Li Auto (3Q25 Minutes): Transforming Cars into Embodied Intelligence in the Physical World

The following are the $Li Auto (LI.US) Q3 2025 earnings call minutes organized by Dolphin Research. For earnings interpretation, please refer to "Li Auto: Can the i6's explosive orders save the collapsing ideal amidst the range extender's collapse?"

1. Review of Core Financial Data

Performance Overview:

Total Revenue: RMB 27.4 billion, down 36.2% year-on-year, down 9.5% quarter-on-quarter.

Automobile Sales Revenue: RMB 25.9 billion, down 37.4% year-on-year, down 10.4% quarter-on-quarter (mainly due to a decrease in delivery volume, partially offset by an increase in average selling price due to changes in product mix).

Gross Profit: RMB 450 million, down 51.6% year-on-year, down 26.3% quarter-on-quarter.

Operating Loss: RMB 1.2 billion (operating margin -4.3%), compared to a profit of RMB 3.4 billion in the same period last year; Net Loss: RMB 624.4 million.

Vehicle Marketing Expense Ratio: 15.5% (19.8% excluding Li MEGA recall costs).

Cash Position: Cash balance at the end of the quarter was RMB 98.9 billion; Free Cash Flow: -RMB 8.9 billion.

Business Outlook:

Q4 2025 delivery guidance of 100,000 to 110,000 vehicles, total revenue guidance of RMB 26.5 billion to RMB 29.2 billion.

2. Detailed Information from the Earnings Call

2.1 Management Remarks

Li Xiang:

The third quarter of 2025 is also the first quarter of Li Auto's second decade. We have faced various challenges brought by product cycles, public relations, supply chain, production ramp-up, and policy changes, which have impacted our deliveries and operations. However, today I want to talk more about our long-term thinking.

As Li Auto enters its second decade, what are the three most important key choices? They are organization, product, and technology.

The first key choice is organization:

The choice we face is: the management model of a startup or the management model of professional managers? In the past decade, Li Auto's first seven years were under the management model of a startup. As the scale expanded—reaching a revenue scale I had never experienced in my previous ventures—around 2022, many people suggested we move towards the management model of professional managers. Historically, century-old automotive companies like Mercedes-Benz and BMW, as well as tech giants like Microsoft and Apple, have achieved great success under this model.

Over the past three years, we have worked very hard to transform ourselves into a governance system of professional managers. After real experience and implementation, we realized that startups and professional managers are two completely different governance systems. It is not about processes or organizational structure, but the differences in management philosophy and elements, and each is suitable for different stages and industry environments.

The management style of professional managers can be very successful, but it requires three elements:

① The industry and technology cycles are relatively stable;

② The company's industry position is leading and stable;

③ If these conditions are met, the management model of professional managers is a good choice. Apple and Microsoft both grew from billion-dollar companies to trillion-dollar companies after professional managers took over.

The conditions for startups are exactly the opposite:

① The industry and technology cycles are undergoing significant changes;

② The industry landscape is uncertain, and the company is not yet a leader;

③ The founder and startup team are still working hard every day, full of motivation.

In today's world where artificial intelligence is reshaping various industries, the environment we are in and our own characteristics are more in line with the conditions of a startup. The core of startup management is four things:

More dialogue, deep dialogue, rather than more reporting. In a highly changing environment, deep dialogue is a key element for enhancing cognition and making bold decisions.

Focus on user value, not just delivery. Only things that are truly valuable to users are worth delivering, not just completing various delivery tasks.

Continuously improve efficiency, rather than occupying more resources. What cost RMB 10 last year should cost RMB 8 this year, so that there are more resources for long-term investments and capability building that do not generate revenue in the short term.

Identify key issues, not create information asymmetry.

Only when value is increased, efficiency is improved, and key issues are resolved, can we continuously meet user needs in a highly competitive and highly changing environment.

Over the past three years, the startup team and I have worked hard to learn the management system of professional managers, forcing ourselves to accept various changes, but we have become worse versions of ourselves. Yet, companies like NVIDIA and Tesla are still managed in a startup manner today. If the world's strongest companies can use the startup management style, why should we give up the way we are best at? Since 1998, I have 27 years of experience in startup management and have never worked in any large company as a professional manager. We are now facing a highly competitive industry and a highly changing technological environment.

I love cars, love products, love artificial intelligence, and work is my greatest hobby. Why not use my best abilities and methods to manage Li Auto? This is the most important choice for Li Auto as it enters its second decade. Therefore, starting from the fourth quarter of this year, we are firmly returning to the startup management model to face the challenges of the new era and new technologies. The choice of organization is the foundation of everything.

The second key choice is product:

Entering the new decade, the more critical question is what problems we really want to solve for users, what is the product? Where is the technology going, this is the eternal essence. What exactly is our product for the next 10 years? Is it an electric vehicle? Is it a smart terminal? Or is it an embodied robot?

If the product remains an electric vehicle, the logic of competition will become a battle of parameters: 20 kilometers more or less in range, 2 centimeters more or less in car length? If it's just competition among electric vehicles, it's about bigger space, more models, cheaper prices, and replicating already proven successful designs, like the Li L9. Any R&D investment beyond this is a waste of R&D costs, stronger sensors, larger models, stronger active suspensions are all a waste of vehicle costs, and even the energy consumption caused by powerful computing power and active suspensions can have a negative impact on range.

If we choose to make smart terminals, we will focus more on what's inside the screen, instinctively becoming a repetition of smartphone functions. Most of the industry's innovations in smart terminals over the past few years have essentially been about moving smartphone functions to the car, putting phone apps into the car system, putting large language models in the car, and even hoping to do coding in the car, do next token prediction in the car. Do users buy cars for work or for a better life? Clearly, there are better experiences and more natural applications on phones and computers, so why put them in the car? These investments have very limited value enhancement for users and are even self-entertainment for the company.

We can also turn the car into an embodied intelligent product, in other words, a robot. The "Transformers" cartoons and movies clearly tell us that the biggest classification of robots is two forms: one like a human, one like a car. "Knight Rider" and "Cars" also clearly tell us that cars are one of the core forms of robots.

How to turn a car into a robot? Give it the perception ability of eyes and ears, the model ability of a brain and nerves, the computing ability of organs like a heart, and transform the hardware to form a stronger body capability. Let the car robot have the ability of a top driver, not only can it drive, but it can also greet you every day, help you park, help you charge, open and close the car door for you, providing you with a convenient experience and care. It can play the role of a parent, assistant, or attendant, providing you with the most convenient life services and space care in the car. Just like the first-class cabin service on a plane, just like the care and attention of a mother always by our side when we were young.

How to define a good embodied robot? From a passive machine to an automatic machine, and then evolve into an active robot. The most valuable product of embodied intelligence in the next 10 years must be a car with automatic and active capabilities, and the competition of the product is also about the extent of automatic and active capabilities. These values will be integrated into high-frequency life and experiences, and once you have them, you can never go back.

The choice of electric vehicles is not bad, but it is not enough; the choice of smart terminals is not bad, but it is not enough. Only by choosing the most difficult problem of embodied intelligence can we truly change users' lives. Providing embodied intelligent robot products with automatic and active services, it happens to be the car form robot of Transformers, the car form robot of "Cars" and "Knight Rider". This is the best opportunity and the highest difficulty challenge given to automotive companies and entrepreneurs by the new era.

The third key choice is technology:

More specifically, it is the choice of complete system technology. Is it language intelligence for the digital world, or embodied intelligence for the physical world? This means building completely different system capabilities.

To do well in embodied intelligence, the most important thing is to build a system different from language intelligence. It includes:

Perception of embodied intelligence (equivalent to eyes and ears)

Model of embodied intelligence (equivalent to the brain)

Operating system of embodied intelligence (equivalent to nerves)

Computing power of embodied intelligence (equivalent to the heart)

Body of embodied intelligence (equivalent to the human body)

Currently, no third-party supplier can provide this complete system, and not even any part of it can be supplied.

The biggest feature of language intelligence is focusing on large models and computing, larger scale models, and larger scale computing can bring stronger capabilities. However, embodied intelligence must increase the perception of the physical world, and the model must be based on the understanding of the physical world, with precision as the primary task, followed by generalization. The operating system needs to ensure the optimal integration of software and hardware, providing higher frame rates, and the overall system needs to be fast and precise. The computing power provided by the perception, model, and operating system of embodied intelligence also needs to be supplied on the end side, not the cloud. Finally, the hardware needs to be transformed to become the body of embodied intelligence. For example, a three-dimensional neural control system and active hydraulic suspension system to improve the efficiency and accuracy of body execution.

From the perspective of embodied intelligence and the entire system, there is so much that can and urgently needs to be changed:

Perception: Based on existing perception models and end-side computing power, the current 3D perception, I am talking about effective perception distance—I am talking about effective, not theoretical maximum—is only over 100 meters, far less than human eyes. If upgraded to a 3D sensor similar to the working principle of human eyes, the effective distance can be expanded 2~3 times, and today more than 50% of common problems in assisted driving can be naturally solved. 3D perception is not limited to the field of autonomous driving, active interaction and service with the car owner outside the car, and active interaction and service with family members inside the car also become possible. This requires major breakthroughs in perception model research and development, and must be efficiently coordinated with chips like M100 designed for embodied intelligence and a strong compiler team to achieve.

Model: With a true understanding of the physical world, the model can better perceive and understand the physical world, human data can be more efficiently used for training, and data generated by visual models can better improve training. The best existing computing platform, a **4B (4 billion parameters)** model runs at a frame rate of only 10 Hz, while the execution system is 60 Hz. If the model's frame rate can be two to three times faster, some comfort issues and sluggish response issues in assisted driving can be effectively resolved. This also requires deep transformation and customization of traditional technical architecture and computing power, as well as a proprietary operating system. The M100 chip is developed to solve these essential problems of embodied intelligence.

Body: The fastest response speed of human braking and steering is about 450 milliseconds, while the complete chain from perception to execution in autonomous driving is about 550 milliseconds, sitting in the driver's seat, the instinctive reaction is much slower (like an elderly person driving). The wire control system can increase the response speed of the entire chain to 350 milliseconds, don't underestimate this 200-millisecond difference, it can reduce the accident rate by more than 50%, and make people feel that it drives better and safer than themselves, both rationally truly safe and emotionally safe.

Based on such needs, the entire control method will become completely different. If only focusing on the scale improvement of models like language intelligence, doubling the model scale, the performance improvement brought by increased computing power is only 5%~10%. From the perspective of embodied intelligence as a whole, solving the most critical problems in each link of the entire system, the next round of skill improvement in autonomous driving should be 5~10 times. The ability of embodied intelligence to provide users with automatic and active fast interaction and real service is a qualitative change of having and not having.

Over the past three years, our technical reserves for the complete system of embodied intelligence have made us full of confidence in the next generation of products. The first year of embodied intelligent robots officially begins with car robots, and trillion-dollar revenue is just the beginning.

The above three key choices determine the foundation of our development in the second decade, which is more difficult and challenging than the previous decade. We know that future competition is not in one city or one pool, but in the long-term choice of direction and the determination of continuous investment. Relying on a solid financial foundation, we will remain focused, using our most beloved startup management model to create leading embodied intelligent products, ensuring that Li Auto can lead technology through cycles and become a company that creates long-term value for users and society.

I also hope to communicate and interact with investors more in this way in the future, rather than in a fixed format report every quarter. Thank you to our investors for supporting and trusting us in our most difficult times, we will do our best to make Li Auto the best-performing company in the field of embodied intelligence and the company with the highest user value in the next 3~5 years.

2.2 Q&A

Q: If we only look forward to the next year (2026), what technologies and products can we expect? And based on considerations for the next 10 years, how far in the future can investors truly see a leap in the company's products and technology, and feel a real difference?

Li Xiang: We believe that in 2026, that is, next year, when our AI system centered on the M100 chip begins to be delivered on products, I think the real value and experience will change.

The specific change is still what I mentioned before, it will turn products like cars from passive use to truly automated and proactive to provide services to users. And this service and experience is not about doing a Deep Research, or doing a Coding, or generating a video, but it will be reflected in our daily life and high-frequency experiences, feeling and experiencing the fundamental changes brought by this value every day.

Secondly, regarding what clear functions or values can be presented in the next ten years. I think based on our understanding of artificial intelligence and embodied intelligence, it cannot give clear functions like traditional programming or hardware functions. When we build the systematic capabilities of artificial intelligence, as long as some of the most critical core problems are solved, or some key performances are broken through, it will bring a series of changes reflected in our specific life or product value. It may even far exceed our expectations.

I think this is the transformation that artificial intelligence and embodied intelligence can bring to the physical world, products, and our daily lives, and the driving force behind it, rather than the presentation method of traditional programming and function models.

Q: Regarding the three electric systems, the company's products are switching from range extenders to pure electric, which will inevitably face many challenges. So, around electric technology and the industry chain, what effective technical reserves and supply chain reserves have we made in the future three electric fields of pure electric vehicles?

Ma Donghui: Regarding the layout of R&D and supply in the three electric fields. In fact, Li Auto has always focused on in-depth self-research on the core technologies of the three electric systems, including breakthroughs in electric drive, battery, and electric control technologies. In the electric drive system, we mainly focus on what is valuable to users, such as energy efficiency and NVH (quietness/vibration) experience. We have self-developed and contracted the production of silicon carbide power chips, and we have self-developed and self-produced power modules and motor controllers. At the same time, Li Auto has also built its own exclusive drive motor factory, thus constructing a full-chain self-research capability from silicon carbide power chips to power modules to motors. Our electric drive technology is not only used in pure electric models but also in all range extender models, ensuring a very quiet and smooth driving process, and effectively optimizing energy consumption levels, improving vehicle range.

Regarding battery technology, we focus on charging speed and safety. We have built a full-stack self-research capability around 5C ultra-fast charging batteries, including the chemical system of battery cells, battery management system, BMS control models and algorithms, and the layout and structural design of battery packs, which can achieve the three core advantages of ultra-fast charging, long range, and long life. In terms of battery supply layout, our plan is to adopt a dual mode of external purchase and self-research joint venture. Next year, we will also mass-produce Li Auto's own brand 5C batteries, further strengthening battery safety performance through the industrialization of self-research technology, and upgrading the user's energy replenishment experience.

The last point is in electric control, we create an ultimate driving texture through full self-research of software and hardware. On the software side, we can achieve full-stack self-research on power control, energy management, and engine calibration. On the hardware side, some core domain controllers, PCBA, and underlying software are all deeply self-researched. Through collaboration with self-researched chassis technology, we can achieve precise optimization of vehicle drivability and smoothness, achieving easy-to-drive and easy-to-handle driving characteristics. Overall, through the self-research of three electric technologies, we hope to bring users a certain experience of fast charging, long range, easy driving, and complete safety with technical hard power.

Q: Regarding the latest orders and delivery status of the i-series models Li i6 and i8, how and when can the current supply chain bottleneck be resolved, and what is the outlook for normalized sales of these two models in the coming months?

A: The i8 and i6 models have laid a solid foundation for the long-term stable growth of the BEV business and have successfully entered core BEV markets such as Beijing, Shanghai, Jiangsu, and Zhejiang, with orders significantly increasing since September. To address production ramp-up challenges, the i6 will officially launch a dual battery supplier strategy in November to ensure consistent performance and quality, with i6 monthly production capacity expected to steadily increase to about 20,000 units from early next year. We sincerely apologize to customers for delivery delays caused by key component supply chain planning, and the team is working hard to accelerate production and delivery.

Q: Regarding the operating cash outflow of RMB 7.4 billion and free cash outflow of RMB 8.9 billion this quarter, which led to a significant decline in cash reserves, what are the reasons and what is the outlook for cash flow in the coming quarters?

A: First, as we guided in the last earnings release, we faced tremendous pressure on deliveries in the third quarter, leading to an overall decline in delivery volume. This caused a difference in revenue and ultimately affected operating cash flow.

Secondly, the impact of shortening the payment cycle to suppliers. Due to the adjustment of the payment cycle caused by the National Day holiday. We attach great importance to our cooperative relationship with supply chain partners and actively respond to their needs.

Currently, the payment period for all our accounts payable is 60 days. Payments are made by bank transfer through bank accounts, without using any physical notes or certificates like bank acceptance bills.

Q: The replacement subsidy policy will change next year, and the purchase tax for electric vehicles will increase from 0% to 50%. If subsidies are phased out, what is the sales strategy for 2026?

A: This marks the industry's shift from policy-driven to market-driven, highlighting the value of stronger players. The policy exit will have short-term fluctuations, and we expect a customer concentration lock-in effect at the end of 2025, leading to a significant decline in delivery volume in the first quarter of 2026. In the long term, we are optimistic about the penetration rate of new energy vehicles, with the penetration rate in the Chinese market expected to reach 55% to 60% in 2026, and over 60% in the high-end market. Our response strategy is to protect user interests and adapt to new standards.

During the transition period, we will provide a purchase tax difference compensation plan for L6 customers who lock in orders in 2025 but deliver in 2026. All 2026 models meet the new energy consumption standards and are eligible for the 2026 incentive policy. In the long term, we will offset the policy impact through technological advancements, such as the comprehensive adoption of the 800-volt high-voltage platform and 5C ultra-fast charging batteries to improve efficiency and reduce energy consumption. The goal is to operate about 4,800 ultra-fast charging stations in 2026, with 35% located in highway service areas. We will continue to deepen supply chain localization and leverage economies of scale to stabilize prices, while accelerating product iteration to maintain the product competitiveness of all 2026 models. The company will achieve a historic breakthrough in delivery volume in 2026 through excellent product strength and user value.

Q: What new features, specifications, and advantages can be expected from the new generation EREV Li series in 2026?
A: Next year will see a major generational upgrade, with changes based on in-depth user research and feedback, as well as years of technological accumulation, aiming to create a powerful product fundamentally different from the current generation to support the goal of regaining EREV market leadership in 2026. In terms of model configuration, we will return to a simplified SKU strategy to balance market coverage and supply chain efficiency, and even the basic version will not compromise on user experience, with all features as standard. In design upgrades, while retaining the iconic design DNA, we will enhance the sense of luxury and craftsmanship, balancing strong brand recognition and fresh user appeal to better meet the needs of family users.

In terms of technology: the entire series will be equipped with 5C ultra-fast charging technology, better coordinating with our current pure electric ultra-fast charging network to better address energy replenishment anxiety. At the same time, relying on the first-mover advantage and technological accumulation in the range extender field, further strengthening the brand recognition of Li Auto as a leader in range extenders.

The core of the major facelift of the L series in 2026 is to respond to market uncertainty with certain technological upgrades, certain delivery rhythms, and certain user value.

Q: Regarding the MEGA recall, why was it accounted for in the third quarter? How was the amount determined and the sharing ratio with the supply chain? What is the impact on the fourth quarter gross margin? And what is the latest progress on the recall and MEGA orders?
A: Let me briefly respond to the question about this cost. The costs associated with the MEGA recall have been fully accrued in the third quarter as a post-balance sheet adjustment item. This cost will be reflected in the latest audit report. I will not repeat most of the details already covered in the announcement. Currently, we have decided to produce and replace all affected battery packs to meet recall requirements. This means we have lowered the gross margin.

The recall details have been announced, and currently, all factory resources are being used to meet recall demands, thus reducing the delivery of the 2025 MEGA, with most battery packs used to replace the 2024 recalled vehicles, in line with the value proposition of customer interest and safety first.

Q: Regarding AI, please update on the latest developments, models, and user feedback of VLA, as well as future goals and upgrade plans?
A: Regarding the current progress and iteration speed. The Li VLA (Vision-Language-Action) end-to-end large model has been fully pushed to the AD Max models in September. And with the migration capability of model iteration, we quickly covered from the new L series to the old Li L9, allowing all users to experience its core capabilities synchronously, enabling old cars to enjoy new technology and new experiences.

Through user feedback and data analysis, the upgrade of intelligent driving has shown significant results. Especially for L series owners, the use of intelligent driving remains strong. User daily activity and MPI (average mileage before takeover) are both on an upward trend. Users also generally feedback that NOA is smoother in longitudinal control, more decisive in detour decisions, and significantly improved in complex intersection games.

We will continue to iterate and continuously break through functions:

OTA 8.0 was the first fully pushed version, with safety as the core optimized experience.

In early December, we will push OTA 8.1 to further enhance perception capabilities and respond more accurately.

At the end of December, we will upgrade the architecture, focusing on strengthening the interaction of language and behavior information, optimizing the decision-making process, and adapting to the 2026 self-developed chip M100 on the car.

We will also have more innovative features landing one after another, launching the industry's first defensive AES (Automatic Emergency Steering) function, upgrading safety protection capabilities. We are also exploring full-scene parking from any parking space to any parking space. And combined with our self-built ultra-fast charging stations to achieve intelligent pile/car finding functions, truly perfecting the intelligent travel ecosystem.

Q: Can you update on the progress and future plans of self-developed chips and self-developed operating systems, and the open-source progress?

A: The operating system is the core cornerstone of smart cars. To achieve high efficiency, the system must be designed as an integrated architecture, not fragmented components. Our self-developed controller hardware and operating system have reduced our development cycle from the industry average of 15 months to 9 months, while costs have been reduced by 20%. Since many modules in the technology stack are still provided by suppliers, to accelerate innovation together, we have open-sourced the Star Ring OS to achieve collaborative development with partners and the ecosystem.

In September, we established the Li OS Technical Steering Committee, and multiple companies in the smart car industry chain signed the community charter, including OEMs, chip manufacturers, software and hardware service providers, and component suppliers.

At the same time, we are advancing the development of a car-end base model for physical AI. Our focus is on enhancing perception, understanding, and response capabilities, enabling the model to have deeper understanding and more stable response capabilities. The AI inference chip is the computing engine of this system. The controller built on our self-developed M100 chip is currently undergoing large-scale system testing. It is expected to achieve commercial deployment next year, and this chip has been collaboratively designed with our base model, compiler, and software system.

We expect that in the next-generation autonomous driving system based on the VLA model, the cost-performance ratio of M100 will be at least three times that of current high-end chips. Based on an efficient AI inference and execution system, we have already started the continuous performance improvement and cost reduction development work of the next-generation platform and chip.

The risk disclosure and statement of this article: Dolphin Research Disclaimer and General Disclosure