Quick answer

AI data centers use ~3-4% of global electricity in 2026, up from 1% in 2022. Per query: a ChatGPT response uses 0.3-3 watt-hours (similar to a Google search). The big energy use is TRAINING — GPT-5 training consumed ~250 GWh (powers a city for a month). Tech companies are signing nuclear and geothermal power deals. The grid is genuinely stressed in some regions.

The AI energy debate is fierce in 2026. Some claim AI is destroying the climate; others say it is overblown. Here are the actual numbers and what they mean.

How much energy does AI actually use?

  • Training GPT-5 — estimated 250 GWh one-time (city of 100k for a month)
  • A single ChatGPT response — 0.3-3 watt-hours (similar to a Google search)
  • Image generation — 10-100x more per query than text
  • Video generation (Sora 2) — 1000x more than text per query
  • Global AI data centers — ~150 TWh/year in 2026 (3-4% of all electricity)

The grid impact

Some regions are genuinely stressed. Virginia (data center capital) is delaying new data centers because the grid cannot handle them. Texas, Arizona, and parts of Europe face brownout risk during peak demand. AI is not the only cause (crypto, EV charging, AC use), but it is the fastest-growing.

How tech companies are responding

  • Nuclear power deals — Microsoft (Three Mile Island restart), Amazon, Google all signed nuclear PPAs in 2024-2025
  • Geothermal — Google partnered with Fervo for next-gen geothermal
  • Direct fusion bets — OpenAI, Microsoft invested in fusion startups
  • Renewable PPAs — most hyperscalers buying solar and wind at scale
  • On-chip efficiency — Nvidia Blackwell, custom inference chips dramatically reducing per-query energy

Honest framing: AI energy use is significant but small compared to existing industrial sectors. The concern is GROWTH RATE — if AI demand doubles annually, the grid runs out of capacity before generation can scale.

Should you feel guilty using AI?

A typical professional using ChatGPT 50 times daily consumes maybe 50-150 watt-hours — about the same as one hour of Netflix HD streaming. Individual AI use is not your big environmental concern. The big use is in training and corporate-scale inference.

Related reading

The State of AI — May 2026 Mid-Year Report What Is AI Inference vs Training? Plain English

Bottom line

AI uses real electricity at meaningful scale, but it is a small fraction of total grid load. The crisis is regional and short-term — the grid cannot scale as fast as AI demand. Long-term, AI is bringing money to nuclear and geothermal that may actually help decarbonisation.