Building AI servers is completely different from traditional servers; it goes much further. It is not limited only to the use of storage, memory, or even the CPU. Machine/Deep learning applications require massive computing power, far greater than that provided by even the best CPUs. The problem is that there are not enough people who can build and manage AI servers well. Because of the information gap and lack of knowledge, it will take a lot of effort, research, and blind attempts based on prior experience in building traditional servers, which leads to wasting time and money and a lot of errors. All of these may cause huge technical debts on various parties, including wastage of resources, failure to meet needs, lack of consistency and stability, and the biggest is the lack of sufficient specialists in this type of server, and thus the difficulty of management, maintenance, and ensuring consistency and stability. Therefore, in this paper, we will seek to develop an appropriate methodology for building and managing on-premise deep learning servers. Seeking to clarify the problems that may be encountered and provide appropriate solutions for them and decision-making support to eliminate the gap between IT and MLOps and to be a method for companies and organizations that want to build on-premise servers for machine/deep learning.