Li Xiang, the AI speaker, standing before a rift?

Li Auto has always wanted to “get on board.”

Get on the AI board.

But on the journey to catch this train, Li Auto doesn’t seem to be taking it easy, always feeling like it’s “just a bit short.”

This “bit short” gives Li Auto and Li Xiang an inexplicable tone of “persuasion” and “self-justification” when discussing AI.

Li Xiang: The “Career Transformation” of a Super Product Manager

This self-justification and persuasion began at the end of 2024.

Around the year-end, Li Xiang gave an extended interview, repeatedly stating, “Li Auto is not an automobile company; it’s an artificial intelligence enterprise.”

To quickly fulfill the promise of becoming an “AI company,” Li Auto was busy throughout 2025: restructuring the organization to “de-Huawei-ize,” open-sourcing the Star Ring OS, advancing in-house chip development, and launching its first AI glasses.

But this doesn’t seem to be enough.

Fast forward to the beginning of this year, Li Xiang decided on short notice to hold an internal all-hands meeting. The meeting lasted nearly two hours. For most of the time, Li Xiang shared his views on AI trends and emphasized several important AI-related milestones.

But after the meeting, employees were left with only a confused remark: “I don’t understand what the boss is talking about.”

And such voices were not few. At that time, many internal Li Auto employees expressed views on the company’s internal social media platform, unanimously stating they “didn’t understand” or questioning the meeting’s significance.

That is to say, at least within Li Auto at that time, Li Xiang had not yet fully convinced employees that “Li Auto is an AI enterprise.”

Why?

“Li Auto’s team has always been biased towards short-term goals,” pointed out an automotive industry analyst. So when Li Xiang suddenly started talking about long-term goals, the employees’ first reaction was that it didn’t fit Li Xiang’s usual “pragmatic” persona.

Li Xiang is indeed very pragmatic.

After all, before 2024, as a product manager akin to a “top hit-making machine,” Li Xiang’s decisions always accurately hit market “pain points” and yielded returns in the shortest possible time.

That’s the pragmatism of a “product manager.”

But in an era where everyone talks about AI, when trends and concepts are laid out on the table, Li Xiang’s identity as a product manager needs to change. He needs to become a technically savvy CTO that the market believes in.

Li Xiang’s new CTO-like identity was fully displayed at the Livis Day Li Auto Software & Embodied Intelligence Launch.

June 15th, Beijing.

Standing on stage, Li Xiang set the tone for Li Auto’s next decade. He said that in the past ten years, Li Auto created a “mobile home,” and in the second decade, it aims to “give life” to the car and the home.

From “building a house” to “finishing and decorating,” besides needing recognition for Li Xiang’s own identity transformation, the tool to give life to Li Auto’s cars—AI capabilities—also needs to be acknowledged.

This launch was extremely information-dense: the in-house chip Mach M100, the large model Mach Mind, the intelligent driving system Mach VLA, the annual OTA roadmap, the promise of “safety and efficiency surpassing humans,” and an appealing upgrade path: pushing a new version in Q3, benchmarking Tesla FSD V14’s capabilities in Q4.

At a glance, comparing these capabilities among automakers with intelligent driving features, they don’t seem particularly extraordinary.

So, what was the significance of this launch?

Li Xiang’s criterion for deciding whether to hold a launch is: “If you don’t hold this launch, users will miss out on knowing one important thing, then it should be held; if holding it doesn’t provide users with more valuable information, then it shouldn’t exist.”

Measuring the Livis Day on June 16th against this standard, it seems that merely understanding these technical parameters is no longer attractive enough, because Li Auto desperately needs a “hardcore tech show” to counter the external stereotype that “Li Auto only knows how to make refrigerators, TVs, and sofas.”

So, who was this embodied intelligence launch really for? Did it achieve the effect Li Xiang wanted?

To a certain extent, it did, but not much.

The primary audience for this launch was Li Auto owners, and the secondary audience was institutional investors in the secondary market and core algorithm teams. It was also a “prevent job-hopping” and “prevent short-selling” roadshow aimed at talent and capital.

A dual first-batch owner of Li Auto L8 and i6 told Auto after watching the launch: “In terms of intelligent driving, Li Auto is in a league of its own.” He also said that his third car would still be unwaveringly a Li Auto.

Clearly, this launch convinced at least one Li Auto owner; the “not much” aspect was reflected in the stock price: Li Auto’s stock fell 1.5% at the close on June 16th and fell 3.75% at the close on June 17th.

The implication is clear.

Not “Sexy” Enough Embodied Intelligence, “New Packaging” for “Old Wine”?

But this is understandable. After all, unlike creating blockbusters, the technology monetization cycle is long. Li Auto, accustomed to short-term returns, inevitably needs to endure a period of solitude.

The question is, for the current market, is the “embodied intelligence” concept formed by Li Auto’s technologies really that sexy?

Before finding the answer to this question, we must first clarify what Li Auto’s embodied intelligence actually is.

The day before the launch, on June 15th, Li Xiang posted on Weibo: “Many people got to know Li Auto starting from ‘refrigerators, TVs, and big sofas.’ Today, please remember this picture.”

Li Xiang said: “The core difference between an embodied intelligent car and an intelligent car is: protecting human safety, completing tasks independently, and being more efficient than humans. This is our definition of an embodied intelligent car, and it’s what Li Auto aims to achieve in the next ten years.”

That is to say, unlike XPeng’s humanoid robot walking the catwalk at its launch, the embodied intelligence concept Li Auto wants to convey seems somewhat “abstract,” and the audience needs some time to accept this not-so-concrete embodied intelligence idea.

In fact, on the surface, this was a launch about technology, but in reality, it was more about “redefining the rules of the game.” Li Auto tried to use the four words “embodied intelligence” to elevate the competition to the dimension of “whose car has more life.”

Li Xiang’s definition of an embodied intelligent car is “four in one”: an electric vehicle, a professional driver, an AI computer, and a life assistant. The electric vehicle and AI computer are the “embodiment,” while the professional driver and life assistant are the “intelligence.”

The subtlety of this definition lies in: it repackages a bunch of terms the industry is talking about into a concept with a sense of ownership. “Professional driver” is autonomous driving, “AI computer” is the computing power base, and “life assistant” is in-cabin AI.

These talking points are used by Tesla, XPeng, and NIO. But Li Auto gave them a collective name called “embodied intelligent car” and then declared: This is true intelligence; the rest are mostly “function-driven” intelligence.

It’s like a restaurant boasting “Michelin three-star chef + organic farm direct supply + constant temperature wine cellar + butler service,” and then slapping on a label of “new dining species.”

It’s just that each thing is done well, but together they don’t create a new species.

1280 TOPS of Computing Power: How to Revitalize Expensive “Bricks”?

If the concept of embodied intelligence wasn’t effectively communicated, it’s just Li Auto not mastering the communication rhythm in the AI wave. But turning back to the heavyweight technologies Li Auto launched, are they impressive enough?

Take the Mach M100 chip as an example. The parameters are indeed solid.

In May this year, the Mach M100 was mass-produced and installed in vehicles, becoming the world’s first mass-produced dynamic data flow AI chip. It uses a 5nm automotive-grade process, with a single-chip computing power of 1280 TOPS and an actual operating efficiency exceeding 82%. Compared to mainstream industry solutions, NVIDIA Orin-X has a single-chip 254 TOPS, making the Mach M100’s computing power density significantly higher. The new Li Auto L9 Livis, equipped with dual Mach M100 chips, has a total system computing power of 2560 TOPS.

Li Auto announced it has achieved full-stack in-house development from chips, compilers, and operating systems to AI algorithms and domain controllers.

But a harsh fact masked by the parameters is: the core of AI is data, and Li Auto’s data scale is being left behind by leading players.

Tesla’s FSD V14 relies on millions of cars on the road, feeding back real driving data every day. Huawei’s ADS relies on a vast matrix of partner automakers.

In comparison, Li Auto’s assisted driving mileage data still has an order-of-magnitude gap with these two. Catching up within half a year presents immense time pressure.

But Li Auto also has its own solution: using “computing power” to compensate for the lack of “data.”

The true value of the Mach M100 is not the number 1280 TOPS, but that it gives Li Auto the ability to define its own data pipeline: no longer constrained by the iteration pace of third-party chips, it can customize the computing power structure according to its own data needs.

Training world models through simulated synthetic data—this is the biggest reason for the Mach M100’s existence: it’s not for handling today’s L2+, but for running tomorrow’s high-density simulations.

This is an expensive “curve overtaking” strategy, essentially admitting: the speed at which I sell cars in the physical world cannot keep up with the speed of data demand.

But this path carries extremely high risk.

XPeng’s Turing chip was already mass-produced in Q3 2025, with cumulative shipments exceeding 200,000 units, targeting nearly 1 million units shipped in 2026, and all models are switching to in-house chips in Q2. Li Auto’s chip has just been “first launched,” mass production verification hasn’t started, and the data loop hasn’t been closed.

If the simulated data path doesn’t work, the Mach M100 will be Li Auto’s most expensive “brick.”

Conclusion:

As mentioned earlier, the day after Livis Day ended, Li Auto’s stock fell for two consecutive days.

This isn’t the market not understanding technology; it’s precisely that the market understands technology too well: it knows how many yield rate climbs separate a 5nm chip from launch to mass verification; it also knows how many gaps between OTA promises and delivery separate a “four-in-one” embodied intelligence entity from definition to mass production experience.

When Li Xiang completed his identity switch from “product manager” to “technology evangelist,” did the company behind him and the users in front of him keep pace? This cognitive gap is becoming the most difficult rift for Li Auto to cross in the AI era.