Automating Electrical Oversubscription Calculation in Data Centers Using AI Agents

Data centers have recently experienced exponential growth, mainly due to the increasing computing power demand of Artificial Intelligence (AI) applications. To meet this demand, data centers have undergone a radical transformation in their electrical supply layout and in the technical specifications of their main machines. This thesis explores the peculiarities of electrical systems in data centers, with regard to the power electronics components, focusing on the latest optimizations that allowed a higher power density and made possible the deployment of state-of-the-art power-intensive computing devices, such as high-performance graphics processing units (GPU). That’s why the dissertation will explore the project Open Rack as part of the open source Open Compute Project (OCP), having Meta (ex-Facebook) and Google as main contributors, with the goal of designing modern solutions for efficient data centers, starting from innovating the electrical supply of the single racks. Finally, generative AI applications to help the automation of daily tasks in Data Centers will be discussed, based on my project at Amazon Web Services as a data center engineering operations intern, about an AI Agent to calculate the oversubscription on new rack arrivals, that is, a safety check concerning the power drawn by the newly installed IT devices, which made possible the testing of AI applications on technical engineering tasks, with promising results for future research in the area.