You may have seen an energy-efficiency label when shopping for a refrigerator or air conditioner. New appliances can often do the same job with less electricity than older ones. The computers that run AI are improving in the same way, and the electricity used for one calculation is falling quickly.
Yet the total amount of electricity needed for AI is growing. That is because we must consider not only how quickly each task becomes more efficient, but also the number of AI users, the number of tasks, and how demanding those tasks are. If a job uses less electricity but is performed far more often, the total can still rise.
This is not only a question about monthly bills or power stations. When a large AI facility is built nearby, demand can become concentrated on the transmission lines and substations in that area. This article separates the electricity used by one task from the electricity used across society.
One device can become efficient while the total rises
Imagine replacing a light bulb with one that uses half as much electricity. One bulb now uses half as much. But if a city installs three times as many bulbs, the city’s total use will be higher than before. The same thing can happen with AI.
The International Energy Agency is an international organization that studies energy around the world. It is known as the IEA. According to its 2026 report, improvements in software and semiconductors have recently reduced the energy needed for one AI task extraordinarily quickly—by at least an order of magnitude, meaning to one-tenth or less, each year.
A short text question now typically uses less electricity than running a television for the same short period. The total can still rise because people are asking AI to do more work, and the work itself is changing.
AI requests do not all have the same weight
Creating one short passage of text and creating a long video require very different amounts of computation. The calculation an AI performs to create an answer is called inference. In simple terms, the system repeatedly calculates what to produce next based on the words or images it receives.
AI is also increasingly used to continue through several jobs after receiving one instruction: searching, writing, using tools, and checking results. A system that works through several stages in this way is called an AI agent. A user may press a button once while many calculations happen behind the scenes.
The IEA says that jobs such as video generation, complex reasoning, and AI-agent work can sometimes use hundreds or thousands of times more energy than producing a short piece of text. A single number for “the electricity used by one AI request” treats a simple question and a long video as if they were the same.
Greater efficiency can also make each use cheaper. Uses that were previously too expensive then spread. A person may compare ten drafts instead of making one, create several videos instead of one still image, or give daily AI access to many employees instead of a small group. If the room created by efficiency is spent on new uses, total electricity does not fall.
Small worldwide can still be large in one region
AI computation usually takes place in a data center. A data center is a large facility that contains many high-performance computers for calculation and data storage. It also needs electricity for cooling and for equipment that protects it during outages.
The IEA projects that electricity used by the world’s data centers will rise from 485 terawatt-hours in 2025 to about 950 terawatt-hours in 2030—almost twice as much. In 2024, data centers were estimated to use about 1.5% of the world’s electricity.
That global percentage may not sound very large. But data centers are not spread evenly around the world. They tend to gather in areas with fast network connections, enough land, and access to large amounts of electricity.
It is like having enough water across a country but being unable to send a large amount at once through one town’s narrow pipes. Electricity must also be carried after it is generated. Transmission lines move it over distance. Transformers and substations change the voltage to suit the next part of the system. Together, this equipment forms the power grid.
Data centers can be built faster than grids
The IEA’s Electricity 2026 report says that planning, permitting, and completing grid infrastructure can take 5 to 15 years. A data center, by comparison, may be built in 1 to 3 years. The building that uses electricity can be finished before the equipment that delivers it.
More than 2,500 gigawatts of projects around the world are waiting to connect to grids. These projects include power stations and batteries, so they are not all AI or data centers. The figure still shows that grid work is failing to keep up with many planned projects.
The equipment itself cannot always be obtained quickly either. The U.S. Department of Energy reports component shortages and long delivery times for transformers. Even after a new power station is built, its electricity cannot reach the right place without transmission lines and transformers.
“Electricity per AI task” is not enough
It is important to measure the electricity used by one AI task. That lets us compare which machines and methods are more efficient. It is not enough, however, to understand the effect on society.
We need to separate electricity per task, the number of tasks, and the type of task. We also need total electricity use for a data center. Its location matters, as does how much it uses at times when electricity is scarce. The IEA says that companies providing AI need to disclose energy use in a more consistent way so that projections can become more accurate.
There is no single solution. More efficient chips and software can help. An AI model should not be larger than the job requires. Work that is not urgent can move to a time or place with spare electricity. Cooling can improve. Existing transmission lines can be used more effectively within safe limits, and expanded where necessary. Batteries can soften sudden changes.
The figure to watch is not only the improvement in one AI task. Is use growing faster than efficiency? Is the share of demanding tasks increasing? Which regions are carrying the burden? Looking at all three reveals why the total electricity used by AI can keep rising.