DeepSeek reconstructs the narrative of domestic AI, and domestic cloud factories may welcome the "GPT" moment. At the end of November 2022, OpenAI released ChatGPT, raising global attention to AI. Driven by the AI strategy, overseas cloud factories such as Microsoft, Google, Amazon, and Meta have continuously increased capital expenditures and increased investment in AI and cloud infrastructure in recent years. Judging from the market trend of the above-mentioned cloud factories in the past five years, the four major cloud factories in North America have achieved a "V-shaped" reversal around November 2022, and the AI strategy has been gradually realized. After the release of DeepSeek-R1, its algorithm optimization is expected to make domestic cloud vendors' investment in AI and cloud more cost-effective, and with Alibaba's expected outlook for future cloud and AI investment, the capital expenditure of domestic cloud manufacturers is expected to accelerate, and they continue to be optimistic about the performance growth logic of the domestic computing power industry chain, among which domestic cloud computing vendors are the most direct beneficiaries.

Step Star and DeepSeek adhere to open source, and the domestic AI ecosystem is expected to further prosper.

On February 21, 2025, Step Leap Stars held the Step UP Ecosystem Open Day, showcasing the layout of AI Agent and vertical AI applications. Recently, Step Star and Geely Automobile Group have open-sourced Step-Video-T2V, the world's largest open-source video generation model with the largest number of parameters and the best performance, as well as Step-Audio, the industry's first product-level open-source voice interaction model. This week, DeepSeek will launch a week-long open-source technology sharing activity, and will open source 5 codebases one after another. On the first day of this week, DeepSeek officially open-sourced FlashMLA. FlashMLA is a high-efficiency MLA decoding core for Hopper GPUs, based on NVIDIA's CUDA 12.6 architecture, which enables the 3000 GB/s memory bandwidth limit and 580 TFLOPS (trillion floating-point operations per second) compute limit on the H800 SXM5 GPU. Prior to this, DeepSeek's base model V3 used MLA, an innovative attention mechanism. By employing MLA and DeepSeekMoE (Hybrid Expert Architecture), the V3 large model can achieve efficient inference and economical training. Through further optimization, the open-source FlashMLA effectively excavates the computing performance of computing hardware, improves the speed and capability of the inference side, and is expected to accelerate AI inclusion.

Investment suggestion: We believe that the current domestic AI algorithm optimization will greatly improve the efficiency of computing resources, so the investment of domestic manufacturers in AI and cloud is more cost-effective. At the same time, domestic AI leaders adhere to the open source strategy to create a good ecosystem for the technical resonance of the domestic AI industry. Maintain the "Recommended" rating for the computer industry. It is recommended to pay attention to iFLYTEK (002230. SZ), Sugon (603019. SH), Shensunda A (000032.SZ), iSoftStone (301236.SZ), Digital China (000034. SZ), Inspur Information (000977. SZ), Runze Technology (300442. SZ), Yunsai Zhilian (600602. SH), Kingsoft Office (688111. SH）。

Risk warning: the pace of substitution of domestic computing power is not as expected; the risk of error in the data cited; AI investment is not as strong as expected; Increased competition for AI products; Focus on the company's performance not meeting expectations; The speed of policy standards is not as fast as expected; The risk of existing technology being replaced due to the upgrading of AI technology.