According to a research report released by CICC, thanks to the innovation of model engineering optimization, DeepSeek-R1 has shown leading cost advantages and comprehensive performance, and its open-source strategy has also lowered the threshold for obtaining cutting-edge AI technologies, pushing up the demand of downstream customers for the local deployment of AI large models, and it is expected that the prosperity of privatized computing power hardware represented by all-in-one machines is expected to rise. It is recommended to pay attention to the top DS all-in-one computer suppliers Inspur Information (000977.SZ), Sugon (603019.SH), Super Fusion, ZTE (000063.SZ, 00762), Lenovo Group (00992), etc., and domestic GPU manufacturers Ascend, Muxi, Haiguang Information (688041.SH), etc.

CICC's main views are as follows:

C-end scenario: DeepSeek-R1 distillation technology achieves a breakthrough in lightweight models

Through knowledge distillation, the R1 distillation technology transferred the decision-making logic and feature characterization ability of the complex "teacher model" (671B parameters) to the lightweight "student model", and generated 6 different versions of the distillation model. These distillation models reduce the need for video memory, memory, and storage while maintaining the performance of large models, making them suitable for running on resource-constrained end devices. CICC believes that PCs are important terminals that carry local models, and the deployment of higher-performance device-side models is expected to become a powerful driver for the upgrading of the AI PC industry.

B-end scenario: DS all-in-one provides a full-stack solution for private deployment

DS all-in-one machine is an integrated computing device designed for large-scale model applications and deployment, which can be implemented based on NV or domestic hardware, and its Chinese-made computing power chips may become the main computing power support due to meeting the mainstream downstream needs.

CICC believes that the current software and hardware collaboration of DS all-in-one machines still faces some challenges: 1) mainstream domestic AI chips lack support for FP8 accuracy, and if FP16 or BF16 accuracy is used, the hardware efficiency will decline; 2) In order to realize the deployment of the full-parameter DeepSeek-R1 model on a single 8-card server, all-in-one manufacturers need to carry out fixed-point quantization and seek a balance between computing power efficiency and model effect. However, thanks to the advantages of the DS model, as well as the advantages of local privatization and rapid deployment of all-in-one machines, the domestic market space of DS all-in-one machines is expected to increase rapidly. DS all-in-one has advantages in meeting enterprise data security and compliance requirements, and has a high degree of adaptability to industries with high data security requirements such as government and finance, and we estimate that the market size of DS all-in-one is expected to reach 54 billion yuan in 25 years under the optimistic scenario.

Earnings forecasts and valuations

Related companies include the head DS all-in-one computer suppliers Inspur Information, Sugon, Super Fusion, ZTE, Lenovo Group, etc., and domestic GPU manufacturers Ascend, Muxi, Haiguang, etc. Maintain the earnings forecast of the relevant companies unchanged; Considering the long-term improvement of Inspur Information AI server, based on 35x 25-year P/E, the company's target price was raised by 34% to 63.07 yuan, which is 5% upside from the current stock price; Considering the long-term improvement of Sugon's server business, based on 40x25 P/E, the company's target price was raised by 59% to 73.20 yuan, which is 2% upside from the current stock price; Considering that ZTE's mobile phone/server and other businesses are expected to benefit from the AI industry trend, ZTE's A/H share target price is raised by 15%/24% to RMB42/HKD 29, corresponding to 20x/13xP/E in 25 years, corresponding to 13% and 7% upside.

Risk factors

Generative AI model innovation, model localization deployment requirements, and AI computing hardware technology iteration are less than expected.