Source | BohuFN

Recently, the arms race in the intelligent driving industry seems to have accelerated significantly. XPeng Motors officially launched its AI Tianji system, XOS 5.2.0, to global users, making it one of only two automakers globally to achieve end-to-end large model mass production.

Li Auto and NIO are also hustling to keep pace. Li Auto publicly unveiled its end-to-end autonomous driving technology architecture for the first time, while NIO officially released its NAD Arch 2.0 intelligent driving technology architecture, upgrading its algorithm layer to introduce an end-to-end architecture.

As the 'end-to-end' large model sweeps through the automotive industry, the automotive sector has entered a 'new AI battlefield,' with not only representatives of emerging forces like NIO, XPeng, and Li Auto but also tech companies like Huawei and Xiaomi entering the automotive space. However, as large models accelerate their integration into vehicles, it remains unclear which force – automakers or tech companies – will emerge victorious, and what kind of 'new game' the automotive ecosystem will face.

01 'End-to-End' Large Models Sweep the Automotive Industry

In April this year, Tesla CEO Elon Musk visited China, sparking speculation about whether Tesla's Full Self-Driving (FSD) would enter the Chinese market. During Tesla's Q2 earnings call, Musk revealed that FSD was expected to be approved in China and Europe by the end of the year.

The official countdown for FSD's entry into China is both pressure and a challenge for the new energy vehicle sector. Tesla has always been in the minority in the intelligent driving field. Last August, Tesla released FSD Beta V12, which Musk described as the world's first end-to-end AI autonomous driving system.

What is 'end-to-end'? Simply put, it involves replacing the traditional perception, planning, and control modules with an AI model. Instead of executing code written by engineers, the system thinks and acts like a human.

This sparked a strong reaction in the autonomous driving industry, with many automakers and autonomous driving companies starting to develop end-to-end technology, and domestic intelligent driving routes also began to shift.

Starting in the second half of 2023, NIO set up a separate large model department to oversee end-to-end model research and development. Lang Xianpeng, VP of Intelligent Driving at Li Auto, stated that when Tesla released FSD V12, Li Auto had already begun pre-research on end-to-end technology, and FSD V12's performance further solidified their confidence.

Entering 2024, end-to-end technology began to accelerate its integration into vehicles. In addition to NIO, XPeng, and Li Auto having released end-to-end intelligent driving technology achievements, domestic automakers, tech companies, and intelligent driving companies have also started taking action.

In April, Huawei released Kunpeng 3.0, which is expected to debut on the Enjoy S9. Huawei's Kunpeng ADS 3.0 version has end-to-end capabilities and will significantly enhance intelligent driving capabilities with the assistance of LiDAR.

In June, Great Wall Motors Chairman Wei Jianjun demonstrated the actual performance of Great Wall Motors' NOA in Chongqing through a livestream. Behind this was Great Wall Motors' latest generation of intelligent driving systems, utilizing a modular end-to-end architecture.

In July, BYD's premium brand Denza announced that it had completed the research and development of an end-to-end solution without relying on maps, marking the first phase of realizing intelligent driving.

Judging from the release dates of these end-to-end products, most automakers likely began researching and developing end-to-end large models in the second half of last year. However, catching up with the 'deadline' does not necessarily mean achieving a one-step solution.

Following Tesla, domestic automakers and intelligent driving companies are rapidly advancing in end-to-end technology. Huawei and XPeng adopt a segmented 'end-to-end' approach, replacing perception and planning modules with models, connected by manually written rules in the middle;

Li Auto, on the other hand, has taken a more significant step by replacing both the perception and planning modules with a single model. The industry describes this as a step back from Tesla but a step ahead of domestic peers.

However, compared to Tesla's end-to-end solution, which boasts "input image, output control," domestic end-to-end approaches can currently only achieve from perception to prediction and decision-making. The final control execution module is still backed up by manually written rules from engineers.

Jiang Haipeng, Senior Director of Great Wall Motors' Intelligent Platform Development Center, stated that almost every algorithm company or OEM is researching end-to-end solutions, but no more than three have implemented end-to-end architectures.

Li Xiang, Chairman of Li Auto, also mentioned the challenges of end-to-end technology, including the need for talent to conduct true end-to-end data training, high-quality data, and sufficient computing power.

Algorithms, data, and computing power are considered the three major challenges for end-to-end implementation. Due to these obstacles, the industry has been unable to reach a consensus on whether end-to-end is the future of intelligent driving.

However, there are always more solutions than problems. In terms of computing power, Li Auto stated that the company has 5,000 A100 and A800 training cards, twice as many as XPeng, and a healthy profit margin to support card rentals;

Regarding data, XPeng stated that based on over 1 billion miles of video training, over 6.46 million cumulative kilometers of real-world testing, and over 216 million cumulative kilometers of simulation testing, XPeng's end-to-end large model can achieve "two-day iterations."

For the highly competitive automotive sector today, being the first to implement end-to-end model mass production and accelerate the rollout of city NOA is a crucial strategy for automakers to compete for market share and enhance brand competitiveness.

02 Imagining the Possibilities of Large Models in Vehicles

However, it's not just automakers feeling the heat; tech companies and intelligent driving companies are also moving swiftly. SenseTime's Pure Vision end-to-end autonomous driving general model, UniAD, can achieve an integrated perception and decision-making "one-stage" solution, similar to Tesla's FSD in its fully visual and autonomous driving approach.

Currently, SenseTime's mass-produced intelligent driving products have been implemented in various brands and models, including GAC Aion LX Plus, Hozon Auto Nezha S, GAC Aion Hyper GT, and Hongqi, with high-speed NOA functionality also starting to roll out.

DeepRoute IO's advanced intelligent driving platform, also applying an end-to-end model, is reportedly responsible for BYD's end-to-end intelligent driving project. If successful, this could establish DeepRoute IO as a key partner for China's leading new energy vehicle sales.

Compared to emerging forces like NIO, XPeng, and Li Auto, which are focusing their firepower on intelligent driving, internet giants and tech unicorns are exploring more applications of AI large models in vehicle scenarios, not just in intelligent driving but also in smart cockpits, production and R&D, and marketing promotion.

In smart cockpits, large models can significantly improve human-machine interaction in smart vehicles, catering to users' habits across various scenarios through voice, visual, and gesture interactions.

NIO's large model, NOMI GPT, has officially launched, enabling open-ended Q&A interactions with users. It can also integrate third-party APIs to build AI agents, achieving seamless connection across NIO's products, services, and communities.

iFLYTEK launched its new "Feiyu Smart Cockpit System," closely integrated with diverse in-vehicle usage scenarios like driving, communication, and entertainment. It has already been applied in over ten automakers, including Chery, GAC, and Great Wall Motors.

In production and R&D, automakers can leverage large models to enhance production efficiency and quality. For instance, China FAW and Alibaba Cloud's Tongyi Qianwen jointly created the automotive industry's first large model BI application, automating tasks that previously took 50-80 days and reducing the timeline to days or even seconds.

In marketing promotion, Tencent launched its "Omni-Intelligent" large model solution for the automotive industry, offering a full-chain service from models, computing power, AI engineering platforms to AI applications, covering core scenarios like automotive R&D, production, and marketing. Tencent's automotive large model is currently being applied in collaboration with over ten automotive industry partners, including Changan, GAC, and FAW Toyota.

Whether emerging automakers, OEMs, or tech unicorns, all are targeting large model applications in vehicle scenarios. According to incomplete statistics, over ten automotive brands have already integrated large models, and an increasing number of companies are announcing related layouts.

From the "Hundred Models War" to the "Application War," why have vehicle scenarios become attractive to large model enterprises? On one hand, it's due to the commercial opportunities presented by automotive scenarios.

Baidu's founder, Robin Li, once mentioned that instead of over-competing in basic large models, it's time to focus on scenario applications. Without applications, basic models alone are worthless.

Large models need to find landing scenarios, but currently, there aren't many viable options. While some large model applications have emerged in areas like text-to-video, text-to-text, and smart home appliances, they're still far from commercial profitability. In contrast, the rapidly developing and mature smart automotive sector presents ready-made commercial demands.

On the other hand, as mentioned earlier, AI large models can empower various aspects of the automotive industry, including production processes, intelligent driving, smart cockpits, and sales promotion. While some areas may not yet be fully 'AI-enabled,' they still offer vast imagination space for large model enterprises.

Finally, large models present new development directions for the automotive industry. For instance, Huawei, which has made it clear it won't manufacture cars, can deeply engage in branding, product definition, vehicle design, data generation, and data closed-loop through HarmonyOS Intelligent Driving.

Under the trend of "Intelligence for All," the concept of a "human-vehicle-home ecosystem" is not just a slogan for Xiaomi but can become the future of any automotive alliance, creating a super-intelligent ecosystem that proactively serves people through seamless hardware connections, revolutionizing the future of the automotive industry.

03 The Game Between Automakers and Tech Platforms

However, in the field of vehicle applications, more competitors also mean the 'crowding-out effect.' Judging from the current speed of large model integration into vehicles, it doesn't seem as difficult as initially thought. The challenge lies in demonstrating value post-integration.

Currently, large model integration into vehicles has yet to achieve large-scale commercialization, with most enterprises still in the 'demonstration' phase. Transitioning from PPT presentations to actual applications and assessing their ultimate effects, value proposition, and ability to translate into performance and profits, remains to be seen.

Firstly, let's consider the effectiveness. Taking intelligent driving systems as an example, evaluation metrics should include aggressiveness, dangerous takeover miles, recognition capabilities, and importantly, driver personality matching.

However, as intelligent driving becomes standard in cars, an increasing number of car owners report erratic lane changes and high-risk driving under the guidance of intelligent driving systems. For owners to willingly pay for intelligent driving, the systems cannot merely be ' chicken ribs .'

Similarly, popular smart cockpits also face questions about whether their scene interaction capabilities, based on large models, will evolve into 'super apps' or remain mere 'response machines.' Much remains unknown.

Next, let's consider value. At this stage, sales can be a relatively 'practical' indicator of how much value large model integration can bring to enterprises.

However, 'intelligence' is still a 'value-added option' rather than a 'must-have' for cars. Among the factors influencing purchasing decisions, quality, performance, and design still take precedence. The premium that large models can add to sales is limited. When factoring in the costs automakers incur for developing and training large models, the question of 'recouping investments' becomes even more crucial for automakers to ponder.

However, in terms of enhancing production and operational efficiency, the value of large models will be more directly experienced. In the long run, large model integration can empower the entire automotive ecosystem, but this depends not only on the models themselves but also on joint exploration and mutual support among automakers, large model vendors, and tech ecosystem platforms.

Currently, vehicle scenarios offer one of the fastest-growing, most diverse, and clearest commercial paths for large model applications. All parties want a piece of the pie, and the game between them seems far from over.

For instance, Geely announced last year that it would explore technical cooperation with Alibaba Cloud in large model-related scenarios, but in January this year, Geely pivoted to launching its full-stack, self-developed, all-scenario AI large model – the Geely Xingrui AI Large Model.

In February this year, XPeng announced a collaboration with Volkswagen to develop B-segment vehicles, with XPeng primarily taking on a technology provider role. Industry insiders speculate that XPeng may intend to sell its XNGP intelligent driving technology, positioning itself as a second 'HarmonyOS Intelligent Driving.'

The competition within the intelligent driving supplier landscape is also intense. With the gradual rollout of 'end-to-end' solutions, the chances of top suppliers securing orders are increasing, while lower-tier suppliers focused on basic intelligent driving solutions are seeing their profit margins significantly compressed.

It's evident that as large models accelerate their integration into vehicles, the automotive industry's internal ecosystem positioning and benefit distribution are in a state of 'chaos.' Each 'player' still lacks clear roles and boundaries, leading to both opportunities for multiple winners and intense internal competition.

For the automotive industry and ecosystem to achieve more sustainable and healthy development, it's crucial to avoid meaningless internal conflicts and establish clear roles and responsibilities for each 'player.' Only when large models can generate profits for all parties post-integration will the ecosystem thrive and grow.

*The cover image and accompanying illustrations belong to their respective copyright holders. If the copyright owner believes their work is not suitable for public viewing or should not be used free of charge, please contact us promptly, and we will rectify the situation immediately.