How Much Energy Does AI Use? Data Centres, Power and Cooling Explained

Why it matters now

On 30 October lawyers for the NAACP announced that a US federal judge had refused xAI’s request to throw out a Clean Air Act lawsuit over the gas turbines that power its Colossus 2 data centre in Mississippi. The case is an early test of a shortcut AI companies are taking: building their own power plants because the grid cannot connect them quickly enough. Our brief has the details. Days later, OpenAI went the other way and promised to cut up to 1GW of demand at its planned Georgia campus when the grid is strained.

The bigger picture comes from the International Energy Agency (IEA). It says electricity demand from data centres rose 17% in 2025, and expects consumption to roughly double from 485 terawatt-hours (TWh) in 2025 to 950 TWh in 2030, around 3% of the world’s electricity. A terawatt-hour is a billion kilowatt-hours, the units on a household bill.

So “how much energy does AI use?” is no longer a niche question. It shapes power bills, local air, climate targets and where the next data centre goes.

How it works, step by step

Step 1: The chips do the work

When you ask a chatbot something, your question travels to a data centre, a warehouse-sized building packed with racks of servers. Inside them, AI chips such as GPUs do billions of multiplications to work out each token of the answer. Every one of those calculations uses a little electricity. (Our AI chips explainer covers what the chips are.)

There are two big jobs. Training builds a model once, running thousands of chips for weeks or months. Inference is every time someone uses the model afterwards. Training is a huge one-off bill; inference is a small bill paid billions of times.

Step 2: The electricity becomes heat

Almost all the electricity a chip uses ends up as heat. Pack enough chips together and the heat is enormous: the IEA says AI data centres need to remove the heat equivalent of “30 natural gas boilers per server rack”.

Step 3: Cooling takes it away

An everyday analogy: think of a busy kitchen with every oven on. The cooking is the useful work, but you also need extractor fans, open windows and maybe air conditioning, all of which use power too.

Data centres do the same. Older halls blow cold air through the racks. Newer AI racks often pipe liquid right up to the chips, because water carries heat away far better than air. Some sites also evaporate water to cool it, which is why AI water use is in the news alongside energy.

How much extra power the building uses beyond the computers is measured by power usage effectiveness, or PUE: the total power drawn by the building divided by the power used by the computing equipment. A perfect score is 1.0. Google said in August 2025 that its data centre fleet runs at an average PUE of 1.09, meaning about 9% extra on top of the computers.

Step 4: The grid delivers the power

A large AI campus needs a great deal of power delivered to one spot. Getting it means new power lines, substations and transformers, and permission from grid operators, all of which can take years. The IEA names gas turbines, transformers, grid connections and planning approvals as bottlenecks.

Step 5: Some build their own power

To move faster, some developers put generators on site. The IEA says many on-site natural gas projects are in the United States, and estimates around 15 to 27 gigawatts of on-site gas could power data centres by 2030. Others are signing deals for wind, solar, nuclear and geothermal power. That speed comes with trade-offs, which is exactly what the xAI lawsuit is about.

A worked example: your day of chatting

In August 2025 Google published a measurement of its Gemini apps: the median text prompt used 0.24 watt-hours of energy and 0.26 millilitres of water, “about five drops”. Google compared the energy to “watching TV for less than nine seconds”.

Say you ask 20 questions a day. Our arithmetic: 20 × 0.24 Wh = 4.8 Wh. That is about the same as leaving a 10-watt LED bulb on for half an hour. Small.

Now scale up. The US Lawrence Berkeley National Laboratory found that American data centres used 176 TWh in 2023, about 4.4% of all US electricity, and could reach between 6.7% and 12% by 2028. Not all of that is AI, but it shows the scale. Tiny sips, taken by billions of people and machines, become a river.

And not every request is a text prompt. Generating video, or running an AI agent that works for an hour, can use far more than a quick question. The IEA notes that energy-intensive new uses such as agents are pushing total consumption up even as each task gets more efficient.

What people get wrong

“Every chatbot answer wastes a bottle of water.” Published measurements for a typical text prompt are far smaller, a few drops in Google’s case. But figures vary by model, company, location and task, and few companies publish them, so treat any single number with care.

“AI uses most of the world’s electricity.” It does not. Even the IEA’s 2030 projection is around 3% of global demand. The real pressure is local: data centres cluster in a few places, so one region’s grid can be stretched hard while the global share stays small.

“Efficiency will solve it.” Efficiency is improving fast. The IEA says that “measured per individual task, the energy efficiency of AI is improving at a rate unprecedented in energy history”, and Google said the energy of its median prompt fell 33 times in a year. But cheaper answers invite more use, so total demand still rises.

“Training is the big cost.” Training one frontier model is a large one-off. Inference happens every second of every day for as long as the model is in use, which is why it now gets as much attention.

Where it is heading

• Roughly double by 2030. The IEA expects data centre electricity use to grow from 485 TWh in 2025 to 950 TWh in 2030, and power use from AI-focused data centres to triple.
• New kinds of power. The IEA says tech companies are backing nuclear and geothermal projects, and that the pipeline of small modular reactors linked to them grew from 25 to 45 gigawatts.
• More batteries. The IEA says battery storage is becoming important for handling the rapid swings in demand that AI workloads create.
• More rules and court cases. The xAI lawsuit shows that local air permits, planning law and grid rules are becoming part of the AI story.

Key terms

• Data centre
• Power usage effectiveness (PUE)
• Inference
• GPU
• Token
• HBM
• Terawatt-hour (TWh)

Sources

• International Energy Agency, “Key Questions on Energy and AI”, executive summary, 2026. iea.org
• International Energy Agency, “Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions” (news release). iea.org
• Lawrence Berkeley National Laboratory, “2024 United States Data Center Energy Usage Report” (summary), December 2024. energyanalysis.lbl.gov
• Google Cloud, “Measuring the environmental impact of AI inference”, 21 August 2025. cloud.google.com; technical paper arXiv:2508.15734
• Southern Environmental Law Center, on the NAACP v. xAI ruling, 30 October 2026, as reported in our brief.