AI queries use less energy. Data centres still want more power
"How much energy does an AI query use?" sounds like a clean question. It is becoming a poor way to judge AI's energy footprint.
The International Energy Agency's 2026 report, Key Questions on Energy and AI, says better software and hardware have cut energy use per AI task by at least an order of magnitude annually in recent years. A simple text query now typically uses less electricity than a television running for the same amount of time.
Global data-centre electricity use still grew 17% in 2025. Use at AI-focused data centres grew 50%. Reasoning, video generation, and agentic tasks can consume hundreds or thousands of times more electricity per query than simple text generation, and the number of users is climbing.
A cheaper unit of computation can sit inside a much larger system. Judging that system requires the task, its frequency, the site where it runs, and the people asked to fund the grid around that site.
Cheaper computation invites more computation
When computation gets cheaper, developers tend to use more of it and put it in more products. A model can answer a short question efficiently, then spend much more on a long reasoning trace, repeated tool calls, generated video, or a workflow that continues after the user leaves.
The IEA estimates that replacing every conventional internet search with a simple AI text query would use less than 4 terawatt-hours a year, below 1% of current data-centre consumption. That comparison avoids viral estimates tied to an unnamed model or old hardware. It also has a narrow boundary. Future demand will not be limited to search with a chat box attached.
Coding agents inspect repositories and rerun tests. Browser agents load pages and call tools. Media systems generate several candidates before a user keeps one. One visible request may trigger dozens of model calls, plus storage, networking, and other computation. Counting prompts misses most of that work.
The grid never sees a prompt
Power systems see large facilities requesting firm capacity in specific places.
The IEA's Energy and AI report estimated that data centres used 415 TWh in 2024, about 1.5% of global electricity. Its base case reaches roughly 945 TWh in 2030. The global percentage is a poor guide to local strain. Nearly half of US data-centre capacity was in five regional clusters, and the agency estimated that about 20% of planned projects could be delayed if grid constraints are not addressed.
Those constraints arrive as bills and planning decisions. A utility must decide how to pay for generation, transmission, transformers, and reserves for a data centre that may install servers gradually or reserve more connection capacity than it uses at first. Residents and other businesses do not experience this as a global percentage.
The Lawrence Berkeley National Laboratory's 2025 update, published in June 2026, puts US data centres at 649 TWh in 2030 in its reference case. Its broader range is 521 to 843 TWh, equal to 9.5% to 15.3% of US electricity use. The width is informative. Results change with assumptions about hardware shipments, AI chip lifetimes, server utilization, idle power, and cooling.
A forecast is a planning model, not a meter reading. Its assumptions belong beside its headline.
Measure the workload, then limit it
Product teams do not need a perfect watt-hour figure for every response to make better choices. They do need to know what they are running.
Separate short text from long-context reasoning, image or video generation, and agents with tool loops. Token counts will not capture every energy cost, but they say more than the number of times somebody pressed Enter.
Agent steps, retries, wall-clock time, and generated media also need limits. A runaway job is already a reliability and cost failure. The energy waste is another reason to stop it.
Use the smallest system that clears the quality bar. Test routing on representative work instead of making the newest or largest model the default. A gateway such as api.ish.chat can help compare model use by task, though the team still has to define the policy and acceptance tests.
Keep location and time when a provider exposes them. A kilowatt-hour used during a constrained local peak has a different effect from one used when the grid has spare low-carbon generation. Most provider dashboards do not make that comparison easy.
Public reporting belongs at the facility level
The European Commission's data-centre energy performance page describes a database for the energy performance and water footprint of data centres with significant consumption. EU rules define common indicators. The Commission is also preparing a rating scheme and work on minimum performance standards.
Facility reporting cannot calculate the exact cost of one agent run. It can show whether improving equipment efficiency coexists with rising electricity and water demand. Regulators also gain common fields for comparing sites instead of leaving each company to choose its own flattering metric.
This separation matters for sovereign AI too. Model access, compute location, electricity supply, and legal control affect each other, but one does not guarantee the others. An energy claim only means something when the boundary is stated.
One number cannot carry the argument
Individual AI tasks are getting much more efficient. Total electricity demand from AI-focused infrastructure is rising quickly. Treating either observation as a rebuttal to the other produces bad product choices and bad public policy.
Product teams can track workload type, model use, retries, and achieved quality. Providers can publish comparable workload and facility data. Utilities and governments can measure location, timing, water demand, generation mix, and who bears expansion costs.
A claim about one efficient AI answer says little about a million users, a reasoning loop, a video generator, or a new grid connection. The query can be efficient while the system around it keeps growing.
