Data Centres: Where the Power and Water Go

What counts as a data centre

The figures in this article are for the United States. They come from a report by Lawrence Berkeley National Laboratory, written for the US Department of Energy and published in December 2024. The authors define nine kinds of data-centre space, from network closets and small server rooms up to hyperscale facilities run by companies that deploy internet services at massive scale. Amazon, Microsoft and Google together own more than half of the hyperscale sites, according to the report’s market source.

Whatever the size, the job is the same. Servers compute, storage devices keep data, network equipment moves it, and a layer of infrastructure keeps all of that powered and cool. The report adds up all four.

How much electricity

The report’s headline numbers are these.

2014 to 2016: about 60 TWh a year, flat, continuing a trend seen since about 2010.
2018: about 76 TWh, 1.9% of US electricity.
2023: 176 TWh, 4.4% of US electricity.
2028: a range of roughly 325 to 580 TWh, or 6.7% to 12.0% of the electricity forecast for that year. At an assumed average utilisation of 50%, that is a power demand of 74 to 132 GW.

Growth has sped up. The report gives a compound annual rate of about 7% for 2014 to 2018, 18% for 2018 to 2023, and 13% to 27% for 2023 to 2028. Between 2017 and 2023, total demand more than doubled.

Why the flat years ended

For most of the 2010s, data centres did far more work without using more power. The report credits a shift from companies’ own server rooms to colocation and cloud facilities, plus better cooling and power management, higher server utilisation and lower idle power. In 2014, 20% of servers sat in colocation or hyperscale sites. The report estimates 74% did in 2023, rising to 85% by 2028.

What changed from 2017 was hardware. Servers built around graphics processors, used for AI, became a large enough share of the stock to push the total up. The authors say the rapid growth in these accelerated servers ended the era of generally flat data-centre energy use.

Cooling, and the water it uses

Almost all the electricity ends up as heat, which has to be removed. The report finds large differences between sites. A common efficiency measure, power usage effectiveness (PUE), is the total energy a facility uses divided by the energy that reaches the computers. Closer to 1 is better. Cooling choices matter, and so does climate.

Cooling also uses water, in two places.

At the site. Direct water use rose from 21.2 billion litres in 2014 to 66 billion litres in 2023. Hyperscale and colocation sites account for 84% of the 2023 total.
At the power station. The report estimates nearly 800 billion litres of water consumed indirectly through the electricity data centres use in 2023, about twelve times the direct figure by our arithmetic, plus 61 billion kilograms of CO₂-equivalent emissions. Per kilowatt-hour that is 4.52 litres and 0.34 kg, against a national average of 4.35 litres and 0.35 kg.

The two are linked. Evaporative cooling is generally more energy-efficient than a waterless system, so a design that saves electricity can use more water, and the reverse. The report warns that a low water figure is not automatically good.

What the numbers cannot tell you

The authors are open about the limits. The estimates come from a bottom-up model built on shipment data and many assumptions, because direct energy data are scarce, and they say that lack of transparency limits the analysis. The 2028 range is a scenario spread, not a forecast. The modelled PUE values assume systems run as designed, which the report says is rarely true, so real efficiency is probably worse. The indirect water and emissions figures assume each site uses its local grid mix, without counting power-purchase deals or on-site generation.

The report also puts the rise in context: electric vehicles, new manufacturing and the electrification of industry and buildings are expected to add much larger demand over the coming decades.

Our take: the useful question about a data centre is not “how big” but “where, on what power, and cooled how”. The same server uses different amounts of water and carbon in different places, and the public data to check any one site are thin.

Questions and answers

How much electricity do US data centres use?

A Lawrence Berkeley National Laboratory report estimates 176 TWh in 2023, which is 4.4% of US electricity. Its 2028 range is roughly 325 to 580 TWh, or 6.7% to 12.0%, depending on how fast AI hardware is built and used.

Why did data-centre electricity use start to rise again?

The report says demand was flat at about 60 TWh a year from 2014 to 2016 thanks to efficiency gains and a move to larger facilities. From 2017, GPU-accelerated servers for AI grew enough to push the total up, and demand more than doubled between 2017 and 2023.

Do data centres use a lot of water?

The report estimates 66 billion litres consumed at the sites in 2023 and nearly 800 billion litres indirectly through the electricity they use. Cooling choices trade electricity against water, so a low water figure is not automatically good.

Can these figures be applied to other countries?

No. They are for the United States, and the report's own estimates rest on a bottom-up model with limited public data. Other countries have different grids, climates and cooling.

Do you agree?

Take it to Chevlume: write your view, add your sources and let others answer. The closed alpha is by invitation.

Discuss this on Chevlume

No invite yet? Ask for a place

Keep reading

ENERGY / EXPLAINER

Sources

2024 United States Data Center Energy Usage Report (LBNL-2001637, Shehabi et al.) · Lawrence Berkeley National Laboratory · Accessed 5 November 2026
DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers · U.S. Department of Energy · Accessed 5 November 2026

Published 5 November 2026

Discuss on X