What is the environmental cost of training an LLM?

Training a large language model takes a genuinely large amount of electricity and water. That cost doesn't stop once training ends — every use afterward draws more power, and multiplied across millions of daily users, it can add up to an even bigger footprint than training itself.

How much electricity does training an LLM use?

Exact figures are hard to come by because companies rarely publish them, but one well-documented case is GPT-3, which researchers estimated took around 1,287 megawatt-hours to train — roughly 100 times what an average US household uses in a year. Today's frontier models are far bigger, so current figures are almost certainly higher. The honest summary is that the numbers are large and mostly hidden, which is part of why the topic draws so much attention.

Why does training an LLM use water?

Water use comes from cooling — data centers full of GPUs generate heat, and cooling systems often use evaporated water to manage it. A single AI response may use well under a milliliter, but multiplied across billions of queries a day it becomes a measurable local draw, and some new data centers face pushback from nearby communities. The impact is very local: it depends on where the data center sits and how water-stressed that region already is.

Does the cost stop once training is done?

No. Training is a one-time cost per model, but inference — every question anyone asks afterward — keeps drawing electricity for as long as the model stays in use. Because usage is so constant and large-scale, some estimates put inference's lifetime energy use several times higher than training's. In other words, the headline training numbers can actually understate a popular model's true footprint.

How big is AI's overall energy footprint?

Zooming out, AI-driven data centers used an estimated 415 terawatt-hours of electricity in 2024 — about 1.5% of global electricity — projected to more than double by 2030. The industry is responding on several fronts:

• Renewable-energy deals to power data centers with cleaner electricity.
• More efficient chips that do more computation per watt.
• Smaller distilled models doing more with less compute (as our card on parameter counts covers).

Total demand keeps rising anyway because AI adoption is outpacing those efficiency gains. Honestly, this is a real cost — not a made-up scare story, but not uniquely apocalyptic either. It's one energy-hungry industry among several, making real efforts to shrink its footprint as its scale keeps growing.