China Securities Co., Ltd.: Domestic computing power sector welcomes intensive catalysis, accelerating the landing of ecology and improving competitiveness.

China Securities Co., Ltd. released a research report stating that the domestic computing power sector is facing intensive catalysis, and Tencent's performance meeting indicated that there are multiple choices for the supply channels of inference chips, and domestic inference computing power chips are expected to provide assistance in the background of fluctuations in the international supply chain. DeepSeek has updated its model to specifically support FP8 precision and the next generation of domestic chips that will be released. The support of the top domestic open source models is expected to promote the rapid implementation of domestic computing power ecology by supporting domestic chips. China Securities Co., Ltd.'s main points are as follows: Events On August 13th, Tencent's performance meeting stated that the company's inference chip supply channels have multiple choices. On August 21st, DeepSeek updated its model version to DeepSeek-V3.1, which is expected to support FP8 precision and domestic chips. Tencent's performance meeting indicated that the future supply of inference chips is expected to diversify, benefiting domestic computing power chip manufacturers. On August 13th, TENCENT (00700) released its 2025Q2 performance and held a briefing session, during which Tencent's President, Liu Chiping, stated that the company's AI chip reserves are sufficient to support the training and iterative upgrades of existing models. The increase in users in various applications such as Yuanbao and AI Search has led to an increase in the percentage of inference loads, and the company is controlling the inference costs through lightweight models and software optimization, while also having multiple choices for the supply channels of inference chips. With significant uncertainties in overseas high-end chip trade policies, Tencent is expected to rely on domestic manufacturers such as Huawei, Cambricon, and Horizon Robotics to meet the related computing needs. DeepSeek V3.1 is released, expected to support FP8 precision and domestic chips. According to the official public account, DeepSeek has officially released DeepSeek-V3.1 today. The official website and the web-based model have been upgraded to the latest version, and the new model has been open-sourced on Huggingface. The main upgrades of the new version model are as follows: (1) a model that simultaneously supports thinking mode and non-thinking mode; (2) compared to the previous models, the new model in thinking mode can provide answers in a shorter time; (3) through Post-Training optimization, the new model has stronger agent capabilities. In addition to functional upgrades, the new version model uses UE8M0 FP8 Scale parameter precision, representing the FP8 precision format designed for AI core operations such as matrix multiplication. Compared to FP16 format, FP8 can achieve double the performance on the same chip area, with lower power consumption and bandwidth requirements. Furthermore, DeepSeek indicates that the new precision format is designed for the upcoming next generation of domestic chips, indicating that the future training and inference based on the DeepSeek model are expected to be more applied to domestic AI chips, helping to accelerate the construction of domestic computing power ecosystem. Huawei's Ascend chips are accelerating their breakthrough, and the competitiveness of the domestic chip market is gradually increasing. Recently, server products based on Huawei's Ascend chips have been successively deployed in large orders in government, financial, and telecommunications industries. In the Shanghai Pudong Development Bank's large-scale computing power expansion project tender, it includes 176 Kunpeng + Ascend configured intelligent servers; in the Xinjiang Jiuqian Wanke hub-type intelligent computing center project tender, it includes 4500 Ascend 910C-2 server clusters; in the China Mobile Limited 2025-2026 inference AI computing equipment tender, it includes 5000 sets of CANN ecological devices, which are server products based on Ascend heterogeneous computing architecture and Ascend chips. Due to the continuous risks in overseas supply chains, the cost-effectiveness of domestic chip manufacturers' products continues to increase, and more computing needs are expected to be supported by domestic chips, leading to the continuous increase in the competitiveness of the domestic chip market. Risk warning Accounts receivable bad debt risk; intensified industry competition; impacts of international environmental changes.