AI Infrastructure

How Much Electricity Do AI Data Centers Use in 2026? The Real Numbers

Newaiera Desk · 2026-08-03 · 7 min read

Global data center electricity consumption hits 565 TWh this year, up 26%. We break down the gigawatt figures, what share is actually AI, and why power — n

Ask what limits AI growth and the answer used to be GPUs. In 2026 it is electricity, and the numbers have stopped being abstract.

Gartner forecasts global data center electricity consumption will reach 565 terawatt-hours in 2026, up from 447 TWh in 2025. That is a 26% increase in a single year, and it is driven almost entirely by AI workloads.

For scale: 565 TWh is roughly what France consumes in a year, across its entire economy.

The consumption curve

A 118 TWh year-over-year increase is not a trend line bending gently. It is the steepest single-year jump the sector has recorded, and the forecast assumes no acceleration beyond what is already contracted.

Power capacity is the tighter constraint

Consumption tells you how much electricity gets used. Capacity tells you whether it can be delivered at all — and that is where the squeeze actually bites.

Worldwide data center power demand rises about 27% in 2026, from 104 GW to 132 GW, with forecasts putting it at 290 GW by 2030. Nearly tripling grid draw inside five years is not something a power system absorbs quietly.

The US picture is sharper still:

A jump from 41 GW to 66 GW in one year means roughly 25 GW of new firm capacity has to appear on the US grid. For reference, a large nuclear reactor is about 1 GW. You do not build 25 of those in twelve months.

What share of this is actually AI?

This is where the reporting usually gets sloppy. Not all data center power is AI — a great deal is ordinary cloud, storage and networking that would exist anyway.

Gartner's estimate: AI-optimized servers account for 31% of data center power consumption in 2026. So roughly 175 of those 565 TWh.

The crossover is what matters. By 2027, AI-optimized servers are projected to consume more power than all conventional servers combined. Inside eighteen months, the exception becomes the majority.

Why this is now a business constraint, not an environmental footnote

Three things follow from these numbers, and they are already visible in how the industry behaves.

Siting is dictated by electricity, not by customers. New capacity goes where power is available and cheap, which is why so much of it is landing next to hydro, nuclear and stranded gas rather than near the users it serves.

Interconnection queues have become the real lead time. Getting a facility connected to the grid now takes longer in many markets than designing, financing and building it. The chips arrive before the electricity does.

Power purchase agreements have turned into competitive moats. A hyperscaler with twenty-year contracted supply can commit to capacity that a rival simply cannot promise, regardless of how much cash the rival has.

The scarce input in AI stopped being silicon some time in the last eighteen months. It is now firm, dispatchable electricity — and unlike chips, you cannot air-freight it.

What this means if you are buying AI, not building it

You will feel this as pricing and availability, not as a headline.

Inference pricing has fallen steadily on a per-token basis, but that decline is fighting rising input costs. If power costs keep climbing at this rate, the per-token deflation everyone has budgeted for slows down. Capacity constraints also show up as regional availability gaps — a model live in one cloud region and queued in another.

If you have a multi-year AI budget built on the assumption that inference gets cheaper every quarter, it is worth stress-testing that assumption against these figures.

The honest uncertainty

These are forecasts, and forecasts in this sector have been wrong in both directions. Efficiency gains are real and persistent — better silicon, better cooling, better utilisation, and model architectures like mixture-of-experts that cut active compute per token substantially.

It is entirely possible 2027 comes in under projection because efficiency outruns demand. It is also possible the projections are low, because they mostly assume no new step-change in model size or agentic workloads that run for hours rather than seconds.

What is not in doubt is the direction, or that the binding constraint has moved from the fab to the substation.