Where the data centre lands

The question is local, not global

The environmental conversation about AI usually happens at two useless altitudes: the individual query, which is negligible, and global emissions, which are too aggregated to act on. The decisions that matter sit in between, and they are decisions about where a building goes.

The International Energy Agency estimated global data centre electricity consumption at around 415 terawatt-hours in 2024 — roughly 1.5% of world electricity — and projected it roughly doubling to about 945 terawatt-hours by 2030, close to Japan's total consumption today, with AI the fastest-growing component. Those are estimates with wide uncertainty bands, and the direction is not in doubt.

But a national or global percentage hides the thing that actually affects people, which is that this load is extremely concentrated. Data centres cluster where land, power, fibre and tax incentives coincide, and a handful of regions carry a share out of all proportion to their size.

What concentration does

Ireland is the clearest case. Its Central Statistics Office reported data centres consuming around a fifth of all metered electricity in the country by 2023 — a larger share than all urban households combined. The grid operator effectively stopped new connections in the Dublin region until the late 2020s. A national climate target now has to be met around a load that grew faster than the plan.

Northern Virginia hosts the largest concentration in the world, and the utility's forecast load growth has become a public planning fight about new transmission lines and who pays for them.

Water is the second constraint and it is more local still. Evaporative cooling consumes water — it does not merely borrow it — and the consumption happens where the building is. Google reported data centre water consumption in the billions of gallons annually, and Microsoft reported a sharp rise in the year large model training scaled up. Several facilities sit in water-stressed regions, including in Chile, Arizona and Spain, where local opposition has focused precisely on this. Some operators have moved to air cooling or closed-loop systems, which use more electricity, which is the trade.

Who pays for the grid. When a large load arrives, transmission has to be built. Whether that cost falls on the company or is spread across every household bill is a regulatory decision, made in a proceeding almost nobody attends, and it is the most consequential AI policy question in many places.

Local generation. Where grid connections are slow, operators have installed their own generation, sometimes gas turbines, sometimes before permits are finalised. A 2025 dispute in Memphis over turbines at a data centre in a neighbourhood already carrying poor air quality is the template for arguments that will recur.

Efficiency, and why it does not settle it

Data centres have become dramatically more efficient. Power usage effectiveness — total facility power divided by computing power — has fallen from around 2.0 two decades ago to about 1.1 in the best modern facilities, meaning overhead is now a tenth rather than a doubling. Chips do far more computation per watt each generation.

And total consumption keeps rising, because efficiency lowers the cost of computation and lower cost raises demand. This is Jevons paradox, observed by William Stanley Jevons in 1865 about coal, and it is why per-unit efficiency claims and total consumption figures can both be honest and point in opposite directions. When a company reports improved efficiency, the useful follow-up is what happened to total consumption.

What is worth arguing about

If you want to engage with this beyond feeling bad about your own usage, the questions with leverage are these.

Is the electricity additional and clean? A facility powered by newly built renewable capacity is a different proposition from one that buys certificates while drawing from a coal-heavy grid. Ask about hourly matching rather than annual offsetting, because a grid at 3am is not the grid at noon.

Who bears the grid cost? Read the utility filing. This is where the money is.

What is the water source and what is the local stress level? Consumption in a water-abundant region is a different fact from the same consumption in a drought region.

Is there a community benefit that is not a jobs number? Large data centres employ few people once built. Tax abatements are often granted against employment figures that never arrive.

And the honest framing for an individual: your chat usage is a rounding error against your flights, your diet and your heating. The place your influence exists is as a citizen in a planning process, not as a user counting prompts.