In the heated arms race of intelligent driving in 2025, Li Auto is quietly advancing its self-developed intelligent driving chip, codenamed M100, which outperforms the industry benchmark, NVIDIA Thor-U, in some performance metrics.

In the first quarter of 2025, Li Auto’s self-developed intelligent driving chip M100 completed its sample return, entering a critical phase before mass production. Within just two weeks, the M100 successfully passed functional and performance tests, as well as stress tests conducted by the R&D team.

Currently, the chip has beeninstalled in a small batch of test vehicles, entering the road testing phase. This means that Li Auto is expected to become another Chinese automotive brand to achieve mass production of self-developed intelligent driving chips next year, attempting to break the monopoly of international giants like NVIDIA in the core hardware field of intelligent driving.

01 Background of Chip Development: Breaking Free from Dependence, Building Core Competitiveness

The global competition in smart vehicles has entered a chip-level contest. Li Auto’s Senior Vice President of Autonomous Driving R&D, Lang Xianpeng, stated that the core reason for developing self-owned chips is that as a dedicated chip, it can be specifically optimized for Li’s algorithms,offering high cost-effectiveness and efficiency. In intelligent driving systems, the chip acts as the “brain,” and its performance directly determines the effectiveness of the algorithms—while general-purpose chips can meet basic needs, they often waste computing power and experience response delays when handling specific scenarios (such as large model inference and multi-sensor fusion).

Li Auto recognizes that as intelligent driving algorithms evolve from “rule-based” to “data-driven large models,” the importance of hardware-software collaborative optimization becomes increasingly prominent. Using external general-purpose chip solutions not only fails to fully unleash the potential of algorithms but may also miss market opportunities due to the misalignment of supplier technology iteration pace with its own needs. For instance, when Li’s VLA driver large model requires higher Transformer computing power support, the architecture of general-purpose chips may not efficiently adapt, leading to a compromise in algorithm performance.

Like other new Chinese forces such as NIO and Xpeng, Li Auto has chosen the “difficult but correct” path of self-developing chips, fundamentally to break free from excessive reliance on external suppliers. In the context of increasing fluctuations in the global chip supply chain and rising geopolitical risks, mastering chip autonomy not only ensures supply security but also builds differentiated competitiveness through deep hardware-software collaboration, allowing it to take the initiative in future intelligent driving competition.

02 M100 Chip Performance: Surpassing Industry Leaders in Certain Scenarios, Targeted Optimization Shows Advantages

According to internal testing data from Li Auto, the performance of the M100 chip in core intelligent driving scenarios is impressive, especially showing significant advantages in optimization for Li’s algorithms:

• Large Language Model ProcessingWhen running inference tasks related to the VLA driver large model, the effective computing power provided by one M100 is equivalent to that of two NVIDIA Thor-U chips. This means that the M100 can handle natural language interactions and complex scenario inferences (such as predicting pedestrian intentions and understanding the behavior of traffic participants) more efficiently, with a response speed improvement of about 30%.

• Visual Task ProcessingWhen processing traditional visual tasks based on convolutional neural networks (such as lane line recognition and obstacle detection), the effective computing power of one M100 can match that of three NVIDIA Thor-U chips. This is due to the customized design of hardware acceleration units for Li’s visual algorithm architecture, which reduces data transfer losses.

This performance is not merely a result of “parameter stacking” but stems from the M100’s “scenario-based design” approach. The chip architecture is optimized around Li Auto’s sensor configuration (such as LiDAR + visual fusion solutions) and algorithm logic, eliminating unnecessary redundant computing power found in general-purpose chips and concentrating more resources on core intelligent driving tasks. For example, the M100 includes a dedicated “multi-sensor data fusion acceleration unit” that can quickly process the fusion calculations of LiDAR point clouds, camera images, and millimeter-wave radar signals, achieving 40% higher efficiency than general-purpose chips.

From a hardware parameter perspective, the M100 chip adopts a “domain controller-level” architecture similar to Tesla’s Hardware 5.0, with approximately40 billion transistors, manufactured using TSMC’s 7nm process, and power consumption controlled within 150W, comparable to NVIDIA Thor-U, but with a higher computing power density (computing power per watt). Currently, the M100 has completed TSMC’s tape-out and is entering the reliability testing phase before mass production, providing solid support for Li Auto’s intelligent driving upgrades next year (such as the implementation of L3 level urban road autonomous driving).

03 R&D Journey and Planning: Three Years to Forge a Sword, Mass Production in 2026

Li Auto’s self-developed chip project is not a “spur of the moment” decision but began with a strategic layout in 2022, taking about three years to complete the critical leap from project initiation to sample return, with a relatively fast R&D pace in the industry.

In the early stages of the project, Li Auto formed a multidisciplinary team consisting of chip architects, algorithm engineers, and software experts, with core members coming from companies like Intel, Huawei, and NVIDIA, possessing rich experience in chip design and automotive adaptation. The team first clarified the core goal of the chip: not to pursue “universal” computing power but to focus on “high efficiency” in intelligent driving scenarios, which guided subsequent R&D directions.

According to the plan, the M100 chip will officiallybe mass-produced and installed in vehicles in 2026, initially equipped in Li’s flagship models (such as the upgraded MEGA and the new pure electric SUV). During the transition period before mass production, Li Auto will continue to adopt a “dual supplier” strategy: all pure electric models (MEGA, i8) will be equipped with NVIDIA Thor-U, while the AD Max version of range-extended models will use Thor-U, and the AD Pro version will adopt Horizon Journey 6M, balancing the competitiveness of current products with the R&D pace of self-developed chips.

To ensure the smooth progress of the project, Li Auto has recently strengthened internal information management within the chip department, on one hand to avoid leakage of R&D details, and on the other hand to minimize the impact on relationships with partners like NVIDIA and Horizon. This “low-key advancement” strategy reflects both the strategic importance of self-developed chips and Li’s cautious attitude towards supply chain management.

04 Technical Architecture Features: Hardware-Software Collaborative Design, Vertical Integration Releases Potential

The core competitiveness of the Li M100 chip stems from its “hardware-software collaborative design” technical architecture, which sharply contrasts with the traditional chip model of “hardware first, software later.”

During the R&D process of the M100, the chip hardware architecture, compiler, operating system, and Li’s Halo operating system were designed and optimized in sync from the beginning. This vertical integration model brings multiple advantages: first, the compiler can fully understand hardware characteristics, mapping algorithm code more efficiently to hardware computing units, reducing instruction execution losses; second, the operating system can dynamically allocate hardware resources based on real-time task requirements, avoiding computing power waste. For example, when the vehicle is in a highway scenario, the system will automatically allocate more computing power to LiDAR point cloud processing; when entering urban roads, resources will be tilted towards visual recognition and large model inference.

This R&D strategy is led by Li Auto’s CTO, Xie Yan, who has previously worked at Intel, Alibaba, and Huawei, possessing a strong background in compilers and operating systems. His core idea is to “maximize the potential of hardware computing power through powerful software scheduling capabilities.” In Xie’s view, the chip’s “effective computing power” (the computing power that can actually be used by algorithms) is more important than “theoretical computing power,” and software is the key to connecting hardware and algorithms.

The architecture design of the M100 also reflects this concept: it adopts a “heterogeneous computing architecture,” integrating multiple dedicated modules such as NPU (Neural Processing Unit), CV accelerator (Computer Vision Accelerator), and DLA (Deep Learning Accelerator), each module can be precisely matched with Li’s specific algorithm modules through software scheduling. This design allows the M100’s computing power utilization rate to reach 85%, far exceeding the average level of around 60% for general-purpose chips.

05 Dual-Track Parallel Strategy: Balancing Present and Future, Ensuring Smooth Transition

Li Auto has adopted a pragmatic “dual-track parallel” model in its chip strategy, neither blindly relying on external suppliers nor hastily pushing immature self-developed chips to market, but achieving a smooth transition through scientific planning:

Vehicle Type	Chip Solution	Application Status	Core Consideration
Pure Electric Models	NVIDIA Thor-U	All MEGA and i8 models equipped, i6 expected to be fully equipped	Pure electric models are positioned as high-end, requiring leading intelligent driving performance; NVIDIA chips have high maturity and can meet current needs
Range-Extended Models	NVIDIA Thor-U + Horizon Journey 6M	AD Max version uses Thor-U (high-end configuration), AD Pro version uses Journey 6M (mainstream configuration)	Balancing different user needs, controlling costs through differentiated solutions while maintaining product competitiveness

The core of this strategy is “not to rush.” Lang Xianpeng explained: “Currently, we are still using the Thor series chips mainly because NVIDIA has good support for new operators, sufficient computing power, and the overall VLA iteration still has potential for change.” Before the self-developed chip matures, relying on external mature solutions ensures product competitiveness while using actual vehicle data to feed back into the optimization of the M100, ensuring that it can perfectly adapt to the latest algorithm versions upon mass production.

The dual-track strategy also reduces supply chain risks. When the M100’s production capacity is insufficient in the early stages of mass production, chips from NVIDIA and Horizon can serve as supplements, avoiding impacts on vehicle delivery due to chip supply issues. This “multi-legged approach” reflects Li Auto’s ability to balance technological innovation and commercial implementation.

06 Team and Investment: Heavy Investment Layout, Talent Ladder Supports Long-Term R&D

Self-developed chips are a typical “high investment, long cycle” field, and Li Auto has invested substantial resources in this regard. According to insiders, the total budget for the project planning is as high asseveral billion dollars, covering all aspects from chip design, tape-out, testing, mass production to team building. In 2025 alone, Li Auto’s investment in chip R&D exceeded 2 billion yuan, accounting for about 17% of its total R&D expenditure for the year.

In terms of the R&D team, Li Auto’s intelligent driving chip R&D involves multiple aspects such as NPU architecture design, SoC integration, compiler development, and automotive adaptation, making it a multi-layered system engineering project. The current team size has exceeded 500 people, with core members being senior experts from companies like ARM, Qualcomm, and Huawei HiSilicon, possessing rich experience in chip architecture, low-power design, and automotive reliability.

This year, despite some personnel changes in the intelligent driving team, the overall organizational structure remains stable. Li Auto has built a complete talent ladder through a combination of internal training and external recruitment, selecting a group of young technical leaders with international perspectives. For example, the hardware architecture leader of the M100 previously led the design of some modules of Huawei’s Ascend chip, while the software leader has years of software development experience for NVIDIA’s autonomous driving chips. This “cross-border” background helps the team integrate the advantages of different technical systems.

Continuous financial investment and talent reserves provide long-term support for Li Auto’s chip self-development, enabling it not only to complete the mass production of the M100 but also to iterate more advanced chip versions in the future, keeping pace with the evolution of intelligent driving technology.

07 Industry Trends: Automakers are Investing in Self-Developed Chips, Core Technology Autonomy Becomes Consensus

Li Auto’s self-developed chip strategy is not an isolated case but a reflection of Chinese automotive brands competing for core technology autonomy in the intelligent era. Currently, more and more Chinese automakers realize that chips are the “soul” of smart vehicles, and mastering chip technology is essential to take the initiative in competition:

• NIO has initially achieved its self-developed chip strategic goals, with its global first 5-nanometer intelligent driving chip “Shenji 1.0” already installed in the ET9 model, achieving a computing power of 1000 TOPS, focusing on LiDAR and visual fusion calculations.

• Xpeng’s self-developed Turing chip has been fully installed in the G7 and new P7 models released this year, with a single chip computing power of up to 750 TOPS.

• BYD is also making strides in power semiconductors, with its self-developed silicon carbide chips achieving mass production, with a voltage of up to 1500V, applied to its high-end models, reducing dependence on overseas suppliers like Infineon.

This trend reflects a change in the competitive logic of the smart automotive industry—from “mechanical performance competition” to “software-defined vehicles,” where chips serve as the carrier of software capabilities. As intelligent driving enters L3 and above levels, the demand for computing power grows exponentially, and the coupling of algorithms and hardware becomes increasingly tight, making the “universality” of general-purpose chips a disadvantage. Automakers developing their own chips can achieve “maximum computing efficiency” through vertical integration of “algorithms – software – hardware,” which is the core barrier in future intelligent driving competition.

08 Challenges and Risks: Multiple Tests in Technology, Mass Production, and Ecosystem

Despite the promising prospects of the M100 chip, Li Auto’s self-development path still faces multiple challenges, and any misstep in any link could affect the project’s success:

• Technical RisksChip design is a “battle of millimeters”; even if sample testing passes, reliability testing before mass production (such as stability under high temperature, low temperature, and vibration environments) may still reveal issues. For example, the cache coherence protocol of the M100 may experience data synchronization delays in extreme scenarios, requiring repeated debugging and optimization.

• Mass Production ChallengesThe transition from samples to large-scale mass production involves multiple links such as wafer manufacturing, packaging testing, and yield control. Any problem in any link could lead to delays in mass production or soaring costs. TSMC’s 7nm process capacity is tight, and Li needs to compete with other chip design companies for capacity to ensure the M100’s mass production schedule.

• Ecological ConstructionThe value of chips lies not only in the hardware itself but also relies on a complete software toolchain (such as compilers and debugging tools) and developer ecosystem. Li needs to provide algorithm engineers with an easy-to-use development environment; otherwise, even if the chip performs well, it will be difficult to fully utilize its potential. Compared to NVIDIA’s mature CUDA ecosystem, the software ecosystem for the M100 still requires time to build.

• Market CompetitionIndustry leaders like NVIDIA are not standing still; their next-generation chips (such as the upgraded version of Thor-U) may be launched when the M100 is mass-produced, further enhancing performance. Li needs to maintain its R&D pace to avoid falling behind right after mass production.

Moreover, the return on investment for self-developed chips has a long cycle, posing a significant test for the company’s financial strength and strategic determination. If the mass production progress of the M100 does not meet expectations or its performance does not meet market expectations, it could not only affect Li’s competitive edge in intelligent driving but also drag down the company’s financial performance due to the massive investment.

09 Strategic Significance: From “Technology Follower” to “Standard Setter,” Reshaping the Competitive Landscape

The strategic significance of Li Auto’s self-developed chip goes far beyond “cost reduction” or “supply assurance”; it is more about its positioning in the global smart automotive industry:

• Supply Chain SecurityIn the context of increasing uncertainty in the global chip supply chain, the mass production of the M100 will reduce Li’s dependence on NVIDIA, avoiding impacts on vehicle delivery due to external supply interruptions. Especially in the context of a complex international situation, chip autonomy can provide the company with a “strategic buffer.”

• Cost OptimizationAlthough the initial investment is substantial, once mass production matures, the unit cost of self-developed chips is expected to be 20-30% lower than that of purchasing external chips. Based on Li Auto’s projected sales of 1 million vehicles in 2026, this alone could save hundreds of millions of yuan annually, directly enhancing gross margins.

• Building Technical BarriersThrough hardware-software collaborative optimization, the M100 can fully unleash the potential of Li’s VLA driver large model, leading to superior performance in intelligent driving functions (such as urban NOA and automatic avoidance of complex scenarios) compared to competitors, forming a differentiated advantage that is difficult to replicate through merely purchasing external chips.

• Enhancing VoiceAfter successfully developing its own chips, Li Auto will transition from being a “chip user” to a “technology definer,” capable of leading the direction of chip technology according to its own needs, and may even open some technologies to industry partners, forming a technology ecosystem centered around itself, enhancing its voice in the global industry.

For the Chinese automotive industry, the self-developed chip practices of companies like Li Auto help promote the upgrading of the entire industry chain. From chip design tools, wafer manufacturing to packaging testing, Chinese companies’ capabilities in these areas will gradually improve with the increasing demand for self-developed chips from automakers, reducing dependence on overseas supply chains and achieving a transformation from “import dependence to self-control.”

10 Future Outlook: Synergy Between Chip Self-Development and Global Strategy, Supporting Long-Term Growth

Li Auto’s path of self-developing chips is closely integrated with its medium- to long-term globalization strategy, forming a mutually supportive pattern:

According to Li’s plan, its development will be divided into three stages:

• 2020-2024Focus on the domestic market and range-extended products, completing technological accumulation and brand building;

• 2025-2027Market expansion from domestic to overseas, covering both range-extended and pure electric products; during this stage, the mass production of the M100 chip will provide core technical support for overseas models;

• After 2027Focus on L4 autonomous driving and new forms of intelligent products, with chip self-development capabilities becoming key to achieving this goal.

2025 has been designated as Li Auto’s “overseas year,” with the company establishing R&D centers in Germany and the United States, working on building overseas sales and after-sales service systems. Entering mature markets like Europe and the United States, local adaptation of intelligent driving functions (such as autonomous driving logic compliant with local regulations) is crucial, and the hardware-software collaborative advantages of the M100 chip will allow Li to optimize algorithms based on local needs more quickly than competitors relying on external chips.

For example, when Li enters the European market, it will need to optimize intelligent driving algorithms for scenarios like narrow streets and complex roundabouts; the customized architecture of the M100 can quickly adapt to these new requirements, while general-purpose chips may require longer debugging cycles due to architectural limitations. Additionally, self-developed chips can reduce patent risks and supply chain costs in overseas markets, enhancing product competitiveness.

Li Auto plans to upgrade the AD Max version of its range-extended models to the VLA intelligent driving assistance system by September 2025, with performance improvements comparable to the breakthrough from ChatGPT 3.5 to 4.0. This upgrade will further validate the advanced nature of Li’s algorithms, while the mass production of the M100 chip will provide the hardware foundation for fully unleashing algorithm performance.

When the M100 officially goes into vehicles in 2026, Li Auto will become one of the few automakers that simultaneously master intelligent driving algorithms, software systems, and chip hardware. This path is full of challenges, but once successful, it will lay a solid foundation for Li’s competition in the “deep water zone” of intelligent vehicles and promote the Chinese automotive industry to leap from “following” to “leading” in core technology fields.

If you are interested in new energy vehicles, please follow the 36 Strategies WeChat public account.

Disclaimer: The images are for communication and learning purposes only, copyright belongs to the original author. If there is any infringement, please contact the author for deletion. Thank you for your support!