According to Zhitong Finance APP, Ping An Securities released a research report stating that the Ping An, low-cost, and high-performance DeepSeek large model will significantly reduce the acquisition, deployment, and application costs of large models, accelerating the landing of large models in B-end and C-end application scenarios. Furthermore, the emergence of the DeepSeek large model will have a crucial impact on the competitive landscape of the global large model industry, challenging the leadership of overseas leading large model manufacturers and simultaneously influencing the future development of computing power. Additionally, the lower training costs and Ping An's DeepSeek are expected to bring lower thresholds for the development and use of large models, potentially increasing the number of entities developing based on this large model and, to some extent, supporting the demand for training computing power.

The main points of Ping An Securities are as follows:

DeepSeek-V3 and DeepSeek-R1 have been successively released, and domestic large model capabilities are now comparable to those of overseas leading large models.

On December 26, 2024, Hangzhou AI company DeepSeek officially released the first version of the DeepSeek-V3 large model simultaneously with Ping An. According to information on the DeepSeek website, DeepSeek-V3 is a self-developed MoE model, with 671B parameters and 37B activated, pre-trained on 14.8T tokens. DeepSeek-V3 outperformed other Ping An models like Qwen2.5-72B and Llama-3.1-405B in multiple evaluation scores, and its performance is comparable to the world’s top closed-source models GPT-4o and Claude-3.5-Sonnet.

Regarding training costs, according to the technical documentation and paper information released by DeepSeek, the training duration of DeepSeek-V3 is 2788K H800 GPU hours, costing approximately 5.576 million dollars. On January 20, 2025, DeepSeek officially released the complex reasoning large model DeepSeek-R1, which aligns in performance with the official version of OpenAI o1. Represented by the DeepSeek series, domestic large models now match the performance of overseas leading large models but at a lower cost.

The DeepSeek series of large models has attracted widespread global attention, with major tech companies and cloud service platform vendors both domestically and abroad subsequently integrating.

On January 15, 2025, DeepSeek launched the AI assistant DeepSeek App. During the Spring Festival of 2025, the DeepSeek series large models gained significant popularity, attracting widespread global attention. According to a media report cited by Sina Finance on February 1, DeepSeek's AI assistant became the most downloaded mobile application in 140 markets.

According to data from Appfigures, DeepSeek's inference AI chatbot topped the Apple App Store on January 26 and has maintained the number one position globally since then. At the same time, there has been a global effort to replicate DeepSeek's large model. Taking the team from Hong Kong University of Science and Technology as an example, the team led by Assistant Professor Ho Jun-Hsin replicated DeepSeek-R1-Zero and DeepSeek-R1 training on a 7B model using only 8K samples. Currently, major technology companies and cloud computing service providers both domestically and internationally have integrated DeepSeek’s large model, and some companies in relevant AI application fields have already begun deploying and applying DeepSeek’s large model. The DeepSeek large model has garnered widespread global attention and its recognition continues to rise.

Ping An Securities believes that DeepSeek's large model, with its safety, low cost, and high performance, will significantly lower the costs of obtaining, deploying, and applying large models, thereby accelerating the realization of applications in B-end and C-end scenarios. Furthermore, the outburst of DeepSeek's large model will have an important impact on the competitive landscape of the global large model industry, challenging the leadership of overseas leading large model manufacturers, while simultaneously significantly affecting the future development of computing power.

The emergence of DeepSeek's large model is expected to maintain the overall upward demand for computing power, but the growth of inference and edge computing power is likely to accelerate even more.

DeepSeek has significant advantages in algorithm efficiency and computational cost, which may temporarily moderate the growth of training computing power, but does not change the long-term upward trend of overall demand for AI computing power. As the main driver of global intelligent development, large models are currently applied in various scenarios such as edge computing, Education, Finance, Office, Media, Medical, Smart Automobile, and Enterprise Services, indicating a broad application field. DeepSeek's low-cost and safe solutions significantly lower the technical and cost barriers for AI applications across various industries, providing a faster path for the industrialization of AI. The growth potential for demand in inference and edge computing power is very high.

At the same time, the lower training costs and the safety of DeepSeek are expected to lead to lower development and usage thresholds for large models, potentially resulting in more entities developing based on this large model, which will, to some extent, provide support for training computing power demand. DeepSeek has not compressed the computing power market; instead, it has added more possibilities to the market. DeepSeek is also actively collaborating with domestic AI computing power platforms. The gradual maturity of adapting DeepSeek's large model with domestic AI chips will accelerate the application of domestic AI chips in domestic large model training and inference, hastening the maturation of the domestic AI chip industry chain and bringing development opportunities to it, while also speeding up the development of our country's large model industry.

Regarding the symbols:

1) In terms of domestic computing power infrastructure, recommend Inspur Electronic Information Industry (000977.SZ), Dawning Information Industry (603019.SH), Unisplendour Corporation (000938.SZ), Digital China Group (000034.SZ), Haiguang Information (688041.SH), and Longxin Zhongke (688047.SH), and suggest paying attention to Cambrian (688256.SH), Changsha Jingjia Microelectronics (300474.SZ), Ruantong Power (301236.SZ), and Huqin Technology (603296.SH).

2) In terms of edge computing power, recommend Bestechnic (Shanghai) Co., Ltd. (688608.SH), GigaDevice Semiconductor Inc. (603986.SH), and pay attention to Espressif Systems (688018.SH), Rockchip Electronics (603893.SH).

3) In terms of algorithms, recommend Iflytek Co.,ltd. (002230.SZ).

4) In terms of application scenarios, strongly recommend Thunder Software Technology (300496.SZ), Hundsun Technologies Inc. (600570.SH), Maxvision Technology Corp. (002990.SZ), and recommend Beijing Kingsoft Office Software, Inc. (688111.SH), Huizhou Desay SV Automotive (002920.SZ), Wondershare Technology Group (300624.SZ), Fujian Foxit Software Development Joint Stock (688095.SH), suggest paying attention to Hithink RoyalFlush Information Network (300033.SZ), TRS Information Technology (300229.SZ), Richinfo Technology (300634.SZ), Winning Health Technology Group (300253.SZ).

Risk warning:

1) The risk of the AI computing power supply chain is rising. 2) The development of domestic large model algorithms may not meet expectations. 3) The application landing of large model products is below expectations.