Data Center Environmental Impact

A data centre is a building full of servers, storage and networking equipment, kept running and cool around the clock. Because computing and cooling both draw continuous power, these facilities have quickly become one of the most closely watched pressures on energy and water systems — and the recent surge in artificial intelligence has sharpened the debate.

This page pulls the credible public numbers into one place. The headline figures are genuinely large, but they come from different studies, different years and different geographic scopes, so we keep them clearly separated rather than blending them into a single, misleading “impact” number.

From ~1.5% to ~3%: the energy trajectory

Globally, the International Energy Agency estimates that data centres consumed about 415 TWh of electricity in 2024, roughly 1.5% of world electricity. In its Base Case, demand climbs to about 945 TWh by 2030 — still just under 3% of the global total, but growing at about 15% a year, more than four times the growth rate of electricity use across all other sectors combined.

Most of that net increase is driven by accelerated servers — the power-hungry hardware behind AI — which account for almost half of the projected growth. Geographically it is concentrated: the United States and China together make up nearly 80% of the global increase in data-centre electricity demand to 2030. These are scenario estimates with sensitivity cases, not metered global totals, so treat the trajectory as a well-reasoned projection rather than a measured fact.

The United States is where the most detailed measurement exists. A 2024 Berkeley Lab / DOE report puts U.S. data-centre use at 176 TWh in 2023 — 4.4% of national electricity — and models a 2028 range of roughly 325–580 TWh, or 6.7%–12.0% of forecast U.S. electricity. The width of that range is itself the point: how much data centres will draw depends on choices about efficiency, hardware and how fast AI is deployed.

Four footprints, four different scopes

“Environmental impact” is not one number. The same Berkeley Lab / DOE report separates four distinct footprints for U.S. data centres in 2023, each measured on a different boundary. Keeping them apart is essential to reading the data honestly.

Indirect water and emissions assume the local balancing-authority grid mix; they do not incorporate facility power-purchase agreements or behind-the-meter generation. The four rows measure different things on different boundaries and are not additive.

What determines the impact?

Two facilities drawing the same electricity can have very different environmental outcomes. The main variables are:

• Grid carbon intensity. Nearly all of a data centre's operational carbon footprint comes from the electricity it uses, so whether that power is fossil-based or low-carbon matters more than almost anything else. The U.S. average works out to about 0.34 kg CO₂e per kWh, but a facility on a coal-heavy grid and one on a hydro- or nuclear-rich grid can differ materially. This is why the renewable electricity mix of the host grid is decisive.
• Cooling design and climate. Air cooling, water cooling, free-air cooling and newer liquid-cooling approaches trade electricity against water very differently, and local climate changes the maths. A dry, hot region may cut power use with evaporative cooling but at a real cost in water stress.
• Utilisation and hardware. A well-used modern server does far more computing per watt than an idle or ageing one; AI accelerators push both performance and power density higher.
• Life-cycle burden. Operating energy is only part of the story. A full resource-use view spans chips, servers, cooling gear, buildings, backup power and grid infrastructure — their manufacturing, construction, replacement and eventual e-waste, which connects to the wider problem of waste and electronics.

PUE and WUE: useful but partial

Two efficiency ratios dominate industry reporting:

• PUE (Power Usage Effectiveness) is total facility energy divided by the energy delivered to the computing equipment. A perfect 1.0 would mean every watt reached the servers; as an illustration, a PUE of 1.5 implies the site draws about 50% more power than the computing alone, for cooling and losses.
• WUE (Water Usage Effectiveness) is on-site water consumed per unit of computing energy, in litres per kWh.

Both are genuinely useful for comparing operations, but neither captures life-cycle impact. PUE says nothing about whether the electricity is clean; a very efficient facility on a dirty grid can still have a large carbon footprint. WUE counts only on-site water and ignores the often-larger indirect water used to generate the electricity. Efficiency ratios describe how well a building runs — not the full environmental cost of what it consumes.

Why per-query AI numbers are unstable

It is tempting to quote a single figure for the energy or water “cost” of one AI query, but such numbers are unreliable and usually unsuitable for universal claims. They swing by orders of magnitude depending on the model's size, the hardware it runs on, how heavily that hardware is utilised, the cooling system, and the carbon and water intensity of the local grid. Estimates also differ in what they count — training versus inference, whether facility overhead (PUE) and indirect water are included, and which model generation is assumed.

The honest framing is at the level of systems and grids, not individual prompts. AI is clearly pushing data-centre electricity demand upward — accelerated servers account for much of the projected growth — but converting that into a precise per-query footprint that applies everywhere is not something the current evidence supports.

Community and local effects

Beyond global totals, data centres are felt locally:

• Heat. Nearly all the electricity a data centre consumes ends up as heat that must be rejected to the surroundings. Yale researchers highlight a Phoenix-area study that modelled and observed local land-surface warming of up to 16°F, with nearby air warming up to 4°F as far as about 0.5 km from a facility. This is one local study and must not be generalised to every site, but it shows that heat rejection is a real, place-based effect.
• Water in stressed regions. Direct cooling water totalled about 66 billion litres across the U.S. in 2023, with 84% at hyperscale and colocation sites. National totals aside, on-site water can matter a great deal where a facility sits in a water-stressed basin.
• Grid and air. Because impact tracks the power source, a facility that pushes a region to run more fossil generation raises both carbon and local air pollutants; one on a clean grid does not. Backup generators and construction add further local effects.

Yale's experts stress that carbon and water outcomes are highly dependent on grid timing and location, cooling design and local water stress — averages hide large site-to-site variation.

How impacts can be reduced

The UN Environment Programme frames the response as a set of levers rather than a single fix, noting that energy demand is rising broadly while water impacts vary greatly by design, cooling technology, climate and location. Practical measures include:

• Cleaner power — siting on low-carbon grids and procuring renewables, so growth adds to clean energy rather than fossil generation.
• Energy and resource efficiency — higher utilisation, better hardware and lower PUE.
• Smarter cooling — matching cooling technology to climate to balance electricity against water, and avoiding heavy water use in stressed regions.
• Sustainable procurement and heat reuse — longer equipment life, recycling to cut e-waste, and capturing waste heat for nearby buildings.
• Transparent, common metrics and smart-grid management — shared reporting and demand flexibility so facilities help balance the grid.

Whether these levers add up to a smaller footprint depends heavily on whether new power is low-carbon — which ties data centres directly to the pace of the wider energy transition and the shrinking carbon budget.

Methodology and limitations

This page synthesises four authoritative public sources: the IEA's *Energy and AI* analysis for global electricity; the 2024 Berkeley Lab / DOE *United States Data Center Energy Usage Report* for U.S. electricity, water and emissions; UNEP for context and mitigation; and Yale School of the Environment for local heat and life-cycle framing. Figures are reproduced as published and dated to their source year.

Key limitations to keep in mind:

• Scenarios, not measurements. Global and future figures (IEA 2030; U.S. 2028) are modelled scenarios with sensitivity cases, not metered totals.
• Bottom-up modelling. The U.S. report is a bottom-up estimate constrained by limited public data; its indirect water and emissions assume the local grid mix and exclude facility PPAs and behind-the-meter generation.
• Scopes are not additive. Direct and indirect water — and operational versus life-cycle impact — measure different things and must not be summed.
• Local studies do not generalise. The Phoenix heat figures come from one location and are attributed as such.

We update this reference as new editions of these reports are published. It is an independent editorial synthesis for education and reference, not a facility-level life-cycle assessment; consult the original sources for authoritative figures.