Writing

Are data centers bad for the environment?

Published 22 September 2026

The question bundles five separate things: electricity, emissions, water, land and waste, and they have five different answers. Worse, the honest answer to the biggest of them is that it depends more on which grid the building is plugged into than on anything the operator does inside it. A facility can be identical in every respect and differ tenfold in carbon by virtue of its postcode.

Taking the question apart

Most coverage of data center environmental impact picks whichever of these is most alarming in a given week and treats it as the whole story. They are not interchangeable:

Notice that three of the five are dominated by a decision made before the building opened: where to put it, and how to cool it. That is the consistent finding across all of this, and it is why a single national or global figure explains so little.

Electricity: the global number is not the problem

The International Energy Agency put global data centre electricity use at about 415 TWh in 2024, roughly 1.5% of world electricity consumption, and projects something near 945 TWh by 2030.

That 1.5% is genuinely small, and quoting it as reassurance is as misleading as quoting the growth rate as alarm. The reason is concentration. Data centres do not distribute themselves evenly across a grid: they cluster where land, fibre and power interconnects are available, which means a handful of counties absorb a share that looks nothing like the national average. A region where three hyperscale campuses land has a genuine capacity problem that "1.5% of global electricity" does not describe at all.

This is the substance behind most local opposition, and it is a fair complaint even when the global framing is reassuring. The two facts are not in conflict; they are about different things.

Separately, most published capacity figures are not consumption at all: how much electricity a data center uses covers why an announced 500 MW site may be drawing 40.

Emissions: the grid decides, not the operator

Data center emissions are almost entirely indirect: they come from generating the electricity, not from the building itself. The IEA put them at roughly 180 Mt CO₂ in 2024, about 0.5% of energy-related emissions, growing by nearly 80% across this decade.

The thing that actually moves that number is grid carbon intensity: grams of CO₂ per kilowatt-hour where the facility draws power. The spread is enormous:

That is better than a tenfold range. The same servers, the same workload, the same PUE, sited in two different places, produce carbon figures an order of magnitude apart. No efficiency measure available to an operator comes close to mattering as much as that siting decision.

Which means a data center carbon footprint is mostly a statement about geography. Quoted without the grid attached it is not a measurement of the facility at all, and comparing two operators' figures without checking where their capacity sits compares their electricity suppliers rather than their engineering.

Water: covered by the cooling design

The short version is that water consumption ranges from effectively zero to millions of gallons a day at the same computing load, because it is set by the cooling method rather than the size of the site. And it trades directly against energy: evaporative cooling is used precisely because it lowers a facility's PUE, so the sites using the most water are frequently the ones optimised hardest for efficiency.

How much water a data center uses works through the whole trade, and the cooling load and water use estimator shows both sides at a load you choose.

Land and waste: the ones that are genuinely smaller

Land use is real but modest by industrial standards. A large campus occupies tens to low hundreds of acres: substantial locally, unremarkable against agriculture, logistics or retail parks. The sharper objections are usually about what the land was before, and about facilities that consume local services while employing relatively few people once built.

Hardware waste is the least well measured of the five. Servers are replaced on roughly three-to-five-year cycles, and large operators resell or recycle at high rates because there is money in it. The honest position is that this is a smaller problem than the electricity, and that the data is poor enough that confident claims in either direction are not supportable.

What efficiency did, and why it stopped working

Between roughly 2010 and 2020 data centre workloads grew enormously while total energy use stayed remarkably flat. Hyperscale facilities displaced thousands of inefficient server rooms, virtualisation raised utilisation, and PUE fell sharply. It is one of the better efficiency stories in any sector, and it is routinely left out.

Two things ended it. PUE has largely stopped improving: the industry average has sat near 1.5–1.6 for about a decade, because the easy gains were taken and the remaining ones are expensive. And AI training raised rack density from 5–10 kW to 40–130 kW, which is new demand rather than the same demand served more efficiently.

So the flat decade was not a law of nature. It was efficiency outrunning growth, and growth has taken the lead back.

The honest answer

Are data centers bad for the environment? At a global scale, they are a small and rapidly growing share of electricity and a smaller share of emissions, and they replaced something less efficient. At a local scale, a facility on a coal-heavy grid using evaporative cooling in a water-stressed basin is genuinely doing harm that a facility in Sweden on closed-loop cooling is not.

Both sentences are true, which is why the argument never resolves. The useful questions are not about data centres in general:

1. Which grid? This dominates the carbon answer by a factor of ten.
2. Which cooling design? This dominates the water answer by a factor of thirty.
3. Consumption or announced capacity? Usually the latter.
4. Is the renewable claim annual or hourly?
5. What is the local grid's spare capacity? The global share is irrelevant to a county.

Timing is becoming another design variable. A data center demand-response plan can shift eligible compute away from a constrained hour or use stored energy while the grid is stressed. That can lower cost and sometimes emissions, but the result depends on what generation replaces the load and whether the delayed work creates a rebound later.

A note for people entering the field

Most of what is decided here is engineering, not policy. Whether a facility lands on a clean grid, whether it uses water or electricity to reject heat, whether racks run at 8 kW or 120 kW and what that does to the network inside. These are design decisions made by people, and the build-out is creating a great many of those jobs.

It is a reasonable field to enter for anyone who wants infrastructure work where the constraints are physical and the arguments are winnable with numbers. The networking side in particular has changed shape: the skills that matter in a modern fabric are not the ones that mattered in a campus network, and that gap is where the demand is.

Sources

• Energy and AI: International Energy Agency, 2025. Global consumption, emissions and projections.
• Five charts that put data-centre energy use and emissions into context: Carbon Brief
• 2024 United States Data Center Energy Usage Report: Lawrence Berkeley National Laboratory
• Carbon dioxide emissions coefficients: US Energy Information Administration

Figures are ranges from published studies read in September 2026. Grid intensity figures move year to year and vary by region within a country; they are given as approximate bands rather than point values for that reason. Where sources disagree, the wider range is given rather than the more quotable number.