The Physical Stack of AI · Energy and grid
You can cite the IEA's 2030 datacenter demand projection, separate AI from the rest of datacenter load, and explain why rack-level density turns average energy forecasts into site-level power questions.
Two kinds of evidence matter for this chapter, and they show up over and over in 2026 reporting. The first is system-wide electricity demand: how many terawatt-hours data centers may consume in a year. The second is rack and campus density: how many megawatts a specific AI site must deliver continuously. Together they explain why the conversation has moved from "how many GPUs?" to "where do we energize the campus?"
The IEA's Energy and AI puts global datacenter electricity at around 485 terawatt-hours in 2024 and projects it to roughly 945 TWh by 2030 — almost a doubling. The US alone passes 250 TWh in 2026 and 300 TWh in 2027. The growth rate is the headline: about 15% per year, four-plus times the overall electricity-demand growth rate.
The three lessons that follow take you through:
Chapter contains 3 lessons.