AI and data-centre energy consumption

Energy · 9 min read · Updated 2026-09-02

Almost every widely-quoted figure for AI energy use is a projection, an estimate or a single-vendor disclosure generalised beyond its basis. Understanding what is measured and what is modelled is more useful than any number.

Key takeaways at a glance

Topic	Key point
Why this page has no headline number	The figure you are looking for — how much electricity AI uses — does not exist in a reliable, comparable form, and pages that give you one are usually
Four boundary problems that make figures incomparable	Problem Effect on the number Training versus inference Training is a large one-off; inference is small per event and continuous.
Why per-query figures are the least trustworthy of all	The energy of a single model query depends on model size, quantisation, hardware generation, batching, context length, output length, cache behaviour
Questions that produce honest answers	Is this measured, modelled or projected — and by whom?
What is not in dispute	The direction is clear even where the magnitudes are not: compute demand is growing, and it is growing in specific places rather than evenly — which i

Why this page has no headline number

The figure you are looking for — how much electricity AI uses — does not exist in a reliable, comparable form, and pages that give you one are usually repeating a projection without its assumptions.

That is not evasion. It is the actual state of the evidence, and knowing why is more useful than memorising a number that will be revised.

Four boundary problems that make figures incomparable

Problem	Effect on the number
Training versus inference	Training is a large one-off; inference is small per event and continuous. Totals combining them without saying so are unreadable
Facility versus IT boundary	Does the figure include cooling and power conversion, or only the servers? This alone moves totals substantially
Embodied versus operational	Manufacturing the hardware carries energy and emissions that operational figures exclude
Attribution inside a shared facility	A data centre runs many workloads; splitting consumption between AI and everything else requires assumptions that are rarely published

Why per-query figures are the least trustworthy of all

The energy of a single model query depends on model size, quantisation, hardware generation, batching, context length, output length, cache behaviour and how loaded the cluster is. Every one of those varies by an order of magnitude across real deployments.

A per-query figure without those parameters stated is not a measurement, it is an anecdote. Treat any comparison of a query to a household activity as illustration rather than evidence — the comparison usually crosses boundaries the two figures did not share.

Questions that produce honest answers

1. Is this measured, modelled or projected — and by whom?
2. Does it cover training, inference or both?
3. Is the boundary IT load or whole facility?
4. What year is it for, and is it a forecast?
5. How was AI separated from other workloads in a shared facility?
6. Is the underlying data published, or only the conclusion?

A source that answers all six is worth citing. Most cannot answer the fifth.

What is not in dispute

The direction is clear even where the magnitudes are not: compute demand is growing, and it is growing in specific places rather than evenly — which is why the constraint shows up as grid connection queues and local generation capacity rather than as a global electricity problem.

That regional concentration is the part with practical consequences for power systems, and it is measurable in a way that global AI energy totals are not.

Frequently asked questions

Related guides