Soochow: Maintaining XPENG-W (09868) "Buy" rating, accelerating physical AI progress.

Soochow released a research report stating that XPENG-W(09868) is currently undergoing a new round of strong new car cycle+globalization+Robotaxi+Siasun Robot&Automation four-fold resonance phase. The company maintains its forecasted operating revenue for the years 2026/2027/2028 to be 91.4/126.5/154.6 billion yuan respectively, with year-on-year growth of +19%/+38%/+22% respectively. The company also maintains its forecasted net profit attributable to shareholders for the years 2026/2027/2028 to be -2.1/1.5/5.8 billion yuan. The EPS for the years 2026/2027/2028 are -1.08/0.78/3.03 yuan respectively, with corresponding PEs of 65/17 times for 2027/2028. XPENG Group's layout in the field of physical AI is forward-looking and leading, with the current valuation being significantly undervalued. The company maintains a "buy" rating. Soochow's main points are as follows: The world model makes its debut at the top AI conference CVPR, the second generation VLA and the world model complement each other, accelerating the progress of physical AI. In June 2026, XPENG Group's related research on the world model was presented at the International Conference on Computer Vision and Pattern Recognition (CVPR). The second generation VLA and the world model are not mutually exclusive or competing, but are two aspects of the same physical world model under different training signals: VLA is supervised by human driving behavior actions, solving "how to act", while the world model predicts the next frame in a self-supervised manner, teaches dynamics and causality, and solves how the world will change after action. Both are applied together to train a unified base using Scaling Law. The second generation VLA omits the language intermediate layer, enabling rapid iteration from the cloud to the vehicle, and the effectiveness of the push is beginning to show. The second generation VLA combines vision and language modeling at the base layer, omitting the explicit translation of the intermediate layer in language to reduce information loss and improve reasoning efficiency. It is trained on about 30,000 GPU clusters in the cloud, with research and development investment of over 20 billion yuan, and a full cycle iteration approximately every five days. The second generation VLA was pushed in March 2026, with the first month of push contributing to over 50% of the assisted driving mileage, and the company plans to push a new version in the third quarter of 2026 to further enhance the capability. X-World, X-Foresight, X-Cache, the three works combined with DrivePTS officially included in CVPR, constitute the academic support of the world model. X-World achieves controllable multi-camera scene generation through VAE latent space diffusion (spatial compression by 16 times, temporal compression by 4 times); X-Foresight, built on top of VLA, predicts multiple camera images simultaneously while outputting actions through blockwise autoregression; X-Cache uses cross-block caching to achieve up to approximately 2.7 times faster inference speed and reduce about 70% of repetitive calculations in denoising the main chain; DrivePTS achieves controllable generation of rare scenes through progressive learning + mutual information constraints + frequency-guided structural losses. Commercialization of three major applications of physical AI accelerates. The second generation VLA, Robotaxi, and humanoid Siasun Robot&Automation IRON are advancing synchronously. Turing chips began mass production and shipment from 2025Q3, with a target of nearly 1 million units shipped in 2026, and all models switched in the second quarter, with Volkswagen becoming the first external customer of Turing chips and the second generation VLA; Robotaxi completed the production of the first mass-produced vehicle in Guangzhou on May 18, 2026 (China's first full-stack self-developed vehicle factory), with the company expected to start demonstration operations in the second half of 2026; IRON is planned for mass production by the end of 2026, with a monthly production target of thousands of units, and on June 10, 2026, He Xiaopeng announced in an internal memo personally taking on the role of the head of Siasun Robot & Automation. Risk warning: Passenger car price wars exceed expectations; terminal demand recovery is lower than expected; the pace of the introduction of L3/L4 level intelligent driving policies is slower than expected; the industrialization progress of humanoid Siasun Robot & Automation is slower than expected.