Guosheng Securities: Maintain "Buy" rating on XPENG-W with a target price of HKD 126

According to Zhitong Finance APP, Guosheng Securities released a research report stating that it is optimistic about XPeng-W (09868) due to its strong product cycle, overseas expansion, advancements in intelligent driving, and emerging business opportunities such as robots/Robotaxi. The firm expects the company's sales from 2025 to 2027 to be approximately 450,000/680,000/1,010,000 units, with total revenue reaching 80.6 billion/133.3 billion/188.5 billion yuan, and a non-GAAP net profit margin of -1.5%/3.4%/5.2%. The target price is set at HKD 126 and USD 32.4. Additionally, the company's other growth curves are expected to gradually be incorporated into the valuation system as subsequent product developments progress, maintaining a "buy" rating. XPeng will announce "significant breakthroughs" in the field of physical AI at this year's AI Technology Day, with its world base model achieving "key progress" in its ability to simulate the world.

Guosheng Securities' main points are as follows:

The XPeng AI team has been engaged in the research and development of the physical world AI base model for over a year.

As early as April this year, XPeng held an AI technology sharing session in Hong Kong, revealing that it is developing a super-large-scale autonomous driving model with 72 billion parameters, known as the "XPeng World Base Model," which has approximately 35 times the parameter count of mainstream VLA models. 1) In terms of model training speed, XPeng has independently developed the underlying data infrastructure (Data Infra), increasing the scale of data uploads by 22 times and improving the data bandwidth during training by 15 times; by jointly optimizing GPU/CPU and network I/O, the model training speed has ultimately increased by 5 times. 2) In terms of the training data volume and scale of the model, the parameter count will reach 72 billion, and the training data volume will increase to 200 million Clips. Dr. Liu Xianming, head of the world base model, stated: "Over the past year, we have conducted a large number of experiments and observed significant scale law effects on models with 1 billion, 3 billion, 7 billion, and 72 billion parameters: the larger the parameter scale, the stronger the model's capability. For the same model size, the larger the training data volume, the stronger the model's capability." Reports in October indicated that the base model being developed by XPeng uses the largest model data volume ever recorded, claiming to be "the most advanced" physical AI large model in the country. Recently, Dr. Liu Xianming of XPeng was promoted to head of the autonomous driving center, highlighting the importance of the world base model within XPeng.

The physical AI large model + Turing chip constitute XPeng's core competitiveness in the fields of smart vehicles, Robotaxi, and robotics.

1. On the software side, the models in XPeng's mobility and robotics sectors share the same origin and will deploy the base model to vehicles through cloud distillation of small models, equipping "AI vehicles" with a new brain, allowing XPeng to quickly deploy Turing AI intelligent driving in various countries, while also empowering XPeng's AI robots, flying cars, and more. 2) On the hardware side, XPeng's self-developed Turing chip achieves higher effective computing power at a lower cost through customized design, with the computing power of a single Turing chip equivalent to three Orin-X chips. The current new P7 and G7 Ultra versions are already equipped with three Turing chips, and the L4 model in 2026 will be equipped with 3000 TOPS of computing power. XPeng's vehicle-end models, supported by higher computing power, are expected to bring far superior assisted driving experiences, potentially achieving a leading edge among domestic automakers In terms of robotics, the firm believes that the company possesses full-stack self-research capabilities, an excellent brain, better hardware-software integration, and cost advantages. The firm looks forward to the next generation of robots on Technology Day.

Similarly, global new force leader Tesla is also actively promoting the research and development of high-performance computing chips, showcasing the "World Simulator," and demonstrating the versatility of cars and robots.

1. In terms of smart driving chips, Tesla stated that relying on the fourth-generation hardware, it will elevate AI parameters by an order of magnitude in the future, making cars "like living organisms." The next-generation hardware AI5, leveraging Tesla's deep understanding of software and hardware, is expected to enhance computing power by another tenfold at the same cost. Additionally, according to South Korean media reports, Tesla's next-generation core chip "AI6" is also in planning, which will adopt Samsung's second-generation 2-nanometer (SF2P) process. 2) In terms of software algorithms, on October 26, U.S. time, Tesla officially disclosed a neural network system called the "World Simulator," which is a "twin world" entirely composed of neural networks. It can generate continuous, multi-perspective virtual driving scenarios with extremely high fidelity based on massive data from the real world. Tesla claims that through this method, its AI system can learn the equivalent of 500 years of human driving experience in just one day. This underlying AI engine and simulation platform are versatile, with the "World Simulator" used for training cars also being utilized to train its "Optimus" humanoid robot. The firm believes that XPeng and Tesla have a high degree of similarity in technical routes and business layout ideas, and looks forward to breakthroughs in physical AI and achieving a leading position