
XPeng challenges the computing power ceiling in the automotive industry

Arriving at the future first?
Author | Wang Xiaojun
Editor | Zhou Zhiyu
This year, XPeng Motors Chairman He Xiaopeng can be considered a busy person in the automotive circle.
With new car models and facelifts constantly being released, press conferences are happening one after another; sales have reached new heights, and XPeng's stock price has also seen significant growth compared to the beginning of the year.
On the evening of June 11, He Xiaopeng passionately introduced the computing power and other strengths of the XPeng G7 at its launch event; the next day, he hurried to Hong Kong for the first auto expo, continuing high-intensity operations for XPeng's sales in Hong Kong and overseas markets.
Compared to last year's low pressure and atmosphere of skepticism, XPeng, which has entered a positive cycle this year, is clearly more spirited.
XPeng's product strength is being recognized by more consumers. As the most important flagship model this year, XPeng has set a pre-sale price of 235,800 yuan for the G7. Within just 46 minutes at this price, 10,000 pre-orders flooded into the system.
For this new car, after experiencing the drawbacks of overly complex SKUs in the past, the G7 is divided into two versions based on different intelligent hardware: the XPeng G7 Ultra version equipped with three Turing AI chips and the XPeng G7 Max equipped with two Orin-X chips, both versions standardly feature a long range of 702 km.
As the first car equipped with AI Turing chips, the highlight of the XPeng G7 this time is its computing power.
On the hardware level, the XPeng G7 Ultra version is the first to be equipped with three Turing AI chips, equivalent to nine mainstream intelligent driving chips, with effective computing power exceeding 2200 TOPS, which is 3-28 times that of other flagship models in the industry, capable of simultaneously driving intelligent driving large models and cockpit large models, with effective AI computing power in the intelligent cockpit being 26 times that of industry flagships, meaning users will enjoy a smoother intelligent driving experience and instant cockpit response.
In He Xiaopeng's view, computing power is the primary standard for measuring AI cars. High computing power not only significantly enhances intelligent assisted driving capabilities but also greatly improves safety limits.
He also revealed that XPeng is working on many interesting features; when the G7 is first released, it may not have all the functions, but new features will be added every month through OTA iterations. These updates require the support of computing power.
It can be said that this computing power is directly aimed at the future, which is also why XPeng refers to the G7 as "a car from the future."
In comparison, the current mainstream effective computing power in the industry ranges from 80 to 700 TOPS, while the XPeng G7 is the world's first AI car to reach L3-level computing power, launching the industry's first fully local VLA+VLM model, defining the upper limit of AI cars with high computing power + physical world large models + big data.
Currently, L2+ intelligent assisted driving has become standard for high-end models, while L3 is still in the transition period from technological research to commercial validation. Therefore, at the press conference, He Xiaopeng also emphasized that although the XPeng G7's computing power has reached L3 level, the XPeng G7 is not an L3-level car. To become an L3-level car, in addition to L3-level computing power and AI software, L3-level hardware redundancy and legal certification are also required In terms of intelligence, XPeng has also introduced AR-HUD for the first time.
This time, XPeng collaborated with Huawei to create the Light-Chasing Panoramic AR-HUD. He Xiaopeng stated during the demonstration that the role of AR-HUD becomes particularly evident in rainy or snowy weather when lane markings are covered or blurred. He introduced that future upgrades will also consider unpaved road scenarios. For example, when users drive to Xinjiang or Qinghai for self-driving trips, AR-HUD can also draw lane markings on unpaved roads.
Since it is positioned in the 200,000 yuan SUV market, it inevitably enters the territory of the Tesla Model Y. XPeng faces certain challenges in wanting to carve out a share.
At the end of last year, there were over six SUVs in the automotive circle attempting to encircle the Tesla Model Y, but all failed. This year, XPeng also encounters the Xiaomi YU7, which is another formidable competitor that cannot be ignored.
Now, He Xiaopeng, who has already won some battles, appears quite confident. He introduced that, at present, the two have extremely different core technologies, "I believe this is the biggest difference; we have raised the 'ceiling' of all autonomous driving assistance companies."
In terms of user perception, He Xiaopeng also believes that the XPeng G7 has advantages in space and other aspects. From actual parameters, the dimensions of the XPeng G7 are 4892/1925/1655mm, with a wheelbase of 2890mm, which is slightly larger than the Tesla Model Y and comparable to the Xiaomi YU7.
The competition XPeng faces is also intense. In addition to the Xiaomi YU7, the LeDao L90, and the ZEEKR 7X, which could all become competitors, internally, there are the G9 above and the G6 below. If users are not particularly interested in intelligence, they may compare these models, which is something XPeng needs to balance.
However, the most difficult days for XPeng are already behind. In the first five months of this year, XPeng Motors delivered over 160,000 new vehicles, a year-on-year increase of nearly 300%. Moreover, He Xiaopeng is very confident about profitability in the fourth quarter; in an interview after the press conference, he revealed that XPeng's second-quarter financial report was quite good.
At the first Hong Kong Auto Show, He Xiaopeng also stated that XPeng aims to be among the first car companies to bring advanced intelligent driving assistance to Hong Kong.
To this end, XPeng is doing three things: first, preparing relevant vehicles, hoping that all vehicles entering Hong Kong next year will have advanced intelligent driving assistance capabilities; second, waiting for policy developments; if the Hong Kong government allows it, XPeng will be among the first to enter; third, adapting the software.
At the same time, XPeng is also heading from Hong Kong to more overseas markets. Morgan Stanley's latest report on the global electric vehicle market for April shows that in April 2025, global sales of pure electric vehicles increased by 38% year-on-year to 1.0797 million units, with XPeng ranking seventh among global electric vehicle manufacturers in April.
The elimination round in the Chinese automotive market is still ongoing, with each company striving to play its cards right. From seeking survival to actively initiating challenges, XPeng is attempting to gain greater initiative.
The following is a dialogue between XPeng Motors Chairman and CEO He Xiaopeng, XPeng Motors P/G series product head Nick, and Wall Street Journal, edited without affecting the original meaning: Question: XPeng announced that it will invest at least 4.5 billion yuan in AI this year. How does it plan to balance short-term profitability with the construction of long-term technological barriers?
He Xiaopeng: The competition in the automotive industry over the next five years will be extremely fierce. I believe that to become a successful company, one must scale up and also have sufficient research intensity, which definitely relies on profits. This year, our R&D expenditure of around 4.5 billion yuan is all focused on AI, with nearly 10 billion yuan in total R&D expenses.
Our goal is to achieve profitability in the fourth quarter, even with such a large R&D expenditure. I am very confident about this, and our financial situation in the second quarter is very good.
Question: XPeng has also committed to paying suppliers within a 60-day period. What impact will this have on XPeng? Secondly, is it true payment and not a commercial acceptance bill?
He Xiaopeng: All car manufacturers are currently thinking about how to develop sustainably and how to make the competition in the industry more orderly. I believe this is not only a policy expectation but also a user expectation, as well as the expectation of innovative companies like ours that hope to change the world and lives through technology.
In the first half of last year, XPeng's payment terms were relatively long, but they were significantly shortened in the second half of last year and further reduced in the first quarter of this year. I believe this is a clear benefit for the entire industry and certainly a benefit for users. Everyone needs to not only do it but also do it quickly, and XPeng will definitely be one of them.
Question: What advantages does the XPeng G7 have compared to the Model Y? By the end of this year, what is our goal for standing among the top new energy vehicles at this price point?
He Xiaopeng: First, we have extremely different hardcore technology. I believe this is our biggest difference; we have raised the "ceiling" of all autonomous driving companies' computing power. Similarly, the future upper and lower limits will be greatly expanded.
Second, this car offers the largest space in pure electric vehicles and the best comfort among SUVs in this category. I believe the first row, second row, and trunk are all very well done, so it is a car for families, a car for technology, and a car that makes oneself and family members very happy.
There is no specific goal to be the first or second in the industry. Of course, past models like MONA, P7+, and X9 have performed well, but I don't have many demands internally. Because XPeng hopes to form a systemic capability, I think it is meaningless to occasionally excel in a single point.
Question: You mentioned that your effective computing power is 3-28 times that of your competitors. Can you elaborate on what "effective computing power" means? Why can XPeng's self-developed chips achieve higher effective computing power than general-purpose chips? I've heard that some companies encounter issues when using self-developed chips to run their algorithms, such as slow inference speed, low accuracy, and severe heat generation. Has XPeng encountered such problems, and how did you solve them?
He Xiaopeng: A chip itself has computing power, but it depends on what kind of computing power you want to use. For example, if a competitor's chip has 250 Tops of computing power, its composition might be achieved by using A+B multiplied by 2. If you use its B part, you basically don't use the A part, or the A part has little value, so the truly valuable computing power is limited XPeng's chips are fully self-developed, so the computing power we provide is all useful.
Secondly, through compilation optimization, there is actually still strong potential that can be further improved. So if we can turn "1 chip equals 3 chips" into "1 chip equals 4 chips" in a year and a half, we are still exploring the possibility of enhancing this computing power. From a trend perspective, once we truly reach large-scale production, we will gradually discover the capabilities related to memory usage, reliability, heat generation, etc., which will have tremendous value for us.
I believe computing power is fundamental. The only thing I can say is that we are working on many interesting features. When G7 starts to roll out, there will be some features, but not many. As we continue our OTA iterations, new features will be added every month.
This year, I expect to launch extremely significant features, and there will be several major ones next year. Because over the past year, we have been developing for more than a year, and what is the main foundation? You need to run a large model on a new computing power and run it well; that is the foundation. Then, you also need to optimize for performance, stability, and reliability. You need to input a lot of data for training, and you need to create basic capabilities. You need to combine different foundational capabilities to create a valuable scenario. The first two or three have already been completed, and we are working on turning them into interesting scenarios with higher capabilities; this is what we are currently doing.
The reason we are equipping such large computing power and memory is that we want to make significant progress in the coming years; we are indeed thinking far ahead.
Q: This time, you adopted an 80.6 kWh battery as standard across the entire lineup, but you declared two versions, one of which is 68.5 kWh. Why is that? Is it to give users a higher sense of value, or is there another reason? At this stage, how can users perceive the overall capability brought by the two Turing chips, as well as the capability of a single cabin's VLM, and encourage users to choose more Turing configurations?
Nick: In fact, we also have our own insights. From last year to this year, the penetration rate of new energy has exceeded 50%, which means that the entire new energy market is no longer just competing within itself; it has entered a broader competition with fuel vehicles.
We have observed that since entering this broader competitive landscape, traditional and previously conservative users have started to enter this market. For these users, the first consideration when thinking about new energy vehicles is range, especially in the 200,000 yuan segment. If they are first-time car buyers or are replacing their vehicles, they have deep concerns about the range of new energy vehicles, especially pure electric vehicles.
Although XPeng has done exceptionally well in terms of range, to capture more market users, we ultimately decided to offer two battery options but chose to only launch the 700 km range. We believe this is the core purchasing point for users in the 200,000 yuan market.
In the entire 200,000 yuan market, the 200,000 to 250,000 yuan segment basically has pure electric SUVs with ranges of 600 km and 700 km. However, we believe that since XPeng has a solid range, along with our 5C battery and the full-domain 800V architecture, we can provide users with ultra-long range, high charging efficiency, and low energy consumption, allowing more users to enjoy these benefits and experiences So based on future judgments, competition, and users, we ultimately chose a range of 700 kilometers.
Q: The appearance of the G7 is indeed a bit different from that of the P7. What is the reason for this difference? Will the appearance change in the future?
Nick: Both models are under my responsibility, including the P7+. My personal thought is that we believe the P7+ and G7 are more oriented towards family users. For family users, we consider space as the primary factor, so whether it's the P7+ or G7, we have made many additions in terms of space, and we have fully equipped the seats and comfort configurations, considering that everyone in the family is in the C position. This is how we designed these two family-oriented models.
The all-new P7 will be released soon. We believe that the P7, which entered the market in 2020, is a leader in the entire new energy market. We hope the all-new P7 can also lead the entire market. It not only leads in functionality, speed, and performance but also in styling and exploration of the future. This is how we think about the two categories.
Q: The G7 has two versions based on chips. Besides the difference in computing power, what are the differentiating points in terms of intelligence performance between these two versions?
Nick: From the experience perspective, the Ultra version is definitely different from the Max version, and there are also differences according to our performance requirements. We hope the Ultra version can take over every 100 kilometers. However, for the Max version, we have a performance indicator of taking over every 50 kilometers. So there are differences in performance indicators.
Q: When the MONA Max version was released at the end of May, XPeng said that 500 TOPS of computing power is actually the threshold for L2-level assisted driving. Today, with the release of the G7, it is said to be 2000 TOPS, which is a threshold for L3 level. How is this so-called threshold standard evaluated internally at XPeng? Also, after XPeng released this so-called threshold concept, has there been any communication with other major car companies regarding their views on the L2 or L3 computing power thresholds?
He Xiaopeng: In fact, in the industry, the high-level computing power for L2 is basically 500T, and now there is also 700T. To some extent, 700T and 500T are the same; there is actually no difference. I believe that a multiple increase is needed for the probability model to achieve a multiple increase.
So I believe that 2000T, everyone doesn't need to focus on a specific threshold. For example, at the next point in time, we have already seen some competitors releasing 2000T, even 4000T of computing power. I believe this computing power will be L3, and some companies even consider it an important starting point for L4.
In fact, from another perspective, why not release more? Because your computing power cannot achieve it. The previous questions are very important. For example, if you stack 50 chips of 100T to reach 5000T, can it run a large model? It cannot. It is also related to chip technology and memory bandwidth; it is a composite issue. So as of today, this is an optimal solution, but it is not necessarily the lowest solution Question: Can you help identify the fundamental differences between XPeng G7 and G9, so that consumers can make better choices?
Nick: Since I am responsible for these three models, I will briefly introduce the relationship among our SUV product line.
First, the G6 is a model aimed at small and medium-sized families, focusing on practicality and technology.
The G7 is targeted at middle-class, young users and young families. It can meet individual needs as well as the needs of two people, and it is very comfortable for three people together. This is the G7. It is a very strong product in terms of family use, space, and technology. My personal summary is that if you love family and technology, the G7 is a very good model to buy.
The G9 is our high-end model, equipped with air suspension, and it has higher requirements for performance, attributes, and experience. We believe that middle-class users who pursue higher quality, including dynamic and premium users, can consider the G9.
Question: Will there be a dual-motor four-wheel drive version of XPeng G7 in the future? Also, will there be some terrain options for the driving mode, which would give the G7 some basic off-road capability? Additionally, regarding AR-HUD, our current product demonstrations are all focused on road assistance; will there be additional support for unpaved roads in the future?
Nick: We currently do not have plans to develop a four-wheel drive version because we believe the existing options are sufficient.
Secondly, the AR-HUD is currently in its first round of deployment, so we will continue to update its functions via OTA based on user usage and some new scenarios. I believe XPeng's software capabilities are very strong, and seeing the current effects of our AR-HUD, we think the future functions will become increasingly powerful. We will address different road conditions and driving situations.
He Xiaopeng: If we have the energy next year, assuming we drive the Ultra version to Xinjiang, Qinghai, and Tibet, for example, in areas with cliffs and deserts, how should our car find the right path on its own and "draw" a road on the AR-HUD, while being able to drive very steadily, even more steadily than others? This is definitely a combination, and I believe it can be achieved; I just don't know in which month it will happen