Training GPT-6 consumed an estimated 1.5 gigawatt-hours of electricity — equivalent to the annual consumption of 140,000 average US homes. The AI industry's energy appetite has grown 10x in three years, and there is no sign of slowing down. This is now a geopolitical, environmental, and infrastructure crisis.
The Scale of the Problem
- US data center electricity consumption is projected to reach 9% of national grid capacity by end of 2026, up from 2% in 2023
- Microsoft, Google, and Amazon have collectively signed renewable energy purchase agreements exceeding 50GW — more than the entire nuclear output of France
- NVIDIA's GB300 GPU cluster (the most powerful AI training system in 2026) requires 12MW of power — enough to power a small city
Nuclear Renaissance: AI's Unexpected Gift to Nuclear Power
Microsoft's contract to restart Unit 1 of Three Mile Island became a turning point. By 2026, every major tech company has either signed nuclear PPAs or is operating on-site small modular reactors (SMRs). AI's insatiable demand for reliable, carbon-free baseload power has single-handedly revived an industry that was in terminal decline.
"The AI industry will do more for nuclear power in five years than climate policy did in thirty." — Energy Secretary, Congressional Testimony 2026
The Efficiency Race
The response to the energy crisis has also accelerated efficiency innovation: Anthropic's model compression techniques reduced Claude 5's inference cost by 60% vs Claude 4. NVIDIA's Blackwell architecture delivers 4x more FLOPS per watt than its predecessor. Liquid cooling adoption in data centers jumped from 15% to 65% in one year, reducing cooling overhead by 40%.
The Geopolitical Dimension
Countries with cheap, abundant energy are becoming the new AI superpowers. Norway, Iceland, and Canada are attracting massive AI data center investments. Meanwhile, countries with fragile grids — including parts of Southeast Asia and Africa — risk being locked out of the AI economy entirely without significant infrastructure investment.