Guosheng Securities has issued a research report maintaining a "Buy" rating on XPENG-W (09868) with a target price of HK$105.7 and a target of $27.0 for its US-listed shares (XPEV.US). The firm is optimistic about the company's strong product cycle, overseas expansion, advancements in autonomous driving, and emerging business opportunities such as robotics and Robotaxi. It forecasts the company's sales for 2025-2027 to be approximately 430,000, 570,000, and 840,000 vehicles, respectively, with total revenues reaching RMB 752 billion, RMB 1,031 billion, and RMB 1,455 billion. The non-GAAP net profit margin is projected to be -1.2%, 2.3%, and 3.2% for the same periods.
Considering the deepening cooperation with Volkswagen, the firm has provided a breakdown of its business forecasts. It estimates core business revenue will reach RMB 100.1 billion in 2026, with profits from the Volkswagen collaboration contributing approximately RMB 2.7 billion. Furthermore, other growth avenues like Robotaxi and robotics are expected to be gradually incorporated into the valuation framework as they achieve commercial viability this year.
The main points from Guosheng Securities' analysis are as follows:
XPeng's VLA 2.0 is a native multimodal large model for the physical world. Unlike the digital world, the physical world presents unique challenges that increase complexity: 1) Input signals are continuous, unstructured data, unlike text which is easily segmented; 2) Signal outputs are continuous, such as steering wheel control; 3) The physical world involves feedback and interactions with inherent uncertainties. VLA 2.0 is a foundational model designed natively for these multimodal physical world challenges. To address the difficulties of physical AI models, XPeng has developed a native multimodal tokenizer—a signal processing unit—that enables more efficient and primitive encoding of all signals, achieving native fusion of multimodal information and avoiding biases from single modalities.
XPeng's visual reasoning Chain of Thought (CoT) has increased reasoning efficiency across the entire thought chain by 32 times, leading to faster cognitive processes and higher prediction accuracy, while reducing prediction errors compared to traditional CoT. Thirdly, the model enables multimodal output, capable of generating video, audio, and ultimately action behaviors. This framework not only underpins VLA 2.0 but also serves as the foundation for world models used in simulation and reinforcement learning. Additionally, XPeng aims to use this base model to integrate cabin and autonomous driving systems, creating a cohesive, intelligent entity.
The model's superior capability is determined by four key factors: model architecture, computing power, data, and hardware embodiment. 1) Regarding the model, VLA 2.0 is a native multimodal large model for the physical world, with the 'L' (language) component decoupled compared to the first-generation VLA. It is undergoing rapid iteration, with approximately four updates per day currently, targeting over 20 billion parameters for the vehicle-side model by the end of 2026. 2) In terms of computing power, the Turing chip is developed with integrated hardware and software. XPeng maximizes computing utilization through an automated compiler and customizes the Turing architecture model for the chip, achieving a computational utilization rate of 82.5%. The effective computing power per vehicle has increased significantly, with one Turing chip delivering effective computing power equivalent to ten Orin X chips. For cloud-based training compute, the company has built robust AI infrastructure to support rapid iteration. From the Tech Day in November 2025 to early March, XPeng updated its models 468 times, averaging four versions per day, with continued rapid iteration expected. 3) Regarding data, each model version is trained on 4 trillion tokens. Furthermore, through world model simulation, one day of testing equates to 30 million kilometers of real-world driving data. The number of simulation scenarios has grown from 30,000 a year ago to over 500,000. 4) The term 'embodiment' refers to hardware manufacturing.
Aiming for unmanned Robotaxi operations by the end of 2026 to compete with Tesla's FSD. Recently, XPeng's GX model began unmanned driving tests in Guangzhou. The company plans to continuously enhance VLA 2.0's capabilities, targeting a 50-fold increase in safe disengagement distance and a 25-fold increase in average disengagement distance by 2026. It also aims to scale the vehicle-side model parameters to over 20 billion, matching the latest capabilities of FSD. If XPeng successfully launches Robotaxi operations by the end of 2026, it would be the first Chinese automaker to advance from L2+ to L4 autonomy and the only autonomous driving company capable of directly competing with Tesla's FSD in the global market.
From a fundamental automotive sales perspective, the company has a strong new vehicle cycle in 2026, with the potential to create another hit model. This year, XPeng will launch four new dual-energy vehicles, including the large six-seater SUV GX and two Mona SUVs. The firm believes the Mona series targets volume price segments while leveraging the product definition capabilities of the Juanma team to deliver superior exterior and interior design. Simultaneously, the Mona SUV is expected to offer class-leading intelligent driving capabilities in its price segment, replicating the success of the Mona sedan.
Risk factors include potential underperformance of new model sales, delays in product launch timelines, slower-than-expected improvements in autonomous driving capabilities and feature deployment, intense competition, and failure to meet cost reduction and gross margin improvement targets.
Comments