He Xiaopeng responds: Is it reasonable that the market value gap with Tesla is 50 times? Did I "harm" Lei Jun by persuading him to enter the automotive industry?

After trending on the hot search, He Xiaopeng responded:

“Advising Lei Jun to enter the car - making industry was really not to ‘harm’ him.”

He also showed off his “muscles” in terms of technological trends and accumulations:

“In China, no one else has truly implemented VLA except us.”

Regarding the feedback from the secondary market and public opinion, He Xiaopeng also directly expressed his dissatisfaction:

“In another six months, when Robotaxi is launched, the outside world's value judgment of us will be completely different.”

After the new XPeng P7 became a hit and sold 10,000 units in just 7 minutes, Auto Intelligence Reference had a face - to - face chat with He Xiaopeng immediately.

How to define the product, how to compete in AI, how to explain the so - called “backstabbing” of car owners, etc... All these hot, sharp and valuable topics are freshly presented.

How were the configuration and pricing of the new P7 considered?

Question: Why is Ultra standard across all versions of the new P7, but not on the G7?

He Xiaopeng: Each model has its own positioning considerations. The P7 is crucial to XPeng, like a totem. It must be very simple, pure and cutting - edge.

Question: How to solve the production capacity problem of the P7?

He Xiaopeng: Later - entering car - making companies, including us, will have stronger production capacity because modularization is getting better and better. I'm quite confident that the goal of the P7 is to be among the top three in its class of sedans, and the production capacity is prepared accordingly.

Question: Why is less emphasis placed on AI and more on performance and appearance for the new P7?

He Xiaopeng: The new P7 is the model I've driven the most among all the cars I've test - driven. Once you start driving it, you don't want to stop. A large part of the discussion focuses on its driving control, chassis handling and acceleration because it really gives you that adrenaline - pumping feeling.

In terms of intelligence, actually XPeng is crossing this era and entering the next one. Many insiders have asked me why we don't use Thor or dual Orin. My answer is to wait until next year. They are only half - generation upgrades, neither good enough for high - end nor suitable for low - end, and it's a struggle.

Question: Around 200,000 units of the first - generation P7 were sold. How about the amortization of mold and R & D costs and the return on investment? With the shortening product life cycle and more frequent model updates of later models, will there be greater sales pressure from a cost perspective?

He Xiaopeng: I didn't understand it when I entered the industry back then. I got into trouble for many years after rushing in. Everyone told me that selling 100,000 units a year would make a profit, but later I found it was impossible.

First, we come from the Internet industry and are used to amortizing costs at once. Second, the cost of a gasoline car sold for 200,000 yuan is only tens of thousands of yuan, but it's completely different for new energy vehicles. Just the three - electric system accounts for 40% - 50% of the cost. That's why companies only focusing on the three - electric system are doing relatively well, while car companies are struggling.

On the other hand, although we've been losing money in the past few years, we've been accumulating R & D capabilities, manufacturing capabilities and quality - improvement capabilities. I believe that car companies will achieve great success and make a lot of money after five years.

One of XPeng's goals in the fourth quarter is to make a profit, from a tiny profit to a small profit, and then to a medium - sized profit... Maybe we can recoup all the previous losses in one or two years. We definitely have a chance. The core is that we're different from traditional car companies. We have self - developed intelligent hardware and software stacks and are also working on cross - domain integration.

For example, the intelligence we see today is single - entity intelligence. In 2026, we'll enter the era of group intelligence. You'll see that the world is changing faster and cooler. What we're doing now is part of the 2027 and 2028 plans.

Question: The starting price of the P7 is lower than that of the G7. How to explain this to the car owners who feel “backstabbed”?

He Xiaopeng: When converting a gasoline - powered car on the same platform into an SUV, the price increase is not high. But the logic is different for electric vehicles. A larger SUV means greater wind resistance and mass, especially wind resistance, which puts pressure on the range. To ensure the range, a larger battery is needed.

Question: What kind of sales volume does the new P7 need to achieve to play the role of a totem and strategic product?

He Xiaopeng: We never publicly promise sales targets for a single model, but we have internal targets. I hope it can be among the top three in the pure - electric sedan market, and the price is set according to this goal.

The most difficult part is to keep the sales volume stable or even increasing three months after the new car is launched. This is very important, so we're not being aggressive internally.

Question: What's XPeng's methodology for creating a hit product?

He Xiaopeng: First, it needs to have good looks. Second, it should have excellent technological attributes in its niche market. For example, the MONA 03 has almost unrivaled intelligent assisted driving capabilities in its price range.

Actually, we've summarized every launched model, but to be honest, I haven't found a universal methodology. There's no one - size - fits - all approach, and every model has its pros and cons. If anyone has such a methodology, please teach me.

“You can't do a good job in VLA without investing billions every year”

Question: Li Auto revealed that it invests 6 billion yuan in AI every year. How about XPeng? How much money is needed to do a good job in VLA?

He Xiaopeng: I'm not very familiar with the investment situation of our competitors, so I can't make a judgment. But from my observation, only we in the whole industry have truly implemented VLA.

The concept of VLA is easy to explain, but to implement it, an investment of billions of yuan is necessary. XPeng's investment this year is about 5 billion yuan. If you only invest a few hundred million yuan a year, you may be able to create a mini - VLA, and it's probably based on an end - to - end logic.

End - to - end is mainly a data logic based on two - dimensional and structured information to build a model, while VLA is a multi - dimensional data type and unstructured logic, directly building a model for output actions. From this perspective, I haven't seen anyone in China truly implement VLA. Most are grafted VLAs, but XPeng is different.

This is why I emphasize that if it's not 10 times better than the current end - to - end experience, it can't be called VLA.

Question: VLM will be pushed two months later than VLA. How do their functions affect each other?

He Xiaopeng: Actually, I think today's VLA is strictly speaking “VA”, with the “L” being very weak. Due to computing power limitations, although it can recognize some traffic signs and text, its logical reasoning ability has declined significantly.

VLA is a fast model, while VLM is a slow model. The VLM we're working on is essentially more like a “VLT” that can derive tasks. For example, if you ask it to get you a bottle of water, this instruction is actually a collection of tasks, and the VLT can reason out the solution steps.

Ultimately, VLA needs to be combined with VLM. I think VLA is more like a smart cerebellum, fast and smart, while VLM is like the brain, a bit slower but more comprehensive.

From the end of this year to next year, people will gradually see that the two - model solution will become a consensus, just like the human brain.

Some people ask why we don't combine them into one model. I think it's not scientific and difficult to control energy consumption if we combine them, or it's not “biological”, and it's impossible to coordinate the subsequent planning of robots.

Question: With the evolution of VLA technology, the capabilities of various companies seem to have been improved. Will there be homogenization? How does XPeng maintain its unique advantage?

He Xiaopeng: XPeng started working on VLA in October last year, but we haven't announced it.

To be honest, I don't believe that AI can be done well with little investment. There may be some short - term tricks, but in the long run, it still requires basic and solid investment.

So I believe that next year, there will be a differentiation in capabilities, and a new first - tier camp will emerge. The current first - tier camp may all fall to the second - tier camp.

Question: How is XPeng's reserve of super - computing GPUs? Is there any layout for self - developed products?

He Xiaopeng: I think we're progressing quite fast in terms of GPUs. We're using both NVIDIA GPUs and domestic GPUs now, with more than 30,000 in total. And our utilization rate is very high. In the past year or so, there have been huge changes in the inference computing power of the cloud and local ends, as well as the cloud and local models.

In the AI competition, the most important thing is for us to grow rapidly and make more profits because the future investment in AI will definitely be huge.

Question: How does XPeng ensure user privacy and security?

He Xiaopeng: VLM can actually better protect everyone's privacy, and basically, the data doesn't need to be uploaded to the Internet.

VLM must be well - developed so that even if the operator's network fails or the company goes bankrupt, the car can still be used normally.

“The market value gap between the ‘Chinese Tesla’ and Tesla is 50 times? He Xiaopeng responds like this”

Question: XPeng is the company most similar to Tesla globally in terms of products. Whatever Tesla has, XPeng also has. But the market value gap between the two on the capital market is 50 times. Do you feel aggrieved? Is XPeng undervalued?

He Xiaopeng: I have two thoughts. First, the valuations of Chinese technology companies are generally about 7 times lower than their American counterparts, which happens to be similar to the exchange rate. This is the case for Internet companies, AI companies, etc.

Second, XPeng's full potential hasn't been released yet. For example, next year, we'll launch our own Robotaxi. I believe that in another six months, people will have a different perception of XPeng's value.

Question: Technology companies can have profit margins of 20%, 30% or even more than 40%, while the average profit margin of Chinese car companies is only 4% - 5%. How can they support the investment of an AI company?

He Xiaopeng: Before the emergence of digital technology companies, the technology companies in the