During the eight-day long holiday, intelligent driving entered a "free-for-all".

During the recent National Day holiday, the quieter the intelligent driving brands were, the crazier the traffic was.

A report from CCTV.com shows that during this National Day holiday, an average of 12.5 million new energy vehicles hit the road every day. This figure represents a 30% increase compared to the same period last year and a 70% increase compared to normal days.

The huge travel demand has brought about complex traffic scenarios, which also present an excellent opportunity to test the effectiveness of intelligent driving products.

For this reason, in the past few years, the National Day holiday has often been a key moment for car companies to showcase their intelligent driving achievements:

In the first week after the National Day holiday in 2023, travel reports from brands such as NIO, XPeng, Li Auto, ZEEKR, and Avita were released one after another;

In 2024, regarded as the "Year of Intelligent Driving" in the industry, the intelligent driving reports of Huawei and Li Auto were released within 48 hours after the holiday ended.

In contrast, after this Golden Week with a sharp increase in travel volume, car companies have remained rather silent.

As of now, among the leading new energy car companies, except for Huawei and Xiaomi, there are no longer any intelligent driving - related reports on the information release channels of most brands.

Correspondingly, the new energy industry has marked the position of the front - line battle in the intelligent driving track with another big piece of news.

Public information shows that in the first week after the National Day holiday, there were significant personnel and organizational structure adjustments in the intelligent driving teams of XPeng and NIO, involving changes among many senior executives. In particular, the person in charge of XPeng's intelligent driving business was directly replaced. Li Liyun is no longer the head of the Autonomous Driving Center, and Liu Xianming, the head of the World Base Model, has taken over the position.

With the silence of the reports and the replacement of senior executives, between the quietness and the change, from the "Year of Intelligent Driving" in 2024 to the "Year of Mass - Adoption of Intelligent Driving" in 2025, behind the simple change of two words, the rules of the game in this track have quietly changed.

1. In the Year of Mass - Adoption of Intelligent Driving, the Industry Has Changed Its Gameplay

From the decline in the volume of reports to the adjustment of business teams, an obvious trend has emerged:

From the "Year of Intelligent Driving" to the "Year of Mass - Adoption of Intelligent Driving", the industry is no longer competing in a numbers game of popularity and penetration rate, but rather in a "bloody fight" at the technical level.

This can be glimpsed from the intelligent driving data of car companies.

Taking Huawei - affiliated models as an example, public information shows that during the Golden Week in 2025, the cooperative models of Huawei Qiankun Intelligent Driving achieved a total assisted - driving mileage of 294 million kilometers, including 243 million kilometers on highways and 51 million kilometers in urban areas.

Compared with the data from last National Day, where are the changes? Two points are worth noting:

Firstly, the total intelligent driving mileage. During this National Day holiday, the total intelligent driving mileage of Huawei - affiliated models was 3.4 times that of the same period last year, and the active proportion of assisted - driving users also reached 90.8%;

Secondly, the proportion of urban assisted driving. During this National Day holiday, the proportion of urban assisted - driving mileage of Huawei - affiliated models was 17.3%, showing only a slight increase compared to the same period last year. Correspondingly, the average speed of urban assisted driving of Huawei - affiliated models during this National Day holiday decreased by 1.5 km/h compared to last National Day.

Based on the above two points, it is not difficult to find that for new energy car companies, convincing users to try intelligent driving on highways is no longer a problem.

Conversely, the key to the real mass - adoption of intelligent driving does not lie on the wide and smooth highways with light traffic, but in the narrow and accident - prone urban areas.

For car companies struggling in the red - ocean market, after competing in areas such as refrigerators, color TVs, battery life, and charging efficiency, the one that can be the first to launch reliable L3 and L4 - level intelligent driving in urban areas will be able to gain an advantage in both the capital market and the consumer market.

From a technical perspective, urban intelligent driving is also the "long - tail" of the previous technical routes.

Since Tesla pioneered the end - to - end route in 2019, the method of training intelligent driving models by collecting actual road test data has been widely emulated by global car companies. Although this technical route has indeed promoted the rapid evolution and iteration of intelligent driving models in the short term, it has also exposed its shortcomings in subsequent implementation scenarios:

Especially in urban areas, suddenly emerging electric vehicles, pedestrians, complex road sections under temporary closure for repairs, and vehicles with illegal driving behaviors... These "killers" that threaten driving safety rarely appear in traditional road test data, and it is difficult for intelligent driving models to learn and get feedback from such fragmented and accidental data.

The limitations of the end - to - end model have made the demand for "route innovation" in intelligent driving more urgent.

This is why, in the past two months, from Li Auto, XPeng, and DeepRoute.ai successively announcing the installation of VLA large models in vehicles to Huawei and NIO frequently speaking out about focusing on the WA model, an era of "great discovery" of new intelligent driving technologies has quietly arrived.

As the industry crosses into a new era, how will the players remaining at the table make their choices?

2. How Do the "Top Students" Pass the Big Test of Intelligent Driving?

As of now, the leading brands have developed three evolutionary ideas in response to the traditional end - to - end model in the industry.

Using learning as an analogy:

The "reformists" represented by Momenta believe that there is a problem in the "learning process" of intelligent driving. In the traditional end - to - end model, the quality of the data fed into the large model is not good enough, and the "trial - and - error" and "reward" in the learning process are not prominent enough.

Therefore, Momenta advocates replacing the traditional route with a one - stage end - to - end model based on reinforcement learning. Compared with the large model that does traditional teaching - assistant exercises, Momenta's R6 Flywheel large model, which is like doing the "Huanggang Test Papers", has indeed improved its performance in dealing with difficult problems (such as detouring around construction areas and avoiding obstacles at night);

In contrast, the "practicalists" represented by Li Auto, XPeng, and DeepRoute.ai focus on optimizing the "test details". Just like in an exam, no matter how many similar questions you have done in road driving, you cannot directly apply the answers, and you must always be aware of the hidden traps in the questions.

Even the top students in the field of intelligent driving with large models must analyze specific problems specifically and build up comprehensive abilities of "observation, reasoning, and decision - making".

This is also the origin of the VLA technical concept. Different from the "data mapping" of the traditional end - to - end model, the VLA system can integrate three modalities: vision, language, and action. It can convert the information perceived visually into language descriptions, then conduct logical reasoning through the language model, and finally output specific action instructions. It can even predict the traffic conditions over a period of dozens of seconds.

The cost of this approach is obvious. Since there is an additional step, the demand for computing power and data of VLA far exceeds that of all traditional intelligent driving models. Some media have calculated that the cost of a single VLA training is 1.5 times that of DeepSeek - V3.

For this reason, brands such as XPeng and Li Auto have to increase their investment in computing power and build large - scale cloud training clusters to support the daily training of the VLA model.

Public data shows that as of August this year, in the global ranking of car companies' cloud computing power, except for Tesla, which leads the way with a cloud computing power of about 100 EFLOPS, the computing power of other leading intelligent driving brands such as Li Auto (8.1 EFLOPS), XPeng (10 EFLOPS), and Xiaomi (11.4 EFLOPS) is generally one order of magnitude lower. As for brands like Nezha, whose computing power is two orders of magnitude lower, their cloud computing power is hardly sufficient to support the efficient training of VLA.

Even so, for some brands, the current performance of VLA still fails to meet their requirements. In the view of the "idealists" represented by Huawei and NIO, both the traditional end - to - end model and VLA essentially rely on the "question - sea strategy" of feeding a large amount of road test data. Instead of just doing a lot of questions, truly understanding the teaching materials and the exam syllabus is the key to enabling intelligent driving to draw inferences from one instance and fear no test.

The most radical technical route in the field of intelligent driving - WA (World Model) was born as a result.

Different from the "vision - text - decision" logic of VLA, the core of the WA route is to simulate the real world in the cloud and create a "virtual digital world" for intelligent driving. This allows the intelligent driving model to fully learn and understand the logic of the real world in the virtual world, so as to handle real - world driving scenarios with ease.

In this regard, Wang Jun, the R & D leader of Huawei ADS, once made a vivid analogy: "If the intelligent driving system is compared to a student, VLA prepares for the exam by doing a large number of exercises and will be at a loss when encountering unseen questions; while WA first understands the knowledge points and can derive answers through logical reasoning no matter what new questions it encounters."

Coincidentally, Li Bin of NIO also said in an internal email: "WA gives the car 'imagination' rather than'memory'."

It is worth noting that in this personnel adjustment, Liu Xianming, who has newly taken the top position in XPeng's intelligent driving, was previously the head of the World Base Model. This is also interpreted by the outside world as XPeng's move closer to the WA technical route in intelligent driving.

Of course, the most radical technical route is bound to face the most severe tests.

In addition to the even more terrifying R & D cost compared to the VLA route, at the current stage, the implementation results of WA are far from being truly usable and user - friendly.

In contrast, VLA has been integrated into products and has become a core selling point to attract consumers. As of now, the first pure - electric SUV i8 launched by Li Auto in July this year has already installed the VLA large model. At the same time, after an OTA upgrade in September, the G7 Ultra under XPeng has also been equipped with the latest VLA model.

In comparison, Huawei and NIO, which embrace the WA model, are still in the "pre - dawn" stage before a technological breakthrough.

For more brands, the final whistle has not sounded yet, and the cruel competition in the intelligent driving industry continues.

3. The Qualifying Round Is Over, and the Elimination Round Begins

If the current intelligent driving competition is compared to a football game, in the past, brands were only competing for the "qualification to enter". Now, the most basic qualifying round has ended, and brands are about to face a real - life elimination - round confrontation. Facing the new battle situation in the "Year of Mass - Adoption of Intelligent Driving", staying in the game means facing a more brutal competition.

As He Xiaopeng said in March this year:

"Today, no company dares to say that it can sit back and relax and has passed the elimination round. Everyone is facing challenges, just to different degrees."

Back to the "battlefield of intelligent driving" that car companies collectively believe in, today's intelligent driving competition is no longer a battle between "maps and lidars", but a new battlefield of full - stack R & D capabilities and ecological collaboration.

Two trends have gradually emerged:

In the short term, since there has not been a truly revolutionary intelligent driving experience for users, the competition between WA and VLA, both in the first - tier, may not end immediately. However, it is certain that behind the fierce competition among leading brands, small and medium - sized intelligent driving manufacturers that "cannot keep up" are facing the "risk of being eliminated".

Previously, some media have reported that among unlisted intelligent driving companies, two have gone through bankruptcy liquidation, two are facing bankruptcy reorganization, and many are involved in acquisition and integration disputes.

Behind this, on the one hand, they are discouraged by the extremely high cost. VLA relies on excessive investment in computing power and data. Public data shows that the cloud training clusters of domestic small and medium - sized car companies generally stay at 0.2 - 0.6 EFLOPS, only one - tenth of Li Auto's 5.39 EFLOPS. Zhou Guang, the CEO of DeepRoute.ai, pointed out that an intelligent driving company needs to deliver 100,000 mass - produced vehicles to have the basic data foundation for building a VLA architecture.

On the other hand, they are discouraged by the requirement of integrated capabilities. In the past, small and medium - sized car companies could stay in the intelligent driving game by "piecing things together", such as buying chips, algorithms, and annotations. However, as the intelligent driving competition enters a deeper stage, leading players are building their own technical barriers through a self - developed closed - loop of "chips - data - models", which is also forcing the supply - chain ecosystem to make choices.