AI Policy Wiki
Dashboard

RAND — AI's Power Requirements Under Exponential Growth (2025)

high confidence · updated 2026-06-06

RAND report (Pilz, Mahmood, Heim, 2025) forecasting AI data-center power demand through 2030 and framing power as a compute-governance chokepoint.

AI's Power Requirements Under Exponential Growth (RAND Research Report RRA3572-1, 2025) is a report by Konstantin Pilz, James Sanders Mahmood, and Lennart Heim that forecasts AI data-center electricity demand through 2030 and argues that power, rather than chips, is becoming the binding constraint on AI scaling. RAND positions the report as a policy-facing companion to Epoch's compute forecasts, and it has been cited by CSIS, the Department of Defense, and BIS.

Summary of findings

Under continued exponential growth, the report projects global AI data-center power rising from roughly 10–15 GW in 2024 to 70–130 GW in 2030, and US AI data-center power rising from roughly 5–8 GW in 2024 to 40–80 GW in 2030. At the upper end, the US figure could represent 8–15% of total US electricity consumption by 2030. The power drawn by a single frontier training run is projected to rise from roughly 100 MW in 2024 to 2–10 GW in 2030, an estimate whose central value sits slightly below, but overlaps with, Epoch's upper bound.

The report's central argument is that power is overtaking chips as the limiting factor on AI scaling, which makes power a first-class compute-governance lever. It gives four reasons. Power is slow to build, with RAND citing 5–15 years for nuclear, 3–7 years for gas with interconnect, and 12–18 months for accelerator shipments. Power is geographically fixed, so unlike chips it cannot be re-routed through shell companies. Power is visible, because interconnect queues and power-purchase agreements generate public paper trails. And power is already politically contested, with utilities, ratepayers, and local governments holding standing.

Policy options discussed

The report surveys policy levers that follow from treating power as a chokepoint: strategic siting on federal land, retiring coal sites, and near existing nuclear capacity; ally-shoring to Canada, the Nordics, and Gulf states; disclosure regimes for large training runs covering site, power draw, and counterparty, modeled on BIS chip-export reporting; transmission buildout functioning as de facto AI industrial policy; and restricting US-origin power equipment such as transformers and turbines to adversary-hosted AI, framed as an analogue to chip export controls. RAND argues that power controls are a natural analogue to chip controls (Export Controls (AI)), and the NEPA and permitting provisions of the America's AI Action Plan are an upstream response to this forecast.

The report identifies three tensions running through these options. A climate-versus-capability tension arises because the marginal AI megawatt is often gas-fired in the near term. A national-security-versus-federalism tension sets state and local siting authority against the federal interest in concentrated frontier compute. And an industrial-policy-versus-free-market tension reflects that direct federal grid and nuclear investment for AI is politically novel in the US.

Relation to Epoch's forecasts

RAND's aggregate, fleet-wide figures align with Epoch's: the 70–130 GW global and 40–80 GW US estimates are consistent with Epoch's roughly 90 GW US-only estimate under continued scaling. On per-run figures, RAND's 2–10 GW central estimate sits slightly below Epoch's 4–16 GW range, a difference RAND describes as within forecasting error and reflecting its median-scenario framing against Epoch's trend-extrapolation bracket. The report characterizes the two efforts as complementary rather than contradictory, with Epoch quantifying the raw trajectory and RAND translating it into policy levers. The report anchors the power-chokepoint framing of Compute Governance with a primary source and supplies US-specific shares for AI Environmental Impact (8–15% of US electricity by 2030).

Relationships

Citation

Pilz, K., Mahmood, J. S., & Heim, L. (2025). AI's Power Requirements Under Exponential Growth. RAND Corporation, RRA3572-1.