AI Data Center Power Scaling

Overview

AI data center power scaling describes the exponential growth in per-facility power requirements driven by the thermal and electrical demands of GPU-accelerated AI workloads. A decade ago, a large data center was considered significant at 10–15 MW. By 2025, industry conversations have moved to gigawatt-scale AI campuses — an increase of two to three orders of magnitude in total facility power.

This escalation is the direct result of GPU TDP increases (now exceeding 1,000W per chip per generation in H-series and B-series hardware), the move to tightly coupled GPU clusters (e.g., NVL72 racks at 130-160 kW), and the scale of AI training and inference deployments required to serve large language models and generative AI systems. At OCP Global Summit 2025, Pegatron identified power as the number-one AI infrastructure challenge, followed by thermal management and scaling cost/ROI. The consequent migration from air cooling to direct-to-chip liquid cooling from ~2024 onward reflects the physical limits being reached with air-cooled infrastructure.