Guotai Junan: Focus on investment opportunities for large-scale deployment of local enterprise models, optimistic about the surge in demand for domestic inference computing power.

Guotai Junan released a research report stating that local deployment may be the most suitable AI invocation method for large enterprises and companies in special industries. As the number of future application scenarios increases, the frequency of model invocation will also significantly increase, effectively driving the demand for enterprise local inference computing power. Optimistic about the outbreak of domestic inference computing power locally, as well as the expansion of new basic software such as vector databases, optimistic about the opportunity for the AI upgrade of enterprise IT infrastructure. Guotai Junan's main points are as follows: Local deployment of large models meets national conditions and regulatory requirements The biggest highlight of this DeepSeek-R1 is open source, significantly reducing the cost of enterprises deploying first-class large models locally. Enterprise local deployment of DeepSeek meets the IT spending habits of domestic enterprises, as well as the requirements for security and privacy in data-sensitive industries such as finance, energy, government affairs, and healthcare. Local deployment may become the preferred AI invocation method for large enterprises and companies in special industries. Local computing power demand is underestimated, optimistic about the outbreak of domestic inference computing power demand at the enterprise end Recently, many individual users have achieved local deployment of models on PCs by downloading distilled models with parameters ranging from 1.5B to 70B, creating the illusion that "DeepSeek inference does not require much computing power." Enterprise-level deployments often need to meet multi-user concurrent scenarios, and the demand for inference computing power will be significantly higher than that of personal deployment scenarios. As the number of future application scenarios increases, the frequency of model invocation will also significantly increase, effectively driving the demand for enterprise local inference computing power. The focus of the future market may shift from the AI capital expenditures of Internet giants to the AI capital expenditures of various enterprises, superimposed on the background of "trusted innovation," optimistic about the outbreak of domestic inference computing power demand at the enterprise end. Vector databases become indispensable tools After enterprises locally deploy large models, they can integrate private data into knowledge bases and use RAG (Retrieval Augmented Generative) to enhance the model's vertical domain expertise. Among them, vector databases are essential infrastructure for vector retrieval in this process. It is expected that in the future, vector databases will no longer be tools exclusive to large model vendors but will generalize to all enterprises deploying local models, optimistic about the opportunity for the AI upgrade of enterprise IT infrastructure. Investment advice: Recommended targets: Inspur Electronic Information Industry (000977.SZ), Unisplendour Corporation (000938.SZ), Transwarp Technology (Shanghai) Co., Ltd. (688031.SH), iSoftStone Information Technology (301236.SZ). Beneficiary targets: Dawning Information Industry (603019.SH), Digital China Group (000034.SZ), Talkweb Information System (002261.SZ), TRS Information Technology (300229.SZ), etc. Risk warning: Technological progress is not as expected, risks of intensified competition.