The rapid evolution of artificial intelligence (AI) workloads has fundamentally transformed the landscape of data center power infrastructure [1], [2]. Contemporary AI training clusters for Large Language Models (LLMs), such as those developed by OpenAI, Google, and Meta, now routinely consume power at megawatt scales (or even beyond) while exhibiting unprecedented power dynamics during operation [3]. This paradigm shift necessitates a comprehensive reexamination of power delivery architectures, emphasizing the fundamental limitations imposed by cascaded power conversion chains. Realizing reliable and efficient AI data centers ultimately requires an end-to-end perspective—spanning facility power, workload management, and grid integration—in which power electronics serves as one of the core enabling technologies. Understanding these constraints is critical not only for current deployments but also for the strategic development of next-generation AI infrastructure.
For more on this article click here.