How Much Energy Does AI Really Use?

AI has no single electricity cost per question. A short text response, a long reasoning task and a generated video require different amounts of computation. To understand AI energy consumption, it helps to separate the work done for one request from the electricity used by the infrastructure serving millions of people.

The distinction matters for climate reporting. A more efficient model can use less electricity for each task while the industry’s total demand continues to grow.

What the latest evidence tells us

The International Energy Agency’s April 2026 report estimates that electricity demand from data centres worldwide grew by 17% in 2025, while consumption at AI-focused facilities rose by 50%. Those figures describe different groups: the first includes data centres supporting services beyond AI.

The same report describes substantial improvements in energy efficiency per task. It also finds that video generation, reasoning and agentic applications can require much more energy than simple text generation. Adoption, efficiency and the kinds of tasks people choose all affect the total. Source: IEA, Key Questions on Energy and AI (2026).

Training, inference and the building around them

Training is the process of adjusting a model using data. Inference means running a trained model to produce an output. A service can incur training costs before launch and inference costs whenever people use it; later model updates add further work.

The processors are only part of the infrastructure. Storage, networking, cooling and power systems also need electricity. The IEA describes substantial differences in the equipment mix and cooling requirements of different facilities. A processor-only measurement therefore cannot automatically be treated as the electricity consumption of the whole site. Source: IEA, Energy demand from AI.

Why a single per-query number can mislead

Consider two requests: “Translate this sentence” and “Analyse these documents, check the calculations and write a detailed report.” Both look like one interaction in a chat interface. They are not equivalent units of work.

Research by Oviedo and colleagues models inference energy using processing throughput, hardware power and infrastructure overhead. Its estimates change with deployment assumptions and longer computation at response time. These are modelled results for specified conditions, not a universal meter reading for every chatbot. Source: Energy Use of AI Inference, Efficiency Pathways, and Test-Time Scaling, revised June 2026.

Before comparing published estimates, check:

• The task: text, image, video, or a workflow involving multiple steps?
• The workload: how much input was processed and how much output produced?
• The measurement: a direct observation, an estimate, a median, or an average?
• The boundary: processors alone, the complete server, or the facility including cooling?
• The date and system: which hardware, model version and operating conditions?

Efficiency and total demand answer different questions

A simple hypothetical example shows why both deserve attention. Suppose a service cuts electricity per task in half, while its number of otherwise identical tasks triples. Its total task electricity becomes 1.5 times the starting level—a 50% increase.

That is arithmetic, not an estimate of any AI provider. It illustrates why an efficiency announcement needs a second question: how has total consumption changed over the same period?

It also helps to keep units straight. Megawatts (MW) describe power at a point in time or rated capacity. Megawatt-hours (MWh) describe energy over a period. A proposed facility’s capacity does not tell us its annual electricity use without assumptions about utilisation and operating hours.

Electricity use is not the same as carbon emissions

The climate effect also depends on the electricity supply. A carbon estimate needs an emissions factor as well as an energy measurement, with compatible locations, dates and accounting boundaries.

The IEA’s energy-supply analysis distinguishes the generation physically supplying data centres from operators’ contractual electricity claims. Its 2025 scenarios show different regional supply mixes and continued near-term contributions from fossil fuels alongside expanding low-emissions sources. These are scenarios, not guaranteed outcomes. Source: IEA, Energy supply for AI.

For readers following our coverage of ASEAN’s data-centre expansion, the useful questions are local: what new load is being connected, what generation will serve it, and what evidence supports the operator’s climate claims?

What users and organisations can ask for

Our practical recommendation is to judge AI services by the useful work they deliver and the evidence they disclose. For routine tasks, choose a suitable tool rather than assuming a more elaborate workflow is always better. For large deployments, request measurements for representative workloads.

Ask suppliers to publish total electricity consumption alongside efficiency metrics, explain their measurement boundaries, and distinguish observed results from forecasts. An energy-saving claim becomes easier to assess when the service quality and comparison baseline are clear.

For more context, explore the evidence behind climate change and our guide to practical climate action.

The takeaway: evaluating AI’s energy footprint requires the task, the scale of use and the source of electricity. One headline number cannot describe all three.