AI & sustainability

The problem is not the single question

A simple question has become cheap. But we don’t ask one question, we ask billions. And we no longer just let AI complete a sentence. We give it documents, let it research, write code, generate images, produce videos and have agents carry out several work steps on their own. That is exactly what changes the calculation.

Epoch AI puts a typical text request at around 0.3 Wh. With very long inputs this can become 2.5 Wh, and with extremely large contexts even around 40 Wh. “An AI request uses X” is therefore about as meaningful as “A car journey uses X”. It depends on where you are driving.

Source: Epoch AI – Energy use of different AI requests

AI is getting more efficient. And still uses more and more.

That sounds contradictory at first, but it isn’t. The models are getting more efficient. The chips are getting more efficient. The data centres are getting more efficient. Google, for example, reports that the median energy demand of a Gemini text request fell by a factor of 33 within a year. This is a manufacturer’s figure and not independently verified, but the direction is fairly clear: AI is becoming considerably more efficient per computing operation.

But at the same time we use it more and more often. In 2025, data centres worldwide used roughly 485 terawatt-hours of electricity, according to International Energy Agency calculations. By 2030 this could reach around 950 terawatt-hours, almost a doubling in five years. The electricity use of AI-focused data centres in particular is growing even faster and could triple over the same period.

So efficiency doesn’t solve the problem automatically. It can even become part of the problem, because whatever gets cheaper and more efficient, we generally use more often.

Source: IEA – Key Questions on Energy and AI 2026

A data centre is now a pretty big power socket

A typical AI data centre can now need as much electricity as roughly 100,000 households. And the largest facilities currently being built can, according to the International Energy Agency, reach twenty times that again.

Beyond the absolute figure, what stands out is the concentration. A new AI data centre isn’t “somewhere on the internet”. It stands in a specific place. There it needs power lines, transformers, cooling and, where applicable, water. The IEA already warns that grid bottlenecks could delay a considerable share of planned data centre projects.

The cloud doesn’t float. It stands in a hall.

Source: IEA – Energy and AI

And then there is the water

Computers, of course, don’t drink water. Data centres do. Water is used for cooling, among other things, and how much is actually needed depends heavily on location, outside temperature, cooling technology and power generation. So here too, it pays to be careful with spectacular average figures.

For the median Gemini text request, for example, Google states a direct water use of 0.26 millilitres, roughly five drops. A widely cited scientific study, on the other hand, estimated the direct water use for training GPT-3 in US data centres at around 700,000 litres. Both figures can be right at the same time. They just describe completely different things.

A single request is something other than training a large model. And a data centre in Finland has different cooling conditions from one in Arizona. That is why “How much water does AI use?” is really the wrong question already. The better one is: which AI runs where, on which infrastructure and with which energy and cooling concept?

Sources: Google – Energy, CO₂ and water use of Gemini, Study “Making AI Less Thirsty”

552 tonnes of CO₂ for one AI model?

That figure is indeed correct, but only with an important caveat. For training GPT-3, an electricity use of around 1,287 MWh and a resulting output of roughly 552 tonnes of CO₂e were calculated. That was in 2020.

The figure therefore shows less “what AI costs today” and more something that is almost more interesting: how strongly location, hardware and electricity mix decide a model’s carbon footprint. For the similarly large open language model BLOOM, for example, only about 30 tonnes of CO₂e were calculated for the training itself. If you include data centre infrastructure and hardware, the footprint rises accordingly.

Besides the size of a model, where and how we compute matters, too.

Source: ARCEP – Generative AI: Environmental Issues and Challenges, 2026

So better not to use AI at all?

That would be too easy. AI uses energy, but it can also save energy. It can optimise power grids, control buildings more efficiently, improve industrial processes, analyse traffic flows or detect waste of material and energy. The International Energy Agency therefore explicitly examines both sides: the additional energy demand from AI and the potential of AI to reduce energy use elsewhere.

The decisive question is therefore not whether AI uses energy. Of course it does. The more interesting question is whether what we do with that energy is worth it.

Sustainable AI doesn’t start with a guilty conscience

Nobody has to wonder before every prompt whether a glass of water is evaporating somewhere. But we should start treating AI like any other technical infrastructure. Do we really need the biggest model for this task? Does every process have to use generative AI? Do ten agents have to research something that a database query could answer? Could a smaller local model do the same job? And does our AI run where energy can be provided as efficiently and climate-friendly as possible?

These are not arguments against AI. They are arguments for sensible computing.

That is the real sustainability question

For decades we built software and mainly asked: can the computer do this? With AI a second question comes in: does it really have to do this every single time? Because a single AI request is now astonishingly efficient. A few billion of them are not.