A single NVIDIA H100 GPU draws around 700W; a 10,000-GPU cluster therefore needs 7+ MW before cooling and networking overhead. Frontier training runs like GPT-4-class models are estimated to consume tens of gigawatt-hours over months. Inference at scale compounds this: serving billions of daily queries across millions of GPUs makes total inference energy consumption rival training. Data center power availability has become a strategic bottleneck, driving investments in nuclear, geothermal, and dedicated power infrastructure.