How much does AI consume: energy, water, environmental footprint
Training and using large models consumes significant and growing amounts of electricity and cooling water, documented notably by the International Energy Agency. Orders of magnitude vary greatly by model and use case: this article provides sourced figures and what they mean at the individual level.
Training and using large AI models consumes electricity and cooling water in significant and growing quantities. The International Energy Agency documents a rapid increase in data center electricity consumption, driven notably by AI. The orders of magnitude vary greatly depending on models, uses and infrastructure, and many figures circulating are uncertain or poorly sourced. This article sticks to what is established and what it means at an individual level.
What consumes
Training large models is intensive: it mobilizes vast computing infrastructures for weeks. It is a one-time but heavy cost.
Daily usage, called inference, consumes less per request but at a very large scale, across billions of requests. This cumulative usage weighs most over time.
Cooling data centers consumes water and energy. The actual footprint depends on the local electricity mix and cooling technologies.
What is established and what is not
Established: data center electricity consumption is increasing and AI contributes to it in a growing way, according to the International Energy Agency. Established: the footprint strongly depends on the electricity source — the same calculation emits much more greenhouse gas depending on the country.
Less established: precise per-request figures, often cited without reliable sources and highly variable across models. Be cautious with spectacular but unsourced comparisons.
At the individual level
A request to an assistant has a real but modest footprint compared to other daily digital and non-digital uses. The issue lies mainly at the collective and industrial scale, in infrastructure choices, energy mix, and usage sobriety.
Useful levers are not individual guilt but provider transparency on their footprint, choosing lower-carbon infrastructure, and sobriety in unnecessary uses.
Frequently Asked Questions
Does a single AI request pollute a lot?
Moderately per unit, but the global scale of usage makes the total significant.
Can AI also help the environment?
Yes, in optimizing grids, buildings, and agriculture. The assessment must balance both sides together.
Where can I find reliable figures?
Reports from the International Energy Agency and peer-reviewed academic work, being wary of unsourced numbers.
ICIA Resource
ICIA offers MentivisOS ICIA, a free lifelong platform to learn AI with a personalized learning path.
Discover MentivisOS ICIA