The Short One AI Power Consumption: The Coming Crisis (Short
Analysis – December 2025)
Executive Summary
Global data center electricity demand is exploding, driven by AI. Current consumption: ~860 TWh/year (~98
GW average load). Projections show doubling by 2030 and tripling by 2035. By mid-century, AI alone could
require 400-600 GW — more than the current electricity consumption of the UK, France, and Germany
combined.
Earth's announced projects cover only ~30-50% of projected demand. The result: a widening "GW canyon" that
risks stalling AI progress unless new infrastructure paradigms emerge.
Key Projections Table (Mid-Range GW Capacity Needs)
Earth Supply Reality
Announced mega-projects (Stargate, Colossus, Hyperion, etc.) deliver ~157 GW by 2042. After that, terrestrial
buildout plateaus due to grid constraints, permitting delays, water scarcity, and land competition. Cost per GW
on Earth: $25-50 billion, with ongoing electricity/cooling opex in the billions annually.
The Canyon
Post-2035, demand outstrips known supply by hundreds of GW. Without innovation, AI scaling hits a hard wall
— compute rationed to highest bidders, frontier model training becomes uneconomic.
Pathways Forward
Efficiency gains help but trigger Jevons Paradox: cheaper compute creates more demand. The release valve:
resilient, low-opex alternatives that bypass Earth's physical and regulatory limits.
This isn't speculation — it's math. The crisis is coming. The opportunity is now.
Phil Cheevers | Pink House Technology | pinkhouse.tech
Year AI/Data Center Demand (GW) Notes/Sources
2025 ~100 Gartner/IEA baseline
2030 150-250 IEA, Goldman Sachs, Gartner
2035 200-350 BNEF, IEA Lift-Off scenario
2042 300-450 Moderate extrapolation
2052 400-600 Tapered high scenario with efficiency gains
pinkhouse.tech Page 1 of 1 January 2026