Spending generously is just the first step

Written by Meng Huiyuan

Edited by Li Wenjie

Typeset by Annalee

"2025 will mark the beginning of the elimination of non-high-level intelligent driving vehicles, and cars without intelligent driving capabilities will gradually lose their competitiveness." In recent days, several executives from new energy vehicle companies have rapidly elevated the value of "intelligent driving" to tens of billions of yuan through their statements.

An industry background worth noting is that, as the reshuffling accelerates, more and more new energy vehicle companies are realizing that intelligent driving is not just a major selling point that influences consumer decisions, but also the key to their battle in the second half. Liu Yudong, an investment manager at Chentao Capital, even made such a judgment: "End-to-end automation has opened up a second growth curve for L4 commercialization."

By stacking computing power, data, and algorithms, suppliers are vying to carve out their territories. Behind this competition lies not only differences in end-to-end strategic planning but also gaps in capital strength. Judging from the continuous heavy investments in supercomputing centers and the expanding R&D teams made by new energy vehicle companies such as Tesla, Huawei, and the "NIO-Xpeng-Li Auto" trio, this end-to-end race has just begun.

01

Full-stack in-house research and development requires substantial investments

"In the era of automotive intelligence, I can confidently say that without an investment of 50 billion yuan, it is impossible to excel in intelligent driving. Therefore, if anyone still believes that intelligent driving can be successfully achieved with just a few billion yuan, I think such a product is likely to become a killer on the road."

"Regarding the 50 billion yuan, we need to determine whether it's a one-time investment or a long-term investment. As mentioned earlier, we invest 1 billion USD annually in intelligent driving R&D, which would exceed 50 billion yuan over ten consecutive years."

"End-to-end automation lengthens the entire chain. We invest 3.5 billion yuan annually in AI, which requires substantial accumulation and time, whether for the construction of computing power or, more importantly, data collection."

Xia Yiping, CEO of Geely Auto

Recently, executives from new energy vehicle brands such as Geely Auto, NIO, and Xpeng have responded to questions about the cost of in-house intelligent driving research and development, inevitably elevating the topic's significance to the utmost level.

In fact, since the emergence of the new energy vehicle sector, "intelligence" has always been the ultimate label that related automakers strive to embody. From the initial smart cockpits to subsequent NOA autonomous assisted navigation driving and, most recently, end-to-end autonomous driving, although the focus of intelligence has shifted, new energy automakers' pursuit of "intelligent driving" remains steadfast and enduring.

From a technical perspective, end-to-end autonomous driving is a data-driven model. Therefore, the importance of training data is increasing. End-to-end automation's requirements for data encompass data volume, annotation, quality, and distribution. In addition to massive amounts of high-quality data, powerful computing power is also necessary to support model training.

As Lang Xianpeng, Vice President of Intelligent Driving R&D at NIO, said, "The core competition in autonomous driving R&D lies in having more and better data, as well as the computing power to match, for model training. Acquiring computing power and data requires substantial investments and resources."

Elon Musk, who was among the earliest to deploy this technological approach, has repeatedly stated that "the iteration of the FSD V12 end-to-end model is primarily constrained by cloud computing resources." He chose to invest heavily in boosting computing power—Tesla plans to invest over 1 billion USD in the DOJO supercomputing center by the end of 2024, aiming to increase its total computing power to 100,000 PFLOPS, equivalent to the combined power of approximately 300,000 NVIDIA A100 GPUs.

Domestically, new automakers represented by NIO, Xpeng, and Li Auto, along with established automakers like Geely and Changan, have either chosen to build their own intelligent computing centers or partner with third parties to prepare for them. For instance, NIO has collaborated with Tencent to establish an intelligent computing center. Although the specific capabilities of this supercomputing center have not been disclosed, Li Bin once described NIO's computing power layout as "insane" and claimed that it would remain at the global forefront for the next one to two years.

Intelligent driving suppliers represented by Huawei, SenseTime Jueying, and Horizon Robotics are also keeping pace. It is reported that Huawei's BU Cloud Intelligent Computing Center's Qiankun ADS 3.0 boasts 3,500 PFLOPS of computing power and trains on 30 million kilometers of daily driving data. Based on a global road network of approximately 64 million kilometers, the system can fully cover this distance in just 2.1 days.

Image source: Autohome

It is evident that advancing end-to-end technology requires substantial investments in high-cost computing power, which not all automakers have the capability and resources to undertake. It is reported that most domestic companies researching end-to-end autonomous driving currently possess training computing power on the order of thousands of GPUs. As end-to-end automation gradually shifts towards large models, training computing power becomes increasingly scarce.

02

Striving for mass production by 2025

"Based on the aforementioned progress, domestic autonomous driving companies' modular end-to-end solutions may enter mass production by 2025," predicts Chentao Capital's "End-to-End Autonomous Driving Industry Research Report" regarding when end-to-end architectures will be integrated into vehicles.

Even though end-to-end autonomous driving is bound to face the challenge of training data, as the hottest technical concept in the automotive industry today, new energy manufacturers' eagerness to compete is clear.

Since the beginning of 2024, news of end-to-end technology being implemented in intelligent driving has been constant: In April, Huawei officially announced its new intelligent automotive solution brand "Qiankun," focused on intelligent driving, at its Smart Car Solutions Conference and unveiled the ADS 3.0, adopting an end-to-end architecture. In June, Great Wall Motors Chairman Wei Jianjun showcased the actual performance of Great Wall Motors' NOA in Chongqing through a live stream, powered by the company's latest-generation intelligent driving system based on a modular end-to-end architecture. At the "AI DAY" event on May 20th, Xpeng announced that it would begin pushing intelligent driving and smart cockpit systems based on end-to-end large models to users immediately. At the 2024 China Automotive Chongqing Forum, Li Auto revealed that it expects to launch a more refined autonomous driving system trained on over 10 million clips by the end of this year or early next year, offering users supervised L3 autonomous driving experiences. In July, BYD's premium brand Denza announced the completion of its "mapless" end-to-end solution, marking the first phase of achieving intelligent driving. In the same month, NIO officially unveiled NWM—NIO World Model, China's first intelligent driving world model, at the 2024 NIO Innovation Technology Day.

Huawei's end-to-end architecture ADS 3.0

From titles like "the first domestically mass-produced vehicle" and "the industry's first dual-system mass production solution" to "the first domestically end-to-end integrated model" and "the first application of end-to-end technology in AEB," while verifying the authenticity of these accolades attributed to related new energy products may be challenging, it is evident that corresponding automakers have made substantial investments in these products—as demonstrated by the growing size of their R&D teams.

In this regard, some automotive bloggers have conducted relevant statistics: Huawei (over 7,000 employees in December 2023, though this figure likely represents the total number of Huawei's automotive BU employees), BYD (over 4,000 employees in Q1 2024), Xpeng (over 3,000 employees in Q1 2024), Zeekr (over 1,500 employees in February 2024), NIO (over 1,300 employees in November 2023), and Li Auto (expected to be below 1,000 employees in May 2024).

Image source: New Channel Observation

Even Xiaomi, a late entrant, has revealed its ambitions through Lei Jun, stating, "Xiaomi's in-house intelligent driving research and development costs over 2 billion yuan annually, which is enormous. Few domestic manufacturers fully in-house develop everything from start to finish." According to Lei Jun, Xiaomi has set a clear goal from the outset to independently develop all intelligent driving solutions. Currently, Xiaomi's intelligent driving team comprises over 1,000 engineers.

Taking NIO as an example, which was exposed to be establishing a professional intelligent driving research institute during its second-quarter earnings call in August, NIO's R&D expenses for intelligence increased by 50% year-on-year in the first half of 2024, "primarily due to increased investments in intelligent driving." Zhu Jiangming, Chairman of NIO, added that NIO's R&D investments would maintain a 50% growth rate throughout 2024. In comparison, NIO invested approximately 3.33 billion yuan in R&D from 2022 to 2023, while Li Auto invested 17.366 billion yuan during the same period.

Who will secure the top tier?

As the end-to-end competition commences, automakers entering this stage will require increasingly higher training computing power, which necessitates greater investments in capital, manpower, and time. When this multi-dimensional competition involving computing power, algorithms, and data reaches its conclusion, it will ultimately pit the comprehensive strengths of the underlying new energy automakers against each other.

One of the most crucial aspects is how much new energy automakers are willing to invest in technological R&D.

For related automakers, funding sources can be divided into external and internal ones. An external case in point is the UK startup Wayve, which secured over 1 billion USD in funding in May 2024 for its end-to-end autonomous driving and large autonomous driving model technologies. This foreshadows the warming of end-to-end technologies in the capital markets, considering the last multi-billion-dollar funding round in the autonomous driving sector dates back to Waymo's 2.5 billion USD funding in the first half of 2021. Since then, the global autonomous driving investment and financing market has entered a downturn.

However, for domestic new energy automakers, relying on the unpredictable favor of the capital markets may be less reliable than betting on technological breakthroughs through self-sustaining capabilities.

From this perspective, domestic automakers can be divided into two camps: those like NIO, Xpeng, and Li Auto, and established automakers transitioning to new energy like BYD. Among the latest financial reports of ten automakers, nine, including BYD, have invested more in R&D than their net profits. SAIC Motor invested 8.96 billion yuan in R&D while earning 6.63 billion yuan in net profits; Geely Auto achieved 3.37 billion yuan in net profits but invested 4.55 billion yuan in R&D; Changan, Dongfeng, and GAC also invested more in R&D than their net profits… These figures also demonstrate the financial strength of established automakers in technological R&D to a certain extent.

Of course, the performance of new automakers is also noteworthy. In the first half of this year, Li Auto delivered 189,000 vehicles, maintaining a clear lead. NIO delivered 87,400 vehicles, still enjoying substantial year-on-year growth. Xpeng delivered 52,000 vehicles in six months, ranking relatively low and facing the risk of falling behind. In terms of profitability, Li Auto is undoubtedly the most profitable among NIO, Xpeng, and Li Auto. In the second quarter, although Li Auto's net profit declined by 51.94% year-on-year, it still recorded 1.102 billion yuan, achieving profitability for the seventh consecutive quarter. NIO and Xpeng continued to incur losses, but encouragingly, both saw substantial narrowing of their losses.

These achievements lay the foundation for new automakers' substantial investments in R&D: In the first half of 2024, NIO invested 6.083 billion yuan in R&D, accounting for 22% of its revenue; Xpeng invested 2.817 billion yuan in R&D and incurred 2.96 billion yuan in sales and administrative expenses; Li Auto invested 6.076 billion yuan in R&D, a 42% year-on-year increase from 4.278 billion yuan in the same period last year, representing an additional investment of approximately 1.8 billion yuan in R&D.

"Before end-to-end mass production, several hurdles must be overcome: first, the preparation of vehicle-end computing power; second, the iteration of end-to-end algorithms; third, the scale of cloud data; fourth, the scale of computing power; and fifth, the validation scheme," said Mao Jiming, Vice President of Engineering at Jiake Technology. Currently, leading automakers and companies such as Tesla, domestic players like NIO, Xpeng, Li Auto, and Huawei have already established capabilities in vehicle-end computing power, cloud data scale, and cloud computing power scale. By the end of this year or early next year, the end-to-end algorithms of several leading automakers will be ready for large-scale vehicle integration. From the second half of next year onwards, the industry will witness a surge in end-to-end mass production for vehicles.

Li Auto's cloud model

According to a research report issued by Cinda Securities, as intelligent driving algorithms gradually converge towards an end-to-end architecture, data and computing power will become core competitive factors. Leading automakers or suppliers can gain access to more and better "data" as well as stronger and faster "computing power." Excellent intelligent driving capabilities are expected to enhance sales conversions, ultimately strengthening the Matthew Effect among automakers and empowering the stronger players in intelligent driving.

As for who will be the first to embark on the second growth curve of L4 commercialization, that remains to be seen by the market.