AI workloads influence data center power systems differently than traditional IT workloads. Most data center electrical infrastructure was engineered for lower rack density, some by a factor of 10x or more compared to the 100+kW racks of today. Modernizing to meet evolving requirements – in a living operating data center - is a real obstacle to leveraging the benefits of AI factories.

We are still in the early stages of understanding how different AI workloads impact data center power systems. Most assume that power system challenges are limited to the pretraining and post-training (e.g., fine-tuning) of large language models (LLMs).

Power profiles are essential for predicting how a data center’s power system will respond to specific AI workloads. While we may not have comprehensive profiles for every workload, in this whitepaper we have identified five key attributes and trends that help us estimate the demands of a worst-case scenario. By designing data center power systems to accommodate these worst-case profiles, we can better verify readiness for future generations of AI workloads.