Start with he premise that energy is not created or destroyed, it only changes state. Pushing processors to perform a task uses energy and creates heat to rearrange numbers and pixels. Combing through huge numbers of data vits takes a lot of energy. If you've ever worked with systems larger than a desktop you know how much they can draw.

My whole home lab uses about 700 watts for basic server duties. To power a single 3090 GPU to run some larger models my PC needed a 1000 watt PSU. Extrapolate that out to a commercial size org and you can see where the major power draw comes in.