Produced by | Bullet Finance

Art Design | Qianqian

Reviewed by | Songwen

On June 15, Li Auto's Livis Day event focused on software and embodied AI came to a close. Li Xiang introduced the concept of an embodied AI vehicle as a four-in-one entity: 'electric vehicle + professional driver + AI computer + lifestyle assistant.' The event showcased technical details of the self-developed Mach M100 data flow chip and the Mach VLA large model, while also announcing a clear OTA (Over-the-Air) update roadmap spanning the entire year.

These achievements represent the culmination of Li Auto's five-year technological marathon. Over the past decade, the company achieved RMB100 billion in revenue through precise product definition. Now, it is building a long-term competitive edge through full-stack self-developed core technologies that are difficult to replicate.

Currently, the entire automotive industry is hyping AI, large models, and advanced intelligent driving concepts. However, most automakers' intelligence efforts remain superficial, relying on external solutions and hardware assembly without true core technological foundations.

For Li Auto, which has surpassed RMB100 billion in annual revenue and maintained three consecutive years of stable profitability, Livis Day signifies a decisive strategic pivot: a formal shift from defining family vehicle products to creating foundational technologies for embodied AI.

Products can only deliver short-term sales benefits, while core technologies determine an automaker's survival ceiling in the next decade. This is the sober answer Li Auto provides to the industry after five years of deep R&D investment.

1. An Eight-Year Full-Stack Layout: A Methodical Technological Marathon

Many observers interpret Li Auto's newly released embodied AI technologies as mere trend-following amid the current AI wave.

However, a review of the complete R&D timeline reveals that Li Auto's systematic layout (strategic layout translates better as 'strategic layout ' in technical contexts, but keeping 'layout' for consistency) of chips, operating systems, large models, and perception systems began as early as 2018—a clear, disciplined, and long-term investment.

In October 2018, the Li ONE was officially launched, establishing a foundational perception hardware system for vehicles. Simultaneously deployed microphones, radar, and in-car cameras enabled preliminary audiovisual perception and human-vehicle interaction capabilities, marking the first step toward full-vehicle intelligence.

By 2021, competition in the new energy industry focused on range, space, and extended-range/pure electric routes. Most automakers concentrated R&D resources on interior, powertrain, and configuration optimizations, while intelligence solutions generally relied on turnkey packages from external suppliers like Qualcomm, NVIDIA, and Horizon Robotics.

That same year, Li Auto made two critical moves: In May, its self-developed in-car perception system debuted with the 2021 Li ONE, achieving autonomous control over environmental perception hardware and algorithms. In October, the self-developed Xinghuan OS (vehicle intelligence operating system) was officially initiated, pre-emptively laying the groundwork for subsequent hardware-software integration and embodied AI infrastructure. At the time, the industry widely misunderstood the value of self-developed operating systems, favoring off-the-shelf solutions for their time and cost efficiency, despite higher R&D investment and longer return cycles.

However, Li Auto's vision extended beyond operating systems—it aimed to secure autonomous control over chip compatibility. The 2021 global chip shortage foreshadowed extended lead times and price surges for automotive chips, causing frequent supply disruptions and prolonged adaptation cycles. Based on the same logic of core technology autonomy, Li Auto initiated preliminary R&D for self-developed automotive AI chips that year, paving the way for the formal chip project launch in 2022.

Thus, the core value of the self-developed Xinghuan OS was not merely optimizing cabin interaction but firmly grasping chip compatibility autonomy to avoid supply chain bottlenecks.

By 2022, the industry became entrenched in an intelligence arms race, competing on urban NOA (Navigate on Autopilot) and highway pilot coverage mileage while still relying on externally sourced core computing chips.

Li Auto chose a more challenging path.

In June, the Li L9 was launched, featuring the first mass-produced self-developed Central Domain Controller XCU. Simultaneously, it transcended traditional vehicle manufacturing by transforming cars into long-term mobile living spaces through multi-scenario hardware design.

In November, the self-developed automotive AI chip project was formally approved. Unlike general-purpose CPU/GPU architectures, this chip was natively designed for massive real-time automotive data stream computing, tailored specifically for autonomous driving and in-car large model AI calculations.

That same year, full-line control chassis R&D commenced, establishing hardware foundations for intelligent agent execution and forming a technical closed loop with self-developed chips for computing decision-making and physical execution.

In 2023, as global large model technologies surged, the industry widely adopted third-party general-purpose large models for cabin voice assistants, achieving simple voice interaction upgrades.

Li Auto again avoided shortcut solutions, initiating R&D for its in-car foundational large model while pursuing a dual-track intelligence evolution strategy combining on-device local inference and cloud-based continuous training.

In March, the self-developed in-car foundational large model project was approved, confirming the dual-parallel path of on-device local inference and cloud-based iteration. By December, the self-developed MindGPT large model was deployed via OTA 5.0, completing in-car verification.

In 2024, Li Auto's self-developed technologies entered mass production: The Li L6 launched in April with the first large-scale deployment of Xinghuan OS. In October, China's first mass-produced end-to-end AD (Autonomous Driving) large model was deployed, upgrading intelligent driving perception and decision-making capabilities.

By 2025, self-developed technologies continued to create value and iterate: In March, Xinghuan OS was open-sourced, sharing self-developed foundational capabilities. In August, the VLA Driver Large Model debuted with the Li i8's mass production, elevating vehicle intelligent decision-making and human-like driving comprehension.

By Livis Day 2026, the Mach M100 chip, complete Mach VLA large model, and next-generation full-domain spatial interaction system achieved full-stack mass production deployment.

From foundational perception in 2018 to the phased deployment of perception, domain controllers, in-car OS, AI chips, and in-car large models over eight years, Li Auto's embodied AI technologies represent not short-term trend-following but a methodical, full-link autonomous long-term technical layout (strategic layout ).

This eight-year layout was supported by organizational restructuring, RMB10+ billion R&D investment, and sustained management strategic focus. In 2025 alone, Li Auto's R&D spending reached RMB11.3 billion, with half allocated to AI, chip, and operating system technologies.

While most automakers still integrate external supply chain solutions, Li Auto has completed the leap from 'assembled intelligence' to 'native intelligence.'

This explains Li Xiang's confidence in presenting a monthly-detailed iteration roadmap at the event: 30% improved intelligent driving efficiency in July with high-frequency features like Tour Guide Agent, Inter-Vehicle Walkie-Talkie, and Low-Power Sentry Mode; September brought complete reverse and complex road intelligent driving capabilities, smart home connectivity, and Super CarPlay with lossless Apple Music; December achieved proactive hazard avoidance, traffic police recognition, and seamless boarding with 0.2-second intelligent driving response.

2. From Product Definition to Technological Definition: Li Auto's Evolving Moat

Reviewing Li Auto's decade-long development, its first leap forward relied on superior product definition.

From the debut Li ONE's instant success to the L-series full-size family SUV lineup dominating the premium new energy market, Li Auto precisely addressed core pain points of Chinese middle-class families: spacious interiors, comfortable home-like cabins, range-anxiety-free extended-range powertrains, and family-oriented configurations.

While the industry pursued sportiness, performance, and minimalist tech interiors, Li Auto carved a niche with its 'mobile home' positioning, implementing family-scene features like refrigerators, large screens, rear comfort seats, child-friendly spaces, and in-car entertainment—creating a unique scenario-based comfort experience that disrupted the market.

This product strategy translated into tangible market results: Li Auto became China's first new energy vehicle maker to produce one million units, achieving RMB112.3 billion in 2025 revenue, three consecutive years of profitability, RMB100+ billion cash reserves, and a leading position in the family luxury segment.

However, Li Auto's management recognized the inherent limitations of product-defined moats.

Product differentiation barriers are relatively low: Competitors can quickly replicate seating, refrigerators, large screens, spacious interiors, and extended-range powertrains within 1–2 years by adjusting R&D priorities.

Recent market shifts confirm this: Multiple automakers now offer family-focused six/seven-seater models with in-car refrigerators, rear entertainment screens, and spacious cabins, gradually eroding Li Auto's former scenario-based advantages.

As all automakers compete on space, comfort, and configurations, product-based competitive edges dilute, leading to homogeneous price competition.

Recognizing these product moat limitations, Li Auto chose embodied AI to elevate its competitive edge. If Li Auto's first growth phase was 'defining a better family vehicle,' its second phase is 'redefining the intelligent vehicle itself.'

This time, Li Auto broke free from traditional automaker thinking, focusing not on 'adding configurations or optimizing cabin experiences' but on Refactoring (restructuring) human-vehicle relationships: Traditional intelligent vehicles passively receive commands, while embodied AI vehicles actively perceive environments, make autonomous decisions, and continuously evolve alongside users.

This new definition rests on Li Auto's eight-year-honed full-stack self-developed technology foundation.

The first foundational layer is the Mach M100 dynamic data flow chip—the 'heart' of the intelligent agent. As the world's first mass-produced dynamic data flow automotive AI chip, it uses 5nm automotive-grade processes, delivers 1,280 TOPS effective computing power, and achieves over82% computing utilization—significantly outperforming mainstream general-purpose intelligent driving chips.

The second collaboration (synergistic) layer is Xinghuan OS—the intelligent agent's 'nervous system.' The self-developed OS integrates vehicle sensors, chips, chassis, and cabin hardware, resolving the traditional fragmentation between cabin and intelligent driving systems to enable full-domain data interoperability across interior spaces, road perception, vehicle control, and lifestyle services.

Through Xinghuan OS, the Mach M100's computing power dynamically allocates across intelligent driving perception, voice interaction, spatial vision, and in-car services, avoiding computing waste and ensuring smooth system operation.

The third intelligent core is the Mach VLA large model—the agent's 'brain.' The model integrates 3D visual perception, road driving logic, and multimodal language interaction, achieving a 0.28-second comprehensive response time—approximately 40% faster than the average 0.45-second human driver reaction and nearing the 0.25-second physiological limit of top F1 drivers.

Deep integration and native compatibility among chips, operating systems, and embodied models form Li Auto's unique technological closed loop.

While product moats are easily imitated, full-stack technological moats face time, capital, and talent barriers that are difficult to overcome quickly. Livis Day sends a clear signal: Li Auto is no longer just building 'better family vehicles' but securing foundational technological authority to define next-generation intelligent vehicle standards.

3. The Ultimate Vision for Embodied AI

By redefining itself as an embodied AI company rather than a mere 'intelligent vehicle company,' what does Li Auto envision?

It means escaping the single dimension of vehicle sales growth.

From a technological reusability perspective, the dynamic data flow chip architecture, embodied AI model framework, and multimodal spatial interaction system are inherently adaptable to all intelligent terminals in the physical world.

Thus, vehicles represent only the first embodiment of Li Auto's embodied AI, with broader technological layout (strategic layout ) for the entire physical AI era.

The previously announced Livis smart glasses enable spatial interaction; humanoid robot R&D projects leverage unified chip, model, and system foundations for motion control and environmental perception development. This technological system could extend to household service robots and commercial mobile intelligent terminals.

Second, the year-round iterative OTA growth plan transforms embodied AI capabilities from a one-time event at product launches into a continuously appreciating long-term user asset.

By defining vehicle delivery as the 'birth moment' of embodied agents, each OTA update becomes an opportunity for users to perceive the agent's autonomous learning and capability evolution.

For users, vehicles no longer rapidly depreciate with age but continuously gain new functions and optimized intelligence through OTAs. For Li Auto, real-world road and family scenario data generated by massive users continuously train the VLA large model, forming a positive data closed loop.

From a decade-long perspective, Livis Day represents a clear demarcation point on the path of ideal development. Before this demarcation point, Li Auto was widely recognized as a top-tier product in the automotive industry. After this point, Li Auto officially transformed its identity, entering the broader arenas of physical AI and embodied intelligence technology as a technological long-distance runner. The competitive boundaries of the industry are no longer confined to the new energy vehicle market but extend to the entire era of physical artificial intelligence.

Many people still tend to evaluate the competitiveness of automotive companies based on short-term indicators such as monthly delivery volumes, model pricing, and end-of-line discounts. However, Livis Day clearly conveys a core message: in the next decade, Li Auto's key to success will extend far beyond just vehicles.