"Intelligent driving" turns into "assistance". Is it just a safety yellow light that's on? | Krypton - Hard Tech

Author | Geng Chenfei

Editor | Song Wanxin

The rapid advancement of "intelligent driving" has been put on hold. Half a month after the accident involving the Xiaomi SU7, the Ministry of Industry and Information Technology issued regulatory documents to restrict intelligent driving functions and their promotion. Some participants at the meeting revealed several key points: L2 intelligent driving systems are not allowed to use the term "takeover" in their promotion; applications for functions such as "valet parking" that drivers cannot fully control will no longer be accepted; and public beta testing is not permitted.

The latest news is that on May 4th, Xiaomi adjusted the wording on the pre - order page of the new SU7, changing "intelligent driving" to "assistive driving".

Before the car accident that triggered an outcry, several new - energy vehicle manufacturers, including Xiaomi, conducted a round of rights issues to raise funds, gearing up to promote "equal access to intelligent driving".

In March, Xiaomi and BYD raised HK$42.5 billion and HK$43.5 billion respectively through rights issues. NIO also announced plans to place approximately 140 million Class A ordinary shares to raise HK$4.03 billion. Among them, BYD and Xiaomi both chose to conduct rights issues at high stock prices, while NIO carried out the placement when its stock price was at a historical low.

The news spread to the secondary market, triggering concerns. After the rights issue announcements, BYD's stock price fell by more than 7%. Xiaomi's market value evaporated by nearly HK$100 billion in a single day, and NIO's stock price also dropped by more than 8% at one point. In particular, Xiaomi's sharp decline further dragged down the Hang Seng Tech Index, which fell by 3.82% in a single day.

Since vehicle manufacturers started promoting intelligent driving, it has always accounted for a large proportion of the total vehicle cost. Data from Gasgoo shows that last year, R & D on intelligent driving accounted for 38% of vehicle manufacturers' total expenditures, becoming the second - largest cost item after batteries.

Taking the three vehicle manufacturers that conducted rights issues as examples, BYD plans to invest up to 100 billion yuan in intelligent R & D. As of the end of 2024, NIO had invested up to 46 billion yuan in intelligent driving R & D. On the Xiaomi side, Lei Jun revealed that the company spends more than 2 billion yuan annually on intelligent driving R & D.

Vehicle manufacturers were eager to expand the scale of intelligent driving through mid - and low - end models to reduce costs, but unexpectedly, a warning sign appeared, and the costly intelligent driving began to become a hot potato.

01 Difficult to Reduce Intelligent Driving Costs

Algorithm expert Fu Cong told 36Kr that, depending on different models and configurations, the overall cost of current intelligent driving systems on the market ranges from a few thousand yuan to 20,000 - 30,000 yuan, accounting for about 5% - 15% of the total vehicle cost. With the addition of high - level functions such as lidar and redundant control, this proportion will further increase.

In terms of R & D models for intelligent driving, new - force manufacturers represented by NIO, XPeng, and Li Auto are advocates of self - research. However, Fu Cong said that in practice, most vehicle manufacturers adopt a hybrid model of "self - research + external procurement".

For example, Li Auto uses its self - developed AD Max intelligent driving system in high - end models, while in mid - and low - end models, it adopts solutions from external suppliers. The low - level version of BYD's Tian Shen Zhi Yan system is also sourced externally.

According to Fu Cong, this approach can ensure the autonomy and controllability of core capabilities and accelerate product launch through mature external solutions. However, external procurement can increase the overall cost of the intelligent driving system in some aspects. For example, the cost of externally procuring a combined navigation system can reach tens of thousands of yuan, while self - research only costs about a thousand yuan.

NIO's management also revealed at an earnings conference that compared with using four Orin chips, its self - developed Shenji NX9031 intelligent driving chip can save about 10,000 yuan in costs.

At the hardware level, lidar is the core cost. Although the price of lidar has been declining in recent years and has reached about a thousand yuan, in the cost - effective automotive market, the use of multiple lidars still incurs a significant cost.

Against this background, vehicle manufacturers have started to adopt differentiated intelligent driving solutions: They mainly promote the more cost - effective pure vision solution in basic models, while in high - level versions, they are equipped with lidar, adopting a "pure vision + lidar" fusion approach.

For example, in Xiaomi's SU7, the Xiaomi Pilot Pro intelligent driving system loaded in the standard NOA version uses a pure vision mode. The Pro, Max, and Ultra versions offer the higher - level Xiaomi Pilot Max intelligent driving system, which uses a fusion perception mode of vision + lidar.

Even NIO, which has always adhered to the vision + lidar fusion solution, has launched a pure vision version for its new brand, LeDao.

These changes indicate that vehicle manufacturers are under great pressure to reduce intelligent driving costs and are trying their best to lower the cost of each hardware component. However, the reality is that reducing hardware costs is only superficial, and the pure vision solution has significant hidden costs.

The pure vision solution relies on a large - scale data training model.

"The algorithm needs to continuously adapt to new cities, new scenarios, and new regulations. The cost of model training, including servers, computing power, and personnel, is relatively high. In terms of data annotation, leading companies often invest tens of millions or even hundreds of millions of yuan. This is because autonomous driving has extremely high requirements for data quality, especially for fine - grained annotation involving multi - modal fusion and 3D semantics. The cost of a single piece of data can reach several hundred or even thousands of yuan."

Fu Cong told 36Kr that algorithm and data annotation are the most difficult parts to cut in the R & D cost of the pure vision solution.

It can be seen that no matter which technical path is chosen, intelligent driving is a money - burning endeavor. So far, most vehicle manufacturers that have bet on intelligent driving as the next decisive factor in the industry have not seen returns.

Xiaomi sold 136,900 vehicles in 2024 but suffered a full - year loss of 6.2 billion yuan, with an average loss of about 45,000 yuan per vehicle. NIO's situation was even worse, with a loss of 100,000 yuan per vehicle.

02 Waiting for the Scale - up of Intelligent Driving

Years of efforts to reduce hardware costs have had little impact on vehicle manufacturers' financial statements. Now, promoting scale - up has become the core path to reducing intelligent driving costs.

At the beginning of this year, BYD launched the "Equal Access to Intelligent Driving" initiative. The competition in intelligent driving has instantly spread from a local war in the high - end car market to a full - scale war in various market segments.

In February, BYD announced that it would equip all models in the Dynasty and Ocean series with the high - level "Tian Shen Zhi Yan" intelligent driving function. All models priced above 100,000 yuan will come standard with this function, and most models below 100,000 yuan will also be equipped with it. From the 78,800 - yuan Seagull to the 249,800 - yuan Song L EV, the first batch of 21 new models all have high - level intelligent driving capabilities, and there is no price increase for the additional features.

This has directly disrupted the original pricing logic of the intelligent driving industry. Before this, the lowest price of vehicles equipped with intelligent driving was generally above 200,000 yuan.

Facing BYD's aggressive strategy, some industry insiders believe that this is a way to "put it differently" to further extend the intelligent driving functions originally concentrated in high - end models to mid - and low - priced models to accelerate the popularization process and drive the scale - up of intelligent driving.

There is a consensus in the industry that mass - scale production is the key to reducing intelligent driving costs by 6%. Only after large - scale implementation can the R & D costs be spread through high sales volume. The real key to the large - scale adoption of intelligent driving lies in whether the market in the mainstream price range of 100,000 - 200,000 yuan can be opened up.

Based on compulsory traffic insurance data, Everbright Securities estimated that the penetration rate of L2+ urban intelligent driving in China was about 5% - 6% in 2024. Among them, the penetration rate of L2+ urban intelligent driving in the 250,000 - 400,000 - yuan price range has reached over 20%. In contrast, the penetration rate of L2+ urban intelligent driving in the 100,000 - 200,000 - yuan price range is still less than 0.2%.

This situation is caused by both cost factors and users' perception. "Ordinary users' understanding of intelligent driving still has a lot of room for improvement," said Fu Cong.

Especially in mid - and low - priced models with downgraded intelligent driving configurations, the reduction of functions further magnifies the safety risks caused by users' lack of understanding.

In the Xiaomi car accident, the standard version of the Xiaomi SU7 used a pure vision solution. "Driving at night and encountering road construction is extremely challenging for a system that relies on a pure vision solution," said an industry insider.

Fu Cong pointed out that currently, intelligent driving systems can "perform excellently" in tests or demonstrations, largely because most demonstration scenarios are included in the model training data, so the system can handle them with ease. However, real - life situations are much more complex than the training environment. For so - called "corner cases", which are extremely rare but highly complex scenarios, due to insufficient data coverage, they are often difficult to solve completely through traditional training methods.

"For example, when the ground markings are unclear and cannot be recognized, the camera is blocked in extreme weather, or there are special traffic scenarios such as traffic police directing with hand gestures, it is difficult for the in - vehicle model to handle all situations comprehensively."

The industry once regarded "intelligent driving capabilities" as the key for vehicle manufacturers to "enter the game". However, now that the scale - up has been forced to slow down, intelligent driving may become a heavy burden on vehicle manufacturers in the short term.

Follow for more information