From NVIDIA to self-research: The crossroads of automotive companies developing their own chips

It is expected that advanced autonomous driving will 迎来 its "iPhone moment" in 2025. We believe that chips are currently the only hardware segment in autonomous driving that has a long-term logic of simultaneous increase in both volume and price.

Standing at a crossroads, we believe that under the strong constraints of performance + cost, future car manufacturers will gradually shift from NVIDIA's general-purpose GPUs to custom-designed ASICs; due to the high binding of ASICs with algorithms, compared to "cutting corners," self-developed chips that are "tailored to fit" will be the necessary path for leading autonomous driving companies. We believe that at this point, self-developed chips for leading companies have considerable feasibility and economic benefits, not only helping to reduce BOM costs but also serving as an important extension of the AI capabilities of car manufacturers, which will help ensure the continuous safety of the supply chain in the long term and reshape the valuation system of car manufacturers.

Chips are currently the only hardware segment in autonomous driving with a long-term logic of simultaneous increase in both volume and price.

Driven by AI-enhanced experience and the popularization of advanced autonomous driving by Huawei, XPeng, Li Auto, and BYD, we believe that advanced autonomous driving in China is about to 迎来 its "iPhone moment" in 2025, with the penetration rate of L2+ and above expected to rise from 14% to 30%.

After three years of an arms race on the hardware front, data + computing power have become the core driving forces for the current development of the industry. Cost reduction on the hardware side is a long-term trend for autonomous driving, and chips are currently the only hardware in autonomous driving systems with long-term price increases: the demand for computing power in autonomous driving chips is expected to grow from the current 200-500 TOPS to over 1000 TOPS with the end-to-end large model deployment, and a comprehensive upgrade of autonomous driving chips is expected in 2025.

We estimate that the proportion of chips in the BOM for autonomous driving will rise from 30% in 2024 to nearly 50% by 2026, with the greatest elasticity during the increase in autonomous driving penetration. We expect that the market size of the domestic autonomous driving chip industry will reach 22 billion yuan in 2025, a year-on-year increase of 121%, and the market size is expected to exceed 53 billion yuan by 2028.

In line with the trend of edge AI applications, ASICs are expected to gradually replace NVIDIA GPUs as a better solution for autonomous driving.

Autonomous driving chips are divided into general-purpose chips and dedicated ASICs. The former uses GPUs as the computing core, which can meet the computing needs of different scenarios but has significant computing power waste to ensure generality. He Xiaopeng stated at the 2024 XPeng AI Technology Day that the utilization rate of general-purpose chips like NVIDIA Orin-X is only 30%-40%; while ASICs use highly customized NPUs as the computing core, which can fully mobilize chip computing power when running specific algorithms, such as Tesla's FSD and Huawei's Ascend 610.

According to Gaishi Automotive, from 2022 to 2024, NVIDIA Orin-X's market share in the third-party market reached over 85%, making it almost the only choice for advanced autonomous driving outside of Tesla and Huawei. Meanwhile, the next-generation chip Thor has also been designated by car manufacturers such as BYD, ZEEKR, and Xiaomi. Investors generally believe that NVIDIA holds an oligopolistic position in advanced autonomous driving and have not given due attention to the trend of self-developed ASICs by companies like Nio, XPeng, Li Auto, and BYD. However, we believe that NVIDIA's monopoly essence to date lies in the fact that its general-purpose chips can fully meet the needs of different algorithm frameworks during the rapid iteration of autonomous driving algorithms.

Since 2024, as large model frameworks converge, ASICs have rapidly risen in AI edge applications such as autonomous driving due to their higher AI computing power density, lower costs, and energy consumption (for example, Broadcom expects the overall scale of ASICs to grow at a compound annual growth rate of over 70% in the next three years). We believe that under the continuous growth of computing power demand and strong cost constraints, more leading automakers will attempt to follow the technological paths of Tesla and Huawei, gradually replacing NVIDIA's Orin/Thor (typical representatives of general-purpose GPUs) with self-developed ASICs to pursue better solutions for autonomous driving chips. Due to their high binding with algorithms, third-party ASICs are difficult to fully adapt to the self-developed algorithm needs of automakers. Therefore, we believe that compared to "cutting corners," self-developed ASICs that are "tailored to fit" will be the necessary path for leading automakers.

Market Divergence: Is it feasible for automakers to develop their own chips? Is it necessary? How to value it?

We believe that the strategic and investment significance of automakers developing their own chips should be fully recognized. In this report, we specifically argue the following about the self-developed chips:

1) Feasibility: The market often compares self-developed autonomous driving chips to self-developed mobile phone chips, thus holding doubts about feasibility. We believe that compared to the latter, the feasibility of self-developing autonomous driving chips will significantly increase.

① Compared to mobile phone chips, the requirements for power consumption (0.1% vs. 50%), area (3-5 times that of mobile phone chips), iteration time (3-4 years vs. 1 year), and integration level are greatly relaxed for autonomous driving chips;

② Self-developing autonomous driving chips essentially means self-developing ASICs. Currently, ASIC design and manufacturing services have become highly marketized, and the improvement of domestic chip design levels has also created conditions for automakers to self-develop;

③ Referring to Horizon Robotics and Black Sesame Technologies' total R&D investment of 5.4 billion and 2.7 billion yuan from 2021 to 2023, we estimate that the average annual investment for automakers developing their own chips over a three-year R&D cycle is about 1 billion yuan. Compared to the hundreds of billions of yuan that automakers spend on R&D each year, the investment scale for self-developed chips is relatively controllable;

④ Autonomous driving ASICs do not strictly rely on process advancements (7nm is better, 14nm is also acceptable). Currently, domestic self-developed autonomous driving chips already have the conditions for full-chain autonomy and controllability.

2) Necessity: Self-developed chips are not only an alternative solution but also a shift in the development path of leading automakers towards integrated hardware and software, while autonomous driving autonomy and controllability are already urgent.

① Reviewing the mobile phone industry, integrated hardware and software is a core element for leading brands to maintain their advantage. In autonomous driving, Tesla and Huawei have set a precedent. Self-developed chips can achieve higher computing efficiency and iteration speed, building a competitive moat;

② Chips are the most fundamental AI capability, with high scalability for robots, flying cars, etc., and are "know-how" that all automakers aspiring to make a mark in AI must master; ③ Chips are the core of intelligent driving systems and currently represent the segment with the lowest domestic production rate in intelligent driving. In the context of increasing U.S. sanctions on Chinese AI, self-developed chips will be the only solution for the long-term development of intelligent driving.

3) Economic Viability: Our calculations show that self-developed chips with a shipment volume in the million-level range throughout their lifecycle will be economically viable.

Risk Factors:

U.S. sanctions on the Chinese semiconductor industry exceed expectations, the progress of automakers in self-developing chips is below expectations, the performance of self-developed chips is not as expected, the penetration rate of high-level intelligent driving is not improving as expected, and there are risks of declining sales in the automotive industry.

Investment Strategy:

High-level intelligent driving is about to 迎来 "iPhone moment" in 2025. We believe that chips are currently the only segment in intelligent driving hardware that has a logic of simultaneous increase in both volume and price. From an industrial trend perspective, with the continuous push for cost reduction, it is expected that specially designed ASICs will erode the market share of general-purpose GPUs represented by NVIDIA Orin/Thor. Self-developed chips from leading automakers will become an important variable in the industry. We recommend two types of investment opportunities:

Complete Vehicles: Self-developed chips can be seen as converting procurement amounts from third parties into sales revenue for the self-developed chip department. In the context of escalating expectations of the U.S.-China AI game, we believe that self-developed chips are not just about economic viability but are a necessary option for the long-term evolution of domestic intelligent driving. The U.S. Department of Commerce's 3A090.a rule restricts Thor, which may be the highest performance overseas intelligent driving chip available domestically in the future. Automakers constrained by NVIDIA's supply chain are likely to face risks of technological blockade in intelligent driving, making self-development the only choice for ensuring long-term development. Automakers' self-developed chips should fully enjoy the valuation premium of being autonomous and controllable.

Since algorithmic leadership is a prerequisite for self-developed chips, and the economic viability of self-developed chips is highly sensitive to shipment volume, self-developed chips are expected to remain a competition among leading automakers. Considering the long-term accumulation in the mobile phone sector, we believe that companies will also have a high probability of joining the ranks of self-developed intelligent driving chips in the future. Based on the progress of implementation, we particularly recommend leading automakers with expansion potential in fields such as robotics.

Components: 1) Leading third-party chip suppliers; companies are expected to achieve cost advantages through larger shipment volumes; 2) Domain controller suppliers: strong positioning effects in the Tier-1 supply chain closely cooperating with OEMs; 3) Wafer fabs: the core bottleneck of self-developed chips lies in advanced processes, and the strategic asset status of wafer fabs is unprecedentedly strengthened.

Author of this article: Yin Xincheng and others from CITIC Securities, Source: CITIC Securities Research, Original title: "Smart Cars | From NVIDIA to Self-Development: The Crossroads of Intelligent Driving Chips." Yin Xinchí S1010519040002

Risk Warning and Disclaimer

The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this is at one's own risk