English · By MundoGood · October 7, 2026
AI-generated conceptual illustration showing computing and facility infrastructure; it does not depict an actual data center.
A proposed data center, a training run and a chatbot response can all appear in an AI energy headline. Each describes a different scale of activity. To understand the number, first ask what equipment is included, over what period, and which unit is being used.
The distinction matters because the sector is changing quickly. The International Energy Agency’s 2026 update estimates global data center electricity consumption at 485 terawatt-hours in 2025 and projects about 950 TWh in 2030. Those figures cover data centers broadly, including workloads beyond AI, and the future figure is a projection. IEA: Key Questions on Energy and AI.
Training and inference describe different work
Training adjusts a model’s parameters using data and an optimization process. Inference uses the trained model to produce an output from an input. Fine-tuning is another training step; generating a reply is typically inference. IBM’s technical explainer describes the distinction and the different deployment arrangements.
For energy accounting, the useful boundary is the workload being measured. A reported training figure might cover one successful run, or a broader development process with experiments and evaluations. A serving figure might cover a batch of requests or continuous operation over a month. The label alone does not settle that boundary.
Repeated use can accumulate substantial electricity consumption even when an individual response is efficient. Conversely, a demanding training run does not tell you how much electricity every subsequent response will require. Ask for both the measured task and the accounting period.
Megawatts measure a rate; kilowatt-hours measure an amount
A watt describes power at a moment. A watt-hour describes energy accumulated over time. The same distinction holds for kilowatts and kilowatt-hours, or megawatts and megawatt-hours. U.S. Energy Information Administration: measuring electricity.
Consider an illustrative facility drawing a steady 10 MW for 24 hours. It would consume 240 MWh during that day. Real facilities have changing loads, so an accurate total requires measurements over the period.
An announcement describing a “100 MW data center” therefore needs a second question: does that mean a proposed grid connection, a design limit, installed IT capacity or measured operating demand? Multiplying the headline capacity by every hour of the year silently assumes continuous operation at that level. It can substantially change the meaning of a comparison.
PUE describes facility overhead
Power usage effectiveness, or PUE, compares total facility energy with IT equipment energy over the same period. Cooling and power-distribution losses contribute to the facility total. The U.S. Department of Energy’s design guide treats PUE as one performance measure alongside other efficiency and sustainability metrics.
For an illustrative PUE of 1.2, every 100 units of IT energy correspond to 120 units at the facility boundary. The additional 20 units are overhead. That is different from saying overhead is 20% of the total; in this example it is about 16.7%.
A low PUE indicates comparatively little facility overhead. It does not reveal how much useful work the servers deliver. Two facilities could have the same PUE while running very different workloads, and a larger efficient facility could still consume more electricity in total.
A per-prompt number needs a measurement label
Google researchers’ 2025 inference study reported 0.24 Wh for the median Gemini Apps text prompt in their measurement setting. Their method included accelerator use, host systems, idle capacity and data center overhead. It is a provider-authored research result with a defined scope, rather than a universal allowance for every AI request.
The median describes the middle of a distribution. It does not show the average or the most demanding requests. Prompt length, output length, model choice and task type can also make comparisons misleading. A short text reply and a long video-generation task should not inherit the same energy estimate.
Before multiplying a published number by millions of users, check whether the statistic supports that calculation. In particular, multiplying the median by the request count does not reliably establish total consumption when requests vary widely.
Use five checks before sharing an energy claim
- Boundary: chip, server, whole facility or a wider lifecycle assessment?
- Activity: training, inference or all data center workloads?
- Unit and period: power capacity, measured electricity, or a modeled annual total?
- Statistic: mean, median, peak or a selected example?
- Status: observed result, provider estimate or future projection?
For a local development proposal, also look for the expected operating load, connection timetable and arrangements for paying for infrastructure. The IEA notes that local price effects depend on system conditions and policy choices. A worldwide electricity percentage cannot answer those local questions by itself.
The strongest comparison uses matching boundaries and useful outcomes. That makes it possible to recognize real efficiency improvements while still asking whether expanding usage is increasing total demand.
Sources and review
Checked October 7, 2026. Sources: IEA’s 2026 update, EIA’s electricity definitions, DOE’s 2024 design guide, IBM’s training/inference explanation and Google researchers’ 2025 preprint. Numerical facility and PUE examples are illustrative calculations, not measurements of an operating site.
Editorial disclosure: AI assisted with research and writing. The linked sources were checked for this article.
Sources checked October 7, 2026. AI-assisted research and writing.