Industries / Technology / Software and SaaS / Cloud and data centers
Energy and water use
This is the work of measuring how much electricity and water data centers use, reporting it, and cutting it.
7
Milestones achieved, from 2007 to 2024.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2007
One number for overhead
The Green Grid, an industry group founded in 2007, introduced power usage effectiveness (PUE): a facility's total power divided by the power its computers use, so 2.0 means one watt of cooling and losses for every watt of computing. Uptime Institute, which has surveyed operators' PUE every year since 2007, puts the average at 2.50 in 2007 and 1.52 in 2026, and says most of the fall came by 2014 from cheap fixes such as warmer supply air, contained aisles and blanking panels.
LessonThe easy savings were found in about seven years; since then only new buildings move the average.
Credit
The Green Grid; Uptime Institute survey
2016
Efficiency flattens the curve
A 2007 report to the US Congress put US data center use at about 61 billion kWh in 2006, 1.5% of the country's electricity, more than double 2000. Lawrence Berkeley National Laboratory's 2016 follow-up found use rose about 90% from 2000 to 2005, 24% from 2005 to 2010 and only 4% from 2010 to 2014, reaching about 70 billion kWh, 1.8% of US electricity.
LessonVirtualization and large, efficient cloud sites absorbed years of growth in demand.
Credit
Arman Shehabi and colleagues, Lawrence Berkeley National Laboratory, with the US EPA's ENERGY STAR program behind the 2007 report
2019
Server heat warms a city
In late 2019 the Tietgenbyen Energy Centre in Odense, Denmark, began taking heat of about 27 °C from Meta's data center and raising it with heat pumps to 70 to 75 °C for the city's district heating network. The engineering firm Ramboll, which designed it for the heating utility Fjernvarme Fyn, gives about 45 MW of heat and more than 12,000 homes served (2023 figure); Denmark abolished a tax on such surplus heat for certified businesses in January 2022.
Credit
Fjernvarme Fyn; Meta; Ramboll
2020
Forecasts corrected
Eric Masanet, Arman Shehabi, Nuoa Lei, Sarah Smith and Jonathan Koomey estimated in Science that the world's data centers used about 194 TWh in 2010 and 205 TWh in 2018, about 1% of world electricity, while their workloads and compute instances grew more than sixfold. They wrote that widely cited claims of doubling energy use came from scaling up old figures by traffic growth.
LessonEnergy use cannot be forecast from internet traffic, because the efficiency of each computation changes too fast.
Credit
Masanet et al., Science, February 2020
2022
Water use made public
The city of The Dalles, Oregon, sued to keep Google's water use secret after The Oregonian asked for it in 2021, arguing it was a trade secret, and Google paid the city's legal bills. In December 2022 the city settled, agreeing to release Google's yearly water use for 2012 to 2021 and future years on request. Google said it would no longer treat site water figures as trade secrets anywhere in the US; The Oregonian then reported its use there had nearly tripled in five years to more than a quarter of the city's water.
LessonA data center's water use became public through a records fight, not a reporting rule.
Credit
The Oregonian/OregonLive (reporter Mike Rogoway); Reporters Committee for Freedom of the Press
2024
Europe makes sites report
Under the EU's Energy Efficiency Directive and Delegated Regulation 2024/1364, data centers of 500 kW and more must report energy, water, heat reuse and renewable energy to an EU database, first by 15 September 2024. The Commission's assessment counts 770 reporting data centers, about 36% of those required, with an average PUE of 1.36, average water use of 0.58 liters per kWh of computing, and only 67 reusing any waste heat. Germany's Energy Efficiency Act goes further: data centers opening from 1 July 2026 must reach a PUE of 1.2 and reuse at least 10% of their energy, rising to 20% for those opening from July 2028.
Credit
European Commission; German Bundestag
2024
AI ends the plateau
Lawrence Berkeley National Laboratory's December 2024 report found US data center use flat at about 60 TWh from 2014 to 2016, then rising as GPU servers for AI spread, to 176 TWh in 2023, 4.4% of US electricity. It projects 325 to 580 TWh by 2028, 6.7% to 12% of the national total. On-site water use rose from 21.2 billion liters in 2014 to 66 billion in 2023, and the water used at power plants to make that electricity was nearly 800 billion liters.
LessonThe efficiency that held use flat for a decade was outrun by AI hardware within six years.
Credit
Arman Shehabi and colleagues, Lawrence Berkeley National Laboratory, for the US Department of Energy
Watch for
An EU data center label
The European Commission has adopted a rating scheme that would give each reporting data center a label, generated from the EU database, covering energy efficiency, water efficiency and low-emission power, and has opened a consultation on minimum performance standards for energy, water, heat reuse and grid services.
Achieved whenthe rating regulation is in force and the first labels are published.
Claimed by
European Commission, September 2026
Open
01
Per-site water ledger
A public table of yearly water use at individual data centers, each figure tied to its source: company environmental reports where they give site figures, and city water utility records got through public-records requests. Where both exist for one site, show them side by side.
The briefCredit · unclaimed02
Per-prompt energy at the wall
An independent, measured version of Google's per-prompt method for open-weight models. Serve each model at realistic load and report energy per median prompt at four boundaries: the GPUs alone, the whole server, the server plus idle capacity held in reserve, and the building (multiplied by the site's PUE).
The briefCredit · unclaimed
Open 01 · the brief
Per-site water ledger
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of yearly water use at individual data centers, each figure tied to its source: company environmental reports where they give site figures, and city water utility records got through public-records requests. Where both exist for one site, show them side by side.
Why it is still open
The EU collects site data but publishes it only in anonymised, national form, and Berkeley Lab's national estimate had to work around "the unavailability of facility-level data". The public tallies are small: TapWaterData's May 2026 set covers five US facilities, and a 2026 Scripps News records investigation covers the Columbus, Ohio area, where it found companies often ask officials to black out water figures as trade secrets.
Built with what exists
Google's yearly environmental report, which Scripps says gives water use for its data centers; public-records requests to city water utilities, using The Dalles settlement as a model; an AI model that reads PDFs to pull figures out, each checked by a person; a Git repository with one row per site and year and a link to its source.
First step
Copy the site water figures from Google's latest environmental report and file records requests with the water utilities of five of those sites.
Done looks like
100 site-years covering at least 30 sites and 4 operators, each with a source document, and at least 10 sites where company and utility figures are compared.
Open 02 · the brief
Per-prompt energy at the wall
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An independent, measured version of Google's per-prompt method for open-weight models. Serve each model at realistic load and report energy per median prompt at four boundaries: the GPUs alone, the whole server, the server plus idle capacity held in reserve, and the building (multiplied by the site's PUE).
Why it is still open
Public benchmarks measure only part of the chain. AI Energy Score, run with Hugging Face, "is focused solely on GPU energy" on H100 GPUs; "How Hungry is AI?" (Jegham et al., 2025) infers energy from public API speeds and company multipliers instead of meters; Google's 0.24 Wh figure covers the whole chain but is its own measurement of its own service.
Built with what exists
Open-weight models from the AI Energy Score text-generation set; vLLM to serve them; NVIDIA's NVML for GPU power; the server's own power readings (IPMI or Redfish) or a metered rack power strip; the site's published PUE.
First step
Rent one GPU server whose provider reports whole-server power, run the AI Energy Score text set through vLLM, and log GPU and server power each second.
Done looks like
Three models at three load levels, at least 1,000 prompts each, energy per median prompt at all four boundaries, raw logs public, and one outside lab reproducing one model's server-level figure within 10%.