AI & energy: the bigger picture

AI's footprint is real — and, today, deliberately hard to see. These are some of the numbers that put your own usage in context. Each one is sourced; follow the link to read it at the source.

Two references carry this page for now: the UNU-INWEH report The Environmental Cost of AI's Energy Use (2026) and FluixAI's breakdown of how data centers are cooled. We'll add more as we find numbers we can stand behind.

Training a frontier model

• Meta's LLaMA 3-405B reportedly used about 21 GWh of electricity to train — roughly 31 million GPU-hours of compute. (UNU)
• GPT-4 likely consumed 50–70 GWh over about 100 days — 40–55 times more than GPT-3, which used 1.287 GWh. (UNU)
• GPT-5's training is projected at around 100 GWh — roughly the annual home electricity use of 770,000 people. (UNU)

Training is the small part

The headline training runs are huge — but they're a one-time cost. The bulk of AI's energy is spent after launch, answering prompts.

• Inference — answering billions of everyday prompts — is estimated to account for 80–90% of total AI energy use. Training is the smaller slice of a model's lifetime footprint. (UNU)
• Generating one AI image takes about 2.9 Wh — 60 times a short text answer, and 1,450 times a text classification. (UNU)
• A high-resolution AI video clip can draw more than 415 Wh — as much electricity as 200,000 text classifications. (UNU)

The data centers behind it

• In 2025, data centers consumed an estimated 448 TWh of electricity — if they were a country, the 11th-largest electricity user on Earth. (UNU)
• On current trajectories, data-center electricity demand could roughly double to 945 TWh by 2030. (UNU)
• Nearly half of the world's data centers are in the United States. (UNU)
• There are roughly 11,000 data centers in operation around the world today. (FluixAI)
• The global AI market is projected to grow from $189 billion in 2023 to nearly $5 trillion by 2033. (UNU)

Cooling & water

Heat is the other half of the story. How a data center sheds it decides how much water it drinks.

• Data centers are cooled three ways — of the ~11,000 today, roughly 5,000 use dry-air, 3,000–4,000 use evaporative, and 1,000–2,000 use liquid cooling. (FluixAI)
• Evaporative cooling is the most water-hungry of the three; liquid cooling can use essentially zero water per year. (FluixAI)
• A typical data center uses about 100 million gallons of water a year; a large evaporative one can use roughly 2 billion. (FluixAI)
• Of the 3,000–4,000 data centers expected to be built this decade, about half will be liquid-cooled. (FluixAI)

Why this is here

You can't act on what you can't see. We show your own AI usage in energy, water, and CO₂ — read how we calculate it, then sign in and run your numbers →.