliers are oligopoly, and the demand for computing power is rational and differentiated
With the upgrade of intelligent driving functions and the evolution of E/E architecture, the computing power demand of intelligent driving chips has increased significantly. We see that with excellent hardware performance, an open ecosystem and the ability to mass produce at the earliest, the NVIDIA Orin chip has almost become the "standard" for high-end intelligent driving. In addition to NVIDIA, domestic chip manufacturers are also actively making efforts. We believe that domestic chips can provide a more cost-effective option, and are relatively independent and controllable, with a low risk of supply interruption.
In terms of the impact of chips on the intelligent driving capabilities of car companies, we believe that in addition to the consideration of chip performance, the stability of chip supply and the continuity of chip platforms are also worth paying attention to. We have seen that many car companies have switched smart driving chip suppliers, collectively switching from Mobileye to NVIDIA in the early days, and some subsequent models have switched to domestic chips. We believe that the chip selection itself is related to the choice of perception hardware and software algorithm architecture, and the continuity of development based on a unified chip platform may be relatively good, while avoiding the incremental cost and customer experience of later switching.
Figure 3: Car companies mainly use intelligent driving chips to review
Source: Company announcement, CICC Research
Some car companies focus on self-development of chips to reduce costs, increase efficiency, and reduce dependence on external supply chains. In terms of the progress of self-research by car companies, the Leap Lingxin 01 has been installed in C01/C11 with a computing power of 4.2TOPS. NIO Shenji NX9031 will be equipped with ET9 (the company expects mass production in 1Q25), with a computing power of more than 1000TOPS. Geely plans to launch a self-developed automotive-grade intelligent driving chip with a computing power of 256 TOPS from 2024 to 2025 to realize the commercialization of L4 autonomous driving. We believe that the difficulty of on-board chips lies in the mass production, which takes about 1-3 years from tape-out to SOP, during which it still needs to go through multiple steps such as vehicle specification certification and model introduction verification, and mass production ramp-up and yield control also need to be accumulated, so we expect that the mature mass production and mass production of self-developed chips by leading car companies will generally be later than 2025.
Perception hardware: Multi-sensor fusion and pure vision solutions are paralleled to match the needs of different market segments
Mainstream autonomous driving sensor hardware includes cameras, millimeter-wave radar, ultrasonic radar, lidar, etc. In addition to Tesla, the perception hardware solutions of existing models equipped with L2+ intelligent driving functions are mainly multi-sensor fusion solutions of camera + lidar + millimeter wave radar + ultrasonic radar, and lidar is usually equipped with 1-3 lidars, which account for the highest cost. Benefiting from algorithm upgrades, suppliers represented by DJI and Momo Zhixing actively promote pure vision (camera + millimeter-wave radar and other sensors) solutions to achieve cost reduction of perception hardware.
► LiDAR: We believe that LiDAR is relatively expensive, but it has high perception accuracy and strong safety guarantee capabilities, and is widely equipped in mid-to-high-end models, which has become a must-have for L3+ perception redundancy.
► The camera configuration is differentiated with the algorithm route, and the pure vision algorithm has high requirements for the number of cameras, pixels and other indicators. At present, the mainstream camera scheme is 8-12, except for the front view camera is generally equipped with 1-2 (monocular, binocular and multi-eye), and the surround view and side view are mainly 4+4 schemes.
► Millimeter-wave radar: It can accurately judge the distance, speed and direction, and respond to the needs of horizontal and longitudinal perception with different wavelengths. Entering the L2 and above stage, car companies are generally equipped with 3-6 millimeter-wave radars (5 are the mainstream scheme, 1 forward + 2 lateral + 2 backward), and if 4D millimeter-wave radar with increased altitude perception function is used, the traditional millimeter-wave radar can be reduced.
► Ultrasonic radar: The perception distance is close, mostly used in parking scenarios, and it has limitations at high speeds. The 4/8/12 ultrasonic radar solution is the mainstream position, and models with valet parking, automatic parking and other functions are generally equipped with 12 as standard.
Algorithm: Actively follow up the "light map" scheme and evolve towards an "end-to-end" large model
Tesla is leading the revolution in autonomous driving algorithms. Since 2020, Tesla's autonomous driving algorithm has undergone three major innovations:
• 2020: BEV (Bird's Eye View) + Transformer was introduced, and feature-level fusion was used to replace post-fusion.
• 2021-2022: Introduce time series networks to form 4D spatial information, and upgrade BEVs to occupancy networks.
• August 2023: Tesla proposes an end-to-end large model on the car, which is based on the deep convolutional neural network model for deep learning, in addition to the perception layer (input) is realized through the model, the planning and decision-making layer (output) is also realized through the model, that is, the perception and decision-making modules are integrated in one model.
We see that domestic car companies have quickly followed up, and most of the leading car companies have switched to the BEV+Transformer model at the perception layer, but the neural network at the planning and decision-making end is still relatively lagging behind. We believe that the data-based, rule-driven end-to-end large model is expected to become the next competitive highland for the intelligent driving algorithm of car companies, but due to the limitations of algorithms, computing power and data, it will still take a long time for domestic car companies to mature and implement large models.
Figure 4: Algorithm solutions of domestic automakers
Source: Company announcement, CICC Research
Data and computing power: The demand for data training has increased significantly, and car companies are actively building their own intelligent computing centers
Large model training drives a significant increase in data demand, and the flywheel effect accelerates the iteration of large models. Data closed-loop refers to the collection of complex scene data (5% corner case) in driving by the car, and backhaul, annotation, and cleaning, and transmission to the algorithm model for training, and the new model is deployed to the vehicle to run the cycle verification.
The premise for the realization of high-end intelligent driving functions is sufficient platform computing power, and it has become a trend for car companies to build their own intelligent computing centers. We believe that the input, analysis and output of massive data are a huge test for computing power requirements, and at the same time, the construction and operation of the digital twin environment in the autonomous driving simulation stage also require a large amount of high-quality computing power. We believe that compared with public cloud services, the self-built computing centers of car companies will be dedicated to autonomous driving applications, which is expected to improve the efficiency of model training.
The ability of intelligent driving has begun to show a watershed, and the market concentration will be further improved in the medium and long term
How to evaluate the current intelligent driving capabilities of car companies?
We believe that the application scenarios of urban NOA and high-end parking are relatively complex, but they also meet the pain points of consumers' actual car use, and are the obvious demonstration of the intelligent driving capabilities of car companies in the current L2+ stage (people-vehicle co-driving). As far as the urban NOA function is concerned, we believe that the landing speed and the quality of the experience are the core evaluation criteria, and in the current stage from 0 to 1, we should first look at the landing speed, and then look at the quality of the experience. As far as high-end parking functions are concerned, the algorithms of various car companies in conventional scenarios are not very different, and the parking efficiency is similar, and complex pain point scenarios are the next competitive highland. In the following, we sort out the progress of major car companies in the field of urban NOA and smart parking, and clarify the differences between the publicity caliber of all parties, and the differences between the comparison of publicity and the actual landing rhythm, so as to form a global overview and analyze and compare.
We believe that the implementation process of urban NOA can be measured from two dimensions, one is the region and the other is the customer group, and the expansion of the region and the customer group represents the generalization ability. For car companies, the first stage is the implementation of the no-map solution, which means that it can break through the regional restrictions, which is the premise of opening the second stage; the second stage is large-scale push, which we believe can be judged from two quantitative dimensions: the number of cities * urban road coverage * unit mileage) and the number of open users (total users * proportion of high-end intelligent driving users * user opening ratio).
Xpeng's opening speed has been accelerated to unlimited XNGP launch, and Huawei has accelerated the coverage of OTAs in all road sections across the country. Starting in February 2024, Huawei will gradually roll out the HUAWEI ADS 2.0 advanced intelligent driving feature package nationwide, covering all cities in China and supporting all roads (trunk roads, secondary roads, and branch roads) in cities, accounting for up to 99% of the available road sections. As of January 2024, Xpeng NGP has been opened in 243 cities, covering 569,000 kilometers of road sections, and on February 29, 2024, Xpeng announced that it will launch unlimited XNGP intelligent driving assistance functions for users with intelligent driving experience, regardless of city and route, and only need to navigate based on SD maps.
As of the end of last year, Ideal City NOA had covered more than 110 urban roads, and as of 0:00 on February 1, NIO's NOP+ had accumulated a verified mileage of 651,600 kilometers (excluding urban expressways) on urban roads, covering 606 cities, and more than 90% of urban trunk roads. Jiyue has opened 4 cities, and Zhiji has pushed the city NOA function in Shanghai.
Looking ahead, Xpeng Ideal NIO accelerated the implementation of NOA in cities across the country in 1H24, and other car companies gradually accelerated from 0 to 1. NIO plans to launch the first round of 1,000-person global navigation assistance for NOP+ users in February 2024, and open the global NOP+ urban navigation assistance function to all Banyan intelligent system users in 2Q24. SAIC plans to launch a nationwide public beta of the pilot version of the urban commuting mode in 1Q24, and launch a mapless city NOA in June 2024 to accelerate the opening of the city. In terms of other independent brands, Leapmotor, Denza, Great Wall Wei, Aion, Haobo and Baojun all plan to start NOA push and testing of some models and some cities in the first half of 2024.
Figure 5: Urban NOA landing and mass production nodes
Source: Company announcement, CICC Research
Xpeng and Huawei have taken over a small number of times, and the performance of other car companies still needs to be tracked. We believe that the actual experience of urban NOA needs to be compared based on the same standard, and it may be biased to compare the unpictured and pictured solutions, so the current models that meet the comparison standards are only Huawei and Xpeng, and other car companies still need to continue to track them. Considering the driving experience and media evaluations, we have observed that the current Huawei series and Xpeng are relatively smooth and sensitive in common urban driving scenarios such as detour avoidance, intersection turning, and lane change and overtaking. In specific scenarios, because the pure graphless scheme highly relies on real-time perception, the corresponding decision-making occasionally "jitter" phenomenon. In terms of the number of takeovers, in September 2023, according to Internet evaluations such as "Know Chedi Original"[1], during the evening rush hour in the main urban areas of Shanghai and Beijing, the number of takeovers of the 20-kilometer Huawei series and Xpeng smart driving models is in single digits, and the overall performance is relatively smooth.
The common perception scheme for intelligent parking is the fusion of surround view camera and ultrasonic sensor, and there is no significant difference in the intelligent parking algorithm of different car companies. When judging the current level of intelligent parking of various car companies, we believe that the mass production rhythm of intelligent parking and the actual parking efficiency are important evaluation criteria, and there is no generational difference as a whole: automatic parking and remote control parking functions are more mature, memory parking is developing to all scenarios, and valet parking is not mass-produced.
In the normal scenario, the level of parking efficiency is similar. Referring to the evaluation of the China Intelligent Vehicle Index of CAERI IVISTA, the intelligent parking sub-index indicators mainly include the number of garage rubbing, parking attitude, and the success rate of remote control of parking in and out. Under this evaluation, NIO ES7, ZEEKR 001, Ora Lightning Cat, AVATR 11, Nezha S, and Ideal L9 all received excellent ratings. Referring to the Internet evaluation, in the conventional scenario (with cars on both sides, moderate vehicle distance and road width), the parking time of the side and vertical parking spaces of AVATR 11, Xpeng G9, Li L8, NIO ES7 and other models is 40s+, which is close to the time taken for manual parking. In extreme scenarios (narrow distance between the two sides and narrow road width), AVATR 11 has better comprehensive performance in the success rate of parking space recognition, parking success rate and smoothness.
Looking ahead, we believe that the applicable distance (usually within 2km) and applicable scenarios of memory parking are relatively narrow, which is more similar to transitional products, and it is difficult for valet parking to be mass-produced in the short term. We judge that improving the parking capacity of difficult scenarios may be more targeted to solve consumer pain points, such as parking on broken roads, parking in ultra-narrow parking spaces, parking in self-seeking parking spaces, parking in three-dimensional parking spaces, etc.
Figure 6: Comparison of parking functions by automaker
Note: There is no relevant information in the blank section
Source: Company announcement, CICC Research
How do you view the trend of intelligent driving hardware technology and configuration?
The decrease in purchase price and the upgrade of algorithm route jointly drive the cost of perception hardware. According to our statistics, the price of a single lidar will drop to about 3,000 yuan at the lowest in 2023, and the decline is considerable and is expected to continue. The price of traditional corner radar has dropped to about 200 yuan, the price of forward millimeter-wave radar has dropped to about 400 yuan, and the price of 4D milliwave radar has dropped to about 1,000 yuan. At the chip level, the unit price of Nvidia Orin-X is about $500, and the Horizon Journey 5 is about 20-30% of it. With the upgrade of intelligent driving algorithms, the hardware solutions of intelligent driving of car companies tend to be reduced, referring to Huawei's ADS hardware solution, compared with ADS1.0, ADS2.0 reduces 2 lidars, 3 millimeter-wave radars and 2 cameras, and the chip computing power is reduced from 400TOPS to 200TOPS, and we expect the total hardware cost to be reduced by about 20,000 yuan.
Under the trend of cost reduction, binocular vision solutions and 4D millimeter-wave radar have cost-effective advantages. A number of suppliers have launched a series of cost-effective intelligent driving solutions, and the price has dropped to the level of 1,000 yuan. The main sensors of DJI's next-generation intelligent driving solution platform are one set of 8 million pixel forward-looking and inertial navigation binocular eyes, four 3 million pixel surround view fisheye cameras, and one 300/8 million camera in rearview mirrors. This solution can realize urban memory driving (32TOPS) / urban pilot driving (80TOPS), and has been mass-produced on Wuling Baojun Yunduo and Chery iCAR 03. The prices of the three configurations of the second-generation HPilot platform are 3,000 yuan, 5,000 yuan, and 8,000 yuan respectively, of which 8,000 yuan can realize the NOH function without a map in urban areas.
Figure 7: Comparison of hardware solutions and costs
Source: Company announcement, Vehicle official account, CICC Research Department
In terms of pricing, the cost is calculated, and the reasonable range of the intelligent driving hardware solution may be 3-5%. In addition to Baojun Yunduo Lingxi, the current price of models equipped with high-speed NOA and urban NOA functions is generally more than 200,000 yuan, and we estimate that the cost of intelligent driving hardware is 20,000 yuan. DJI believes that the reasonable total cost (including software and hardware procurement costs) of L2+ intelligent driving systems ranges from 5,000 to 15,000 yuan, and the upper limit of the cost proportion is 5%, which represents the premium that consumers are willing to pay to alleviate driving fatigue, while the lower limit is 3%, which represents the basic cost of sensors, chips and other basic costs that need to be configured to ensure the availability of intelligent driving systems. The lower limit of 3% can already provide an integrated driving and parking system, a variety of high-resolution sensors, and ensure that the software side has sufficient profits, while the upper limit of 5% can be equipped with higher-cost hardware such as lidar.
The mid-to-low-end models tend to be a pure vision low-cost solution, and the mid-to-high-end models will continue to have a multi-sensor fusion solution. We believe that the choice of intelligent driving hardware for car companies depends not only on the tolerable cost mentioned above, but also on the competitive environment, and the configuration of competing products is largely related to market segment positioning and consumer preferences. Consumers have a high perception of medium and high computing power chips, lidar and other configurations, and mid-to-high-end models need to highlight the intelligent longboard, and the intelligent driving experience should reach the standard of "easy to use", so we believe that there is a certain room for the cost of intelligent driving hardware in mid-to-high-end models, and we judge that the medium and high computing power chips + 1 or more lidar + other perception hardware will be the main solution, and car companies may carry out high and low configuration according to the specific product version. The 10-200,000 yuan market requirements for intelligent driving functions are "useful and usable", and we believe that the reasonable hardware cost should be controlled below 10,000 yuan, with pure visual solutions as the mainstay.
Medium- and low-end models may be more economical, while mid-to-high-end brands tend to source hardware and self-developed software, driving the differentiation of intelligent driving capabilities. We believe that from the perspective of self-development and outsourcing, models below 200,000 yuan have relatively low requirements for technical integration, and there are mature suppliers in the industry with driving and parking solutions, outsourcing may be more economical and efficient, but some car companies with outstanding capabilities may choose to develop their own products to create differentiated selling points. Mid-to-high-end brands have higher requirements for the completion of intelligent driving functions, and need to be supported by more advanced architectures, higher-performance hardware, and more advanced algorithms. Chips, lidar and other hardware have a high technical threshold, at present, except for Leap, other car companies' hardware solutions are basically based on outsourcing, and some car companies develop their own chips; software algorithms are mainly self-developed, supplemented by outsourcing, considering the different layout rhythms of car companies and the differences in capabilities, we believe that the intelligent driving capabilities of mid-to-high-end market brands will be differentiated.
Will the gap in intelligent driving ability diverge or converge?
We believe that the level of intelligent driving of car companies mainly depends on algorithm capabilities and engineering capabilities, and engineering capabilities have gradually become the key to competition. As mentioned above, we believe that under the leadership of Tesla, the technical path of domestic car companies in the field of intelligent driving algorithms tends to converge, BEV+ transformer + occupancy has become the mainstream perception algorithm framework, and the next step is to evolve towards an end-to-end large model integrating perception, prediction, regulation and decision-making. We believe that the advantage of the end-to-end large model is that it can connect all modules, which can achieve global optimization, anthropomorphic effect, higher theoretical technical ceiling, and more concise algorithm code. The challenges of mass production are not only the difficulty of algorithm design, the high demand for computing power, and the higher requirements for the quantity and quality of training data for end-to-end large models, but also the relatively poor interpretability, the risk that problems are difficult to trace, and the division of rights and responsibilities is not clear.
Therefore, we believe that it will still take a long time to truly enter the large-scale mass production of end-to-end large models, and before that, there is no significant generational difference in the ability of various car companies to build algorithms. With the improvement of generalization requirements, intelligent driving solutions need to meet the requirements of controllable cost, fast iteration and update, and strong versatility, and we believe that engineering capabilities have gradually become the key to the competition, among which hardware, software, and test engineering capabilities are generally reflected in the first-mover advantage, while in data engineering, there is the possibility of latecomers gradually catching up.
Figure 8: Automakers' intelligent driving engineering capabilities
Source: Yanzhi Automobile, CICC Research
We believe that hardware and software engineering capabilities are an important foundation, which requires the coordination of organizational structure and talent team, and requires a certain construction period, because it involves supply chain management, organizational structure, personnel team, capital investment and other aspects. Data engineering is the key to improving generalization, and the efficiency is expected to improve rapidly after breaking through key nodes.
In terms of data volume, the stock of sports models is an important foundation, and on this basis, the core requirement for data collection and model training is the unification of sensor configuration, that is, different models adopt basically the same camera configuration, including quantity, parameters, location, etc. Starting from 2022 to 2023, car companies will unify the camera configurations of major models (including different versions such as Max/Pro) to ensure the volume of data. According to our estimates, by the end of 2024, benefiting from the growth in sales, Ideal is leading in the volume of effective return data models, and is expected to accelerate catch-up with its data advantage, while brands such as Huawei and NIO are relatively leading in volume.
The actual implementation progress of NOA in cities without maps, city coverage, and OTA frequency of intelligent driving software can be used as quantitative reference standards. Referring to the OTA update frequency released by the Passenger Car Association, we see that Tesla, Xpeng, Ideal, NIO, etc. as a whole maintain a leading position in the update frequency, and basically push new OTAs every month. In terms of the OTA situation of the intelligent driving software version, Xpeng Xmart OS has been iterated from version 4.2.0 to version 4.5.0, and each version basically has XNGP function updates and upgrades.
Can the charging rate of intelligent driving software be increased and a new business model be opened?
The charging mode of intelligent driving software is diversified and the charging standard is differentiated. We have sorted out the current mainstream brands' intelligent driving charging models, and the model price generally only includes basic intelligent driving functions, such as AEB, ESA and other active safety functions and APA, RPA and other parking functions, and most of the high-speed/urban NOA functions need to be purchased additionally, which can be divided into the following types of modes:
► Hardware standard, software optional: LiDAR and other perception hardware are all standard, the hardware has been fully priced, and the software package is paid. NIO, AVATAR, and Wenjie M9 adopt this form.
► Hardware options, software bundles: New models are usually divided into versions with and without LiDAR, and high-end versions come with software packages (equivalent to bundled sales). This form is adopted by car companies such as Xpeng, Li and Denza, among which the price difference (hardware price) between the Max and Pro models is 2-30,000 yuan.
► Software and hardware are standard: car companies including Leap, Shangtong Wuling, etc., car companies self-developed or pure vision algorithms are used to reduce costs, and software and hardware are standard.
For software payment, car companies can be divided into buyout and subscription:
► Buyout: The user pays the software package fee in one lump sum, and the subsequent OTA function upgrade is generally free. Huawei, Zhiji, Feifan, Zeekr, etc. all adopt this form, and the buyout price before the implementation of rights and interests exceeds 30,000 yuan (about the price of a 5-year subscription), and the price drops to several thousand yuan after the rights and interests are discounted.
► Subscription: Users can choose to subscribe on a monthly or annual basis. NIO's high-end NOP+ subscription price is 380 yuan/month, and Huawei's smart partner model also offers this option, with an ADS 2.0 subscription price of 720 yuan/month.
Intelligent driving payment transitions from the cultivation period to the development period, lowering the threshold to acquire consumers. We believe that it is more difficult for low-end market models to charge for intelligent driving software, so our following discussion is mainly focused on the mid-to-high-end market. Referring to smartphones, we believe that the necessary conditions for improving consumer software payment are: 1) the frequency of use is high enough to solve the pain points of daily car use, such as urban driving, parking in complex scenarios, etc.; 2) the product power is strong enough, and there are differences in the functions and experiences of paid and non-paid; 3) the technology and experience of intelligent driving products of different car companies are differentiated, otherwise it will still be difficult for car companies to charge under the prisoner's dilemma of competing products continuing to acquire customers at low prices or even no fees.
We believe that, as we have analyzed earlier, the current packaging system of hardware payment + software free is more likely to give consumers a "sense of gain", because there is a clear market pricing for hardware, and software is nominally free, which is equivalent to giving away rights. However, if the above necessary conditions are met, we believe that the subscription system is an easier way to achieve the purpose of charging, and its advantage is that the subscription system lowers the threshold for payment, and consumers can choose the subscription cycle according to their own needs, which is easy to develop the habit of payment.
The L3 autonomous driving function will enable software backward payment, or it will be exclusive to mid-to-high-end vehicles. We predict that L3 may support a stronger reason for automakers to charge consumers for backward function subscriptions due to the high cost increase of L3 for a single vehicle, the need for automakers to purchase long-term insurance for their autonomous driving systems, and the functional logic of hiring dedicated drivers. In the long term, we believe that NOA in L2 cities may become mature and unable to provide competitive differentiation for models, while continued cost reduction is expected to equip NOAs in L2 cities in all price ranges.
When intelligent driving really moves towards L3+, theoretically, the intelligent driving charging model is expected to move closer to "substitute driving". When the vehicle is in the L3 autonomous driving level, the human driver still needs to monitor the vehicle's driving, and the L4-L5 level is basically dominated by the intelligent driving system, and the human driver is closer to the passenger role. We predict that due to the high cost increase of L3 for a single vehicle, and the need for car companies to purchase insurance for their autonomous driving systems for a long time, and at the same time, there is a functional logic for hiring exclusive drivers, and at the same time, after the level of intelligent driving reaches L3+, we believe that the marginal cost of new authorized use of intelligent driving software will be greatly reduced, further supporting car companies to charge consumers for backward function subscriptions.
What is the impact of intelligent driving on the market competition pattern?
We believe that the driving factors of the evolution of the market pattern of smart phones and smart cars are similar; in summary, the technology introduction period has a low threshold, poor product competitiveness, and many entrants drive the pattern differentiation; rapid technological change brings about a knockout game, while the technology convergence period examines the scale effect and cost reduction ability, and gradually drives the market pattern to concentrate.
Figure 9: Smartphone and smart electric vehicle market comparison
Source: CICC Research
From 2017 to 2020, the new energy technology revolution was opened, the response speed of the dominant fuel vehicle companies was relatively backward, BYD accelerated the listing of new energy models, and BAIC, Zhidou, Zotye and other car companies actively entered the game, quickly occupied market share, and the market concentration declined. From 2020 to 2023, the sales volume of the new energy industry will enter a period of rapid growth, electrification technology will gradually mature, product strength will be improved, car companies that are highly dependent on subsidies and lack core competitiveness will gradually be cleared, and scale effect and high-quality products will gradually become the core competitive factors, and the share of BYD, Tesla, and new forces will increase, and at the same time, the share of the TOP5 will increase again. In terms of price brought, CR3 in major market segments is more than 50%, and the share of the second echelon in some markets is small and the gap is not large, so there are opportunities for pattern changes.
We predict that in 2025, the function of urban NOA is expected to enter a mature period, and urban NOA will gradually change from "usable" to "easy to use". We believe that before L3 begins to penetrate, urban NOA is the core function of improving consumers' intelligent driving experience and completing intelligent driving education in the era of human-vehicle co-driving. By 2025, other leading car companies are expected to gradually follow up with commercialization, and first-tier players are expected to enter the mature stage of urban NOA.
We predict that L3 penetration will open by the end of the 2020s, with regulations as the core variable. On November 17, 2023, the Ministry of Industry and Information Technology (MIIT) and three other government agencies jointly issued the Notice of the Ministry of Transport on Carrying out the Pilot Work on the Access and Road Access of Intelligent Connected Vehicles (ICVs), which is related to the pilot scheme for the access and on-road driving of intelligent networked vehicles (ICVs). The "Notice" marks that the pilot of autonomous driving roads is expected to be gradually carried out nationwide, and clarifies the division of rights and responsibilities: if an accident occurs in the state of autonomous driving, the responsibility will be borne by the insurance company first, beyond the insurance liability, and if the accident responsibility is on the side of the intelligent networked vehicle, the pilot user (operating platform) will bear it. If the manufacturer, system development unit or infrastructure provider is at fault, the user can recover compensation. We believe that this division of responsibilities is in line with the core demands of L3/L4, and it is expected that the driver will be separated from the driving responsibility, which is expected to allow the driver to truly "take off his eyes and hands". We expect urban L3 penetration to begin by the end of the 2020s, with the expectation that it will gradually reach 40% over a period of several years.
Figure 10: Autonomous driving penetration forecast
Source: CICC Research
Intelligent driving ability has become the winner or loser of the next knockout game, and the concentration of mid-to-high-end has been further improved, and it may be a winner-take-all market in the medium and long term. For car companies, we believe that the contribution of intelligent driving to sales needs to be a gradual process, and the maturity and popularization of urban NOA and the increase in L3 penetration will be two important points. From 2024 to 2025, as L2+ functions gradually mature, consumers use the frequency increases, and the experience is optimized, we believe that the weight of the impact of intelligent driving on consumers' car purchases will increase, and the availability and level of intelligent driving will more directly affect the sales of models.
Referring to smart phones, the process of technological change from differentiation to convergence will form a knockout game driven by intelligence. For cars differentiated according to price gradients, we believe that consumers in low-end models are not willing to pay, and intelligent driving solutions are mainly adopted, but car companies with outstanding capabilities are expected to be equipped with relevant functions to create differentiated advantages, while mid-to-high-end market brands need to highlight intelligent longboards, and software algorithms are mainly self-developed, and the impact of the importance of intelligent driving on the mid-to-high-end market pattern may be more significant, and the mid-to-high-end brands with weak intelligent driving capabilities may be eliminated, driving the market pattern to centralize.
In the long run, when L3 begins to officially penetrate and gradually reaches a certain level, due to the significant reduction in consumer participation in actual driving, the reduction of attention to power, handling, performance, etc., and more attention to intelligent cockpit and intelligent ecology, the car has truly become a new Internet interactive entrance, the attributes of car use will be greatly changed, and a new cockpit interaction mode will come into being. Software has the characteristics of low marginal cost, which may form a more concentrated market structure.
Risk Warning
Technology iteration and functional experience improvement are slower than expected: there is a risk of supply interruption of some core hardware required for intelligent driving, there may be bottlenecks in the construction and training of end-to-end large model algorithms, and the fierce competition in the passenger car market may distract car companies from investing in intelligent driving. All of the above are risk factors for the slower-than-expected improvement of intelligent driving technology iteration and functional experience.
The promotion of regulations and policies is lower than expected: The commercialization of intelligent driving functions is highly related to policies and regulations, and if the necessary regulations such as the testing and verification system, access standards, division of rights and responsibilities, and regulatory standards of intelligent networked vehicles are not as advanced as expected, the pace of promotion of intelligent driving functions will be affected.
[1]https://www.dongchedi.com/article/7286696011471438390
Article source:
This article is excerpted from: "Ten-Year Outlook IV: Intelligent Driving Evolution for Automotive Companies" released on March 21, 2024
Deng Xue Analyst SAC License No.: S0080521010008 SFC CE Ref: BJV008
Chang Jing Analyst SAC Certification Number: S0080518110003 SFC CE Ref: BMX565
Yating Chen Analyst SAC License No.: S0080523120006
Legal Notices