Guosheng Securities: Reiterates 'Buy' Rating for XPeng Motors; VLA2.0 Large Model Demonstrates Smooth Performance in High Computing Power and Big Data

According to Zhitong Finance APP, Guosheng Securities released a research report stating that it maintains the 'Buy' rating for XPeng Motors-W (09868) with a target price of HKD 105.7 and USD 27.0 for (XPEV.US). The brokerage is optimistic about the company's strong product cycle, overseas expansion, intelligent driving advancements, and emerging business opportunities such as robotics/Robotaxi. The brokerage forecasts the company’s sales volume to be approximately 430,000/570,000/840,000 units for the years 2025-2027, with total revenue reaching CNY 75.2 billion/103.1 billion/145.5 billion and non-GAAP net profit margins at -1.2%/2.3%/3.2%. Considering the deepening cooperation with Volkswagen, the brokerage has broken down its business forecasts, estimating primary business revenue to reach CNY 100.1 billion in 2026, with profits from the Volkswagen partnership contributing approximately CNY 2.7 billion in 2026. Additionally, the company’s other growth curves, such as Robotaxi and robotics, are expected to gradually be incorporated into the valuation system following commercialization this year.

The main viewpoints of Guosheng Securities are as follows:

XPeng VLA 2.0 is a native multimodal large model for the physical world.

Unlike the digital world, the physical world presents certain characteristics that make it more challenging: 1) Input signals consist of continuous unstructured data, unlike text which is easy to tokenize; 2) Signal outputs are continuous, such as controlling a steering wheel; 3) The physical world involves certain unknown feedback and interactions. XPeng Motors’ VLA 2.0 is a native multimodal foundational model for the physical world. To address the difficulties of physical AI models, XPeng has designed a native multimodal tokenizer, a signal processing unit that allows it to encode all signals more efficiently and in a more primitive manner, achieving native fusion of multimodal information and avoiding biases introduced by single modalities.

The Chain-of-Thought (CoT) used in XPeng Motors' visual reasoning has increased the efficiency of the entire reasoning chain by 32 times, enabling faster cognitive processes and higher prediction accuracy while reducing prediction errors compared to traditional CoT. Thirdly, it enables multimodal output generation, including video, audio, and final action behaviors. This not only serves as the foundational base supporting VLA 2.0 but also acts as the basic framework underpinning the world model for simulation and reinforcement learning. Moreover, XPeng Motors aims to achieve cockpit-intelligent driving integration based on this foundational model, making the entire vehicle function as an organic intelligent entity.

Model, computing power, data, and hardware collectively determine its superior capabilities.

1) At the model level, XPeng Motors’ VLA 2.0 separates the L component compared to the first-generation VLA model, making it a native multimodal large model for the physical world. It will continue to iterate rapidly (currently averaging four iterations per day), with the goal of increasing the number of parameters in the onboard model to over 20 billion by the end of 2026. 2) In terms of computing power, the Turing chip was developed with integrated software and hardware. The company maximizes computing power utilization through the development of automated compilers and customizes the Turing architecture model for the chip. XPeng Motors has increased computing efficiency to 82.5%, further boosting effective onboard computing power—one Turing chip delivers the equivalent of ten Orin X chips. From the perspective of cloud training computing power, the company has built a robust AI infrastructure to support rapid iteration. From November 2025 Tech Day to early March, XPeng Motors updated its models 468 times, averaging four versions per day, and will continue to iterate rapidly. 3) Regarding data, each version of XPeng's model training uses 4 trillion tokens. Furthermore, through world model simulations, one day of testing generates data equivalent to driving 30 million kilometers, with simulation scenarios increasing from 30,000 a year ago to over 500,000. 4) The term 'ontology' refers to hardware manufacturing.

The goal is to begin unmanned operations of Robotaxi before the end of 2026, competing with Tesla FSD.

Recently, XPeng Motors’ GX model has been undergoing driverless testing in Guangzhou. The company will continue enhancing the capabilities of VLA 2.0, aiming to increase the safe intervention mileage by 50 times, average intervention mileage by 25 times, and onboard model parameters to over 20 billion by 2026, matching the latest capabilities of FSD. If XPeng indeed achieves Robotaxi operations by the end of 2026, it will become China’s first automaker to transition from L2+ to L4 and the only autonomous driving company currently capable of directly competing with Tesla FSD in the global market.

From the perspective of automotive sales fundamentals, the company’s new car cycle in 2026 remains strong, with potential to create another hit model in 2026.

This year, XPeng Motors will launch four new dual-energy models, including the large six-seat SUV GX and two Mona SUVs. The brokerage believes that the Mona series products cater to both high-volume price segments and leverage the product definition capabilities of the Juanma team to deliver superior exterior and interior designs. Meanwhile, the Mona SUV still has the potential to offer class-leading intelligent driving capabilities in the high-volume price segment, replicating the success of the Mona sedan.

Risk Warning

Risks include lower-than-expected sales of new models, slower-than-expected product launch schedules, delays in enhancing autonomous driving capabilities and implementing functionalities, heightened competition, and slower-than-expected cost reductions and gross margin improvements.