How Much Electricity Does AI Use? The Numbers Explained

"How much electricity does AI use?" sounds like it should have a single tidy answer. It does not, because the figure depends on whether you mean one prompt, one model or the entire industry, and because providers publish very little hard data. What we do have are credible estimates from the IEA, peer-reviewed research and company sustainability reports. Put together, they tell a clear story: AI's electricity demand is large, growing fast, and worth understanding.

The big number: ~200 TWh a year

At the industry level, the IEA estimates that AI-related computing already draws on the order of 200 TWh of electricity a year. To put that in perspective, that is comparable to the total annual electricity consumption of a mid-sized country. It covers everything: training new models, running inference for everyday prompts, and the substantial overhead of the data centers that house all of it.

Treat 200 TWh as an estimate, not a precise meter reading. Different studies draw the boundaries of "AI" differently, and the number is a moving target that keeps climbing. But it is the right order of magnitude, and it explains why utilities and governments are suddenly paying attention to data-center demand.

Energy per query

Zoom all the way in and a single text exchange is often cited at a few watt-hours, several times more than a plain web search. That sounds trivial, and per prompt it is. But three things push the total up:

• Model size. Larger frontier models do more computation per token, so they cost more energy per answer.
• Output length. Generating a long, detailed response costs more than a one-line reply.
• Media. Image generation and video are dramatically more energy-intensive than text.

Because providers rarely disclose per-query energy, these figures are estimates with wide ranges. You can explore how prompt length, model and region combine using our AI carbon footprint calculator.

Why data centers are growing

Almost all of AI's electricity is consumed in data centers packed with specialized chips. As demand for AI has surged, operators have raced to build more of them, and each new facility adds steady, around-the-clock load to the local grid. Unlike a home that peaks in the evening, a data center runs flat out all day, which makes its demand easy to forecast but hard to ignore.

This growth is also why water enters the conversation: cooling all those chips consumes large volumes of water, often on the order of millions of litres a day for a big facility. We cover that side in our guide to how much water ChatGPT uses.

PUE and hidden overhead

Not all the electricity a data center draws reaches the chips. Cooling systems, power conversion and lighting all consume energy too. The standard measure of this is Power Usage Effectiveness (PUE): the ratio of total facility energy to the energy that actually reaches the computing hardware.

From kilowatt-hours to carbon

Electricity is only half the story. The same kilowatt-hour can be nearly carbon-free or heavily polluting depending on how it was generated. On a grid near the US average of about 0.395 kg CO2 per kWh, a few watt-hours per query becomes a few grams of CO2. On a coal-heavy grid that figure can more than double; on a renewable-heavy grid it can fall close to zero.

This is why location and timing matter so much. Running the same workload in a region with clean power, or at a time of day when renewables dominate, can cut its carbon footprint dramatically without changing a single line of the model. Water follows electricity too: roughly 3.4 litres are consumed per kWh once generation and cooling are combined.

Where this is heading

The trajectory points up. As models grow and adoption spreads, projections suggest AI could be responsible for around 300 million tonnes of CO2 by 2030, with electricity demand rising in step. The exact path depends on a tug-of-war between two forces: hardware keeps getting more efficient per computation, but we keep asking it to do far more.

The honest takeaway: AI's electricity use is large and growing, but it is also highly variable and increasingly measurable. What you cannot measure, you cannot manage.

That is the gap Ecoia exists to close. We run the same frontier models, measure the energy, carbon and water of every request in real time, and retire verified offsets for more than 200% of that footprint. You can read the full methodology on our how it works page, or see the bigger picture in is ChatGPT bad for the environment?

FAQ