How much energy does AI use?

Updated 2026-07-156,460 searches/mo across 4 ways of asking itRanked #36 of 519· AI and the environment
Short answer

One prompt: about 0.24 watt-hours — less than nine seconds of TV. All U.S. data centers: 176 TWh in 2023, or 4.4% of national electricity, projected at 325–580 TWh (6.7–12%) by 2028. Globally, data centres used ~415 TWh in 2024, roughly 1.5% of world electricity.

Why — the first-principles explanation

There's no single "AI energy" number because AI is not separately metered anywhere. Utilities measure data centers; data centers run AI alongside video streaming, databases, and email. So every credible figure is either a per-query measurement from one company, or a whole-data-center total with AI as an estimated share. Knowing which you're reading is the whole game.

Per query, the best public measurement is Google's, published August 2025: a median Gemini text prompt uses 0.24 Wh, emits 0.03 gCO2e, and consumes 0.26 mL of water. For scale, that's less energy than watching about nine seconds of television. Two crucial caveats. First, the number depends heavily on boundary: counting only the AI chip gives 0.10 Wh, but adding host CPU, idle reserve capacity, and data center overhead gives 0.24 Wh — the same prompt, 2.4x apart, and Google deserves credit for publishing the wider figure. Second, it's the median text prompt — image generation, video, long documents, and reasoning-heavy requests cost substantially more, and the median hides that tail.

In aggregate, Berkeley Lab's December 2024 report to Congress is the authoritative U.S. source: data centers consumed 176 TWh in 2023, about 4.4% of all U.S. electricity, up from 76 TWh (1.9%) in 2018 and 58 TWh in 2014. Its 2028 projection is a range — 325 to 580 TWh, or 6.7% to 12.0% — which at 50% utilization means 74–132 GW of continuous demand. The IEA's global picture: ~415 TWh in 2024 (~1.5% of world electricity), heading to ~945 TWh by 2030, with AI-accelerated servers growing ~30% a year and driving nearly half the increase.

The most important thing in all of this is the width of the ranges. Berkeley Lab's 2028 forecast spans a factor of 1.8. The IEA's 2035 scenarios run from ~700 TWh to over 1,700 TWh, and the agency says plainly that substantial uncertainty exists about consumption today, never mind the future. Meanwhile Google reports per-prompt energy fell 33x in twelve months. When efficiency can move 33x in a year and forecasts span 2.5x, anyone quoting AI's future energy use to two significant figures is guessing.

An example that makes it click

A single AI prompt at 0.24 watt-hours is about what an LED bulb burns in two minutes. Run a hundred prompts a day for a year and you've used roughly 9 kilowatt-hours — less than running one clothes dryer for four loads. As a personal energy decision, it doesn't register.

Now flip the telescope. Take that same trivial 0.24 watt-hours and multiply by billions of prompts a day, plus the training runs, plus the idle capacity held in reserve. You get 4.4% of the United States' entire electricity supply — more than some states use — and possibly 12% by 2028. That's the strange shape of this technology: individually invisible, collectively enormous. Like a single sheet of paper versus a forest. Nobody's page is the problem, and yet.

Key facts

Infographic: How much energy does AI use — short answer and key facts
Visual summary — How much energy does AI use?
▶ The 60-second explainer (script)

How much energy does AI use? Two answers, and they feel like they're about different technologies. One prompt: 0.24 watt-hours. That's Google's own measurement of a median Gemini text prompt, published in 2025. Less energy than nine seconds of television. About what an LED bulb uses in two minutes. And notice the boundary matters — count just the AI chip and you get 0.10 watt-hours. Add the host processor, idle backup capacity, and building overhead, and it's 0.24. Same prompt, more than double, depending where you draw the line. Now the aggregate. Berkeley Lab's report to Congress: U.S. data centers used 176 terawatt-hours in 2023 — 4.4 percent of all American electricity. In 2014 it was 58. By 2028 they project 325 to 580 terawatt-hours: six-point-seven to twelve percent. Globally, the IEA has data centres at 415 terawatt-hours in 2024, about 1.5 percent of world electricity, heading to 945 by 2030. But here's the part nobody says out loud. Look at the width of those ranges. Berkeley Lab's forecast spans nearly a factor of two. The IEA's 2035 scenarios go from 700 to over 1,700 terawatt-hours. And Google made prompts 33 times more efficient in one year. When efficiency moves that fast and forecasts are that wide, anyone quoting AI's future energy use precisely is guessing.

What authoritative sources say

U.S. Department of Energy / Lawrence Berkeley National Laboratory, 2024 Report on U.S. Data Center Energy Usegov — U.S. data centers used 176 TWh in 2023 (about 4.4% of U.S. electricity), 58 TWh in 2014, and are projected at 325–580 TWh (6.7–12%) by 2028. source ↗
Google, "Measuring the Environmental Impact of Delivering AI at Google Scale" (arXiv:2508.15734)official — A median Gemini text prompt uses 0.24 Wh comprehensively or 0.10 Wh chip-only; energy fell 33x and carbon 44x over twelve months. source ↗
International Energy Agency, Energy and AIorg — Global data centres consumed ~415 TWh in 2024 (~1.5% of global electricity), projected ~945 TWh by 2030; accelerated servers grow ~30%/year; 2035 scenarios span ~700 to over 1,700 TWh amid substantial uncertainty. source ↗
Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Reportgov — The 2028 U.S. projection implies 74–132 GW of data center power demand at 50% average capacity utilization. source ↗

People also ask

How much energy does one ChatGPT query use?

OpenAI hasn't published a peer-reviewed per-query figure. Google's comparable measurement for a median Gemini text prompt is 0.24 Wh. Older third-party estimates near 3 Wh predate major efficiency gains and use different boundaries.

Does training or inference use more energy?

Training is a huge one-time cost per model; inference is tiny per query but runs constantly. For widely deployed models, cumulative inference energy typically overtakes training within the model's service life.

Why is there no official 'AI energy' statistic?

Because AI isn't metered separately. Utilities see data centers, and data centers run AI mixed with streaming, storage, and everything else. Every AI-specific number is an estimate of a share.

Is AI energy use per prompt going up or down?

Sharply down. Google reports 33x improvement in twelve months from better models, chips, and serving software. Total demand still rises because usage grows faster than efficiency.

The same question, asked other ways

This page answers all of these. Their searches are counted together in the ranking — one question, 4 phrasings. How we rank →

Related questions