How much electricity does AI use? From one prompt to the whole U.S. grid

How much electricity does one AI prompt use?

For a short text prompt, about 0.3 watt-hours. A watt-hour is the energy a one-watt device uses in an hour. The estimates come from different methods, and they agree more than they differ.

Google's 0.24 watt-hours splits into 58% for the AI chips, 25% for host processors and memory, 10% for idle machines and 8% for data center overhead, with rounding. Counting only the active chips, in the busiest and most efficient 10% of data centers, would give 0.10. Google says that is closer to the narrower method used in earlier studies. Its active chips across the whole fleet use 0.14. Google also says the energy per prompt fell 33 times in the 12 months to May 2025, which is its own claim.

Two cautions apply. A longer output uses more energy, so the figure for a reasoning model or a long document is higher than for a short answer. And water figures are not comparable: Google's counts on-site cooling, and Mistral's is a lifecycle figure.

Why is the total large if each prompt is small?

Because of the number of prompts. At Google's 0.24 watt-hours, one gigawatt running for a year, which is 8.76 terawatt-hours, would cover about 36.5 trillion prompts, by Spezia's arithmetic. That ignores training, which none of these figures include.

Microsoft's researchers calculate that serving a billion queries a day needs about 0.7 gigawatt-hours a day, or about 1.7 if a tenth of the queries are long. That total is larger than the per-query medians because it adds a 31% allowance for uneven demand, spare machines and networking, and averages over a wide spread of query lengths. The same paper says widely cited estimates are overstated by four to twenty times.

How much electricity do U.S. data centers use?

Lawrence Berkeley National Laboratory publishes a widely used U.S. estimate. It is built from equipment shipments, power per device and cooling simulations, not from meters, and it leaves out cryptocurrency mining.

The two reports use different data, so do not subtract one row from the other. On the June 2026 basis, 2023 was about 168 TWh, by Spezia's arithmetic, because the update reports 14% growth from 2023 to 2024.

The December 2024 report projected a range of 325 to 580 TWh in 2028, or 6.7% to 12.0% of U.S. electricity. The 2026 reference case for 2028 sits in the middle of that range. In power terms, the December 2024 report's range for 2028 corresponds to 74 to 132 gigawatts, at an assumed 50% utilization. The June 2026 update puts 2030 at 148 gigawatts on the same assumption.

The June 2026 update says data centers would be responsible for about a third of load growth from 2024 to 2030. It revised earlier years slightly down, mainly because fewer GPUs shipped in 2023 and 2024 than first reported. GPU-accelerated servers, the AI machines, grew from under 2 TWh in 2017 to more than 40 TWh in 2023, counting server electricity only.

And across the world?

The International Energy Agency counts 415 TWh for data centers in 2024, about 1.5% of global electricity, with 45% of it in the United States. Its 2026 update says demand grew 17% in 2025 to 485 TWh, and projects about 950 TWh in 2030, roughly 3% of global demand. It says AI-focused data centers grew 50% in 2025 and would triple over 2025 to 2030. A typical AI-focused data center uses as much electricity as 100,000 households, and the largest under construction use about 20 times that.

What does a gigawatt mean?

A gigawatt is a rate of power, and a gigawatt-year is an amount of energy. One gigawatt running flat out for a year is 8.76 TWh. Dividing by the Energy Information Administration's average household use gives about 810,000 to 840,000 homes, using 10,791 kilowatt-hours a year for 2022 and 865 a month for 2024.

Spezia's tracker, as of November 9, 2026, holds 294 U.S. projects that are not yet operating and report a capacity. Together they plan 285 GW. Some are only announced, and some are blocked or paused. If every one of them ran flat out for a year, it would use about 2,497 TWh, against about 4,200 TWh of electricity the United States consumed in 2025. That is an illustration of scale, not a forecast. Capacity here is planned full build-out, most projects will not be built on their stated schedule, and none run flat out. See the U.S. AI data center pipeline and the gigawatt power wall.

How do you read an AI energy headline?

Six checks help.

1. Per prompt, per task or in total? A watt-hour, a training run and a national total are very different scales.
2. What is inside the boundary? Chips only, whole servers, idle machines, cooling and training give different numbers.
3. Average or median? A few long outputs can lift an average.
4. Measured or modeled? Google measured its fleet. Epoch and Microsoft modeled.
5. Power or energy? Gigawatts measure capacity. Terawatt-hours measure use over time.
6. Which year? Efficiency changed quickly, so a 2023 figure may not describe 2026.