AI's Environmental Footprint: What Machines Consume and What They Save

The International Energy Agency estimated that data centres used about 415 terawatt-hours in 2024, roughly 1.5% of the world's electricity. It expects the figure to more than double to about 945 terawatt-hours by 2030, and data centres to account for about a tenth of the growth in global electricity demand up to that year. These are large numbers, but they do not say by themselves whether AI is an enemy of the environment or its ally.

Carbon: a popular figure that needs correcting

It is often repeated that training a single language model emits as much as five cars over their lifetimes. The figure originates in a paper by Strubell and colleagues (2019), which estimated about 626,000 pounds of carbon dioxide equivalent (about 284 tonnes) for an automated search for the best architecture of a Transformer model, against about 126,000 pounds for the average car over its whole life. Training a single model of the same type was estimated at about 192 pounds (about 87 kilograms). The gap between the searching experiment and one training run is vast, and these figures have dated as hardware and model sizes changed, so they should not be quoted as a verdict on current models.

What matters more today is what happens in use. Luccioni and colleagues (2023) found that multi-purpose generative models are orders of magnitude more costly than task-specific systems, per thousand inferences, even when model size is controlled for. Choosing a small model for a defined task is therefore an environmental decision too.

Water: a cost that does not appear on the bill

Data centres need cooling, and some evaporate water to provide it. A research paper (latest version 2025) estimated that training GPT-3 in Microsoft's US data centres could directly evaporate about 700,000 litres of fresh water, and that AI demand might require withdrawing 4.2 to 6.6 billion cubic metres of water in 2027. These are estimates from a calculation method, not measurements published by the companies, and the paper says such information has mostly been kept secret. In a dry region like the Arabian Peninsula, water becomes a constraint to be weighed before choosing a site and a cooling method.

What does the machine save?

In 2016 DeepMind announced that a machine-learning system trained on sensor data from a Google data centre had cut its cooling energy by 40%, about 15% of overall overhead energy; that is the claim of the party that ran the experiment, for one centre.

More broadly, the IEA estimates that wide adoption of existing AI-based solutions could cut energy-related emissions by about 5% of their 2035 level, provided adoption barriers are overcome. It warns of a rebound effect, in which cheap service encourages overuse that eats part of the gain, and says AI is "not a silver bullet". It also expects data centre emissions to rise from about 180 to about 300 million tonnes by 2035 in its base case, while remaining below 1.5% of energy-sector emissions.

Who bears the cost?

The cost is likely to fall where the centres are, in electricity and water, while the benefit is spread among users far away. That is an inference, not a measured result, but it raises the question of justice: who benefits and who bears the burden? The need grows for uniform, audited disclosure of energy, water and emissions, for what is not measured cannot be held to account.

What to do

1. Make the source of electricity a criterion in choosing data-centre sites and service providers.
2. Choose the smallest model that does the job rather than the largest.
3. Ask your provider for detailed disclosure of energy and water.
4. Direct computing first to what benefits people and the Earth, and review the total impact, not the efficiency of a single operation.

Waste of water and energy, even when far from our sight, is contrary to the trust with which we were made stewards. A tool may hold both benefit and harm; we take its benefit and repel its harm through good management and honest accounting.