The automotive industry can no longer contain NIO, Li Auto, and XPeng.

If you cover the backdrop of the press conference, many people would think it's a developer conference of an AI company.

On June 15th, at the Li Auto Livis Day, Li Xiang held a two - hour press conference.

He talked about chips, AI, Agent, world models, and embodied intelligence. He hardly talked about cars. This might be the press conference in Li Auto's history that least resembles a car press conference.

Interestingly, Li Auto released its self - developed Mach M100, a 5nm automotive - grade chip with a single - core computing power of 1280 TOPS, claiming to be the world's first dynamic data - flow AI chip. The new Li L9 Livis version is equipped with two of them, with a total computing power of 2560 TOPS.

With the launch of the Mach M100, NIO's Shenji, XPeng's Turing, and Li Auto's Mach have all submitted their answers on self - developed chips.

The outside world generally interprets this move as a passive technological pursuit to cope with the supply - chain crisis. But looking beyond the surface, what really deserves attention is that these three companies are declaring the same thing to the world: The automotive industry can no longer contain the ambitions of NIO, XPeng, and Li Auto.

1. What has Li Xiang been talking about in the past three years?

If you line up Li Xiang's public statements in the past three years, the trajectory is very clear.

In 2023, he talked about the Li L series, family users, features like refrigerators, color TVs, and big sofas, range - extenders, and sales. At that time, Li Auto was a typical automotive company, with product definition and family scenarios as its selling points.

Things started to change in 2024. He frequently mentioned AI, AGI, robots, and operating systems. The focus of the topic shifted from how to sell more cars to how large models would reshape the physical world.

In 2025, the context completely leaned towards Silicon Valley. Terms like VLA, end - to - end, and world models, which were originally exclusive to OpenAI, became part of the daily conversations of automotive company founders on their social media.

Today, in 2026.

Li Xiang presented a complete definition of embodied intelligent vehicles at the press conference, proposing the "Four - in - One": an electric vehicle, a professional driver, an AI computer, and a life assistant. To achieve this definition, Li Auto forcibly increased its AI R & D investment to 50% of the annual budget. The R & D organizational structure also underwent a fundamental restructuring. The base model team was vertically split into three secondary departments: embodied engineering, embodied interaction, and embodied behavior. Autonomous driving was downgraded to one of the parallel departments.

In three years, Li Xiang completely switched the company's underlying narrative from car - making to AI - making. The Mach M100 chip is not the starting point but the first step in the implementation of this process.

If you look at these three years as a whole, what he did follows the same logic as Apple's transformation from a computer company to a company focusing on devices, software, and services. First, he completed the cognitive switch internally, and then announced to the outside world with a landmark product: The old era is over.

2. The most expensive part of a car has changed three times

In the past decade, the most expensive and most competitive parts of a car were the engine and the gearbox. Whoever could handle complex mechanical processing and manage the supply chain of tens of thousands of precision parts had the pricing power.

In the past five years, the myth of fuel - powered cars has shattered, and the most expensive asset has become the lithium - ion battery. The production capacity and schedule of CATL directly determined the profit structure and delivery situation of almost all new - energy automotive companies.

Today, the third change is taking place. The most valuable part of a whole vehicle has completely become computing power.

Li Bin, the founder of NIO, once publicly calculated that in a high - end intelligent electric vehicle, the cost of batteries and chips has already exceeded 50%. Relying on overseas general solutions like NVIDIA for a long time, automotive companies have to pay more than 20,000 yuan for chip procurement from Silicon Valley for each high - end model equipped with four Orin chips.

But cost is just the surface. A deeper change is that when the entire vehicle is driven by AI from perception to decision - making to execution, the chip is no longer just a component procurement issue but a strategic power issue.

Let's look at the actual test data provided by Li Auto's Mach VLA system today. The comprehensive reaction speed is 0.28 seconds. In contrast, the average physiological reaction time of an ordinary human driver is 0.45 seconds.

This means that at a speed of 120 kilometers per hour, the AI driven by the self - developed chip can stop the vehicle 6 meters earlier than a human driver. Li Auto plans to compress this indicator to less than 0.2 seconds in the OTA at the end of the year.

When a car's reaction speed starts to exceed that of a human, it is no longer just a car. It is a robot in a car's shell.

What supports this robot is not the steering wheel and brake pads, but the chip.

3. Chips are not about cost but about power

Many people think that automotive companies are making chips to reduce costs. In the short term, it's the opposite. NIO's Shenji team was established in 2021 and it took four years to achieve mass production. Li Auto's chip team has also invested for several years. The huge R & D expenses, the tape - out costs of millions of dollars, and the long automotive - grade verification cycle are all burning money.

The real reason is that if future cars are the carriers of embodied intelligence, the chip is the brain of this intelligent entity. Handing the brain to a third party is like giving away your Achilles' heel.

NIO suffered the most painful setback. In 2021, due to the global chip shortage, the supply of ESP modules was cut off, resulting in a 42% reduction in the production of the ET7 in a critical quarter, causing a revenue loss of about 1.9 billion yuan. Li Bin's statement that the supply - chain lifeline cannot be in the hands of others is a lesson learned with real money.

Another industrial bottleneck that is rarely publicly discussed is the waste of the architecture of general automotive chips.

NVIDIA's Orin and Thor are general architectures, essentially general computing platforms. They are designed very broadly to be compatible with the strange algorithms of all automotive companies.

However, what models NIO, XPeng, and Li Auto run are highly determined internally. Li Auto runs the Mach VLA model, XPeng runs a 30B - parameter large - scale end - to - end model, and NIO runs the NWM world model. Inferring the transistor layout of the chip from their own core algorithms is far more efficient than using general chips.

The actual computing power utilization rate of the Mach M100 is as high as 82%, while traditional general automotive chips usually only have a utilization rate of 30% - 40% when running large models. Under the same computing power indicator, the actual output of self - developed chips can reach three to four times that of general chips.

The difference is more obvious in edge - side inference. When running a large - language model on the Mach M100, its Prefill speed reaches 2.7 times that of NVIDIA's desktop super - computing solution. This means that an intelligent driving chip installed in a car with extremely low power consumption can directly outperform a professional AI workstation worth tens of thousands of yuan in specific tasks.

When automotive chips start to compete with desktop super - computers, the concept of automotive chips itself is outdated. It is just an AI chip that happens to be installed in a car.

4. Three chips, three declarations

NIO, XPeng, and Li Auto are all making chips, but the equity structure and business paths of the chips reveal their different definitions of their future identities.

NIO's strategy is extreme heavy - asset investment and independent commercialization.

The Shenji NX9031 was mass - produced in September 2024, and more than 550,000 chips have been shipped this year, covering models from the flagship ET9 to the LeDao L90. In 2025, the chip business was spun off into an independent Shenji company, with a first - round financing of 2.257 billion yuan and a valuation of nearly 10 billion yuan. NIO achieved its first single - quarter profit in Q4 2025, with an adjusted operating profit of 1.25 billion yuan. The 10,000 - yuan hard - cost savings per car brought by the self - developed chip became the most important financial driver.

Li Bin's ambition is straightforward. In the future, Shenji will not only be used internally but also sold to the entire industry, completely transforming from a cost center to a profit center.

XPeng, on the other hand, chooses to use chips for the deepest global ecological binding.

The Turing chip is not spun off but is packaged as a core technological asset and reverse - exported to Volkswagen in Germany. In March this year, the Volkswagen ID. UNYX and XPeng G08, equipped with dual Turing chips and a total computing power of 1500 TOPS, were officially mass - produced. For the first time in history, a self - developed chip of a Chinese automotive company has redefined the technological foundation of a traditional European automotive giant.

He Xiaopeng's goal is clear. The Turing chip aims to achieve an annual shipment of 1 million units, aiming for the top position in the domestic large - computing - power edge - side AI chip market. His ultimate goal is to build a "physical AI" company. In the future, the Turing chip will take over cars, humanoid robots, and flying cars.

Li Auto is the outlier that submitted its answer last.

The Mach M100 was designed by inferring the chip from the algorithm. After running the VLA model for several years to figure out where the bottlenecks are and how the data flow works, the blueprint of the chip was drawn. Therefore, the M100 uses a data - flow architecture, triggering calculations wherever the data flows, without the need to move data back and forth like in traditional architectures. Li Auto does not make general chips but makes exclusive chips that understand its own models best, defining embodied intelligence through full - stack self - development of chips, models, operating systems, and cars.

The three paths are different, but they point to the same thing. Chips are no longer just parts to be bought and installed but a way for automotive companies to declare who they are.

5. The ultimate competition for the next - generation computing platform

Why did Apple develop its M - series chips?

Because Tim Cook saw that the ultimate battlefield in the future is not smartphones but the personal computing platform in the entire ecosystem. On the day the M1 was released, Apple announced to the industry not just a chip but the official departure of the Mac ecosystem from the control of the Intel era.

Why did Tesla develop the FSD chip and the Dojo super - computer? Because Elon Musk knew that the future lies not in making electric cars but in the dominance of AI training and inference for fully autonomous driving.

Today, when Li Auto releases the Mach M100, the underlying logic is the same.

If you still view this arms race from the perspective of the automotive industry, your vision is too narrow.

Apple is not just a mobile phone company, Microsoft is not just a software company, and OpenAI is not just a chat - tool company. In essence, they are all computing - platform companies that dominate the era.

Today, when NIO, XPeng, and Li Auto make chips, what they are competing for is not better automotive chips but the entry ticket to the next - generation computing platform.

Cars may be the best carrier today. They have enough space to install chips and sensors, complex scenarios to train AI, high enough selling prices to cover R & D costs, and several hours of daily usage time to generate data.

But this carrier is just a starting point.

When AI R & D accounts for 50% of the budget, when the organizational structure is reorganized for embodied intelligence, and when an automotive chip can match a desktop super - computer, these companies are answering a question with their actions:

They never wanted to be better automotive companies. They want to be the next - generation AI platform companies that control the physical world.

Cars are just the shells they chose in 2026.

Words "Beyond the Page":

In the past, Chinese companies were best at innovation in the application layer. Today, more and more companies are moving towards the underlying layer.

Models, chips, operating systems, and infrastructure.

Everyone has realized the same thing. In the AI era, the moat is not in the application layer but in the underlying layer.

Whoever controls the computing power controls the future.

So it's not a coincidence that NIO, XPeng, and Li Auto all end up making chips. It's the choice that all ambitious companies in this era will eventually make.

The greatest company in the next era may not be born in the traditional automotive industry. But when it steps out of the laboratory, it will surely hold the absolute power to define the intelligent underlying layer.

This article is from the WeChat official account "Beyond the Page", author: Ban Jun. Republished by 36Kr with permission.