OSIEN
←Back to blog
12 min read

AI's hidden thirst: water, data centers, transparency

We talk about AI's CO2 emissions. We talk less about water. AI relies on data centers that dissipate dense heat. Water is used for cooling. It also plays a role in the electricity production that powers these infrastructures.

The useful topic is not for or against AI. The useful topic is where, with what water, with what energy, with what evidence.

Two levels of water footprint

On-site water

Server cooling via cooling towers, exchangers, water circuits. Reuters cites an order of magnitude of about 1.4 million liters/day for a medium-sized data center. This volume varies according to climate, computing density, design, site age, engineering choices.

Off-site water

Electricity is not dry. Depending on the mix, the water portion can be dominant. It is rarely published in an exploitable manner.

Numbers exist, comparison is lacking

  • Li et al. (arXiv) project 4.2 to 6.6 billion m3 of water withdrawals linked to global AI demand in 2027, with 0.38 to 0.60 billion m3 of water consumed (evaporated).
  • In the US, the LBNL report estimates 66 billion liters consumed directly by data centers in 2023 (all data centers, not just AI).
  • Google reports 6.1 billion gallons (about 23 billion liters) consumed by its data centers in 2023.
  • NatureFinance highlights geographic exposure: a significant portion of data centers are located in water risk or drought zones, with biodiversity issues.

These orders of magnitude frame the debate. They do not replace public, verifiable, comparable accounting between actors.

Common mistakes we see

  • !Withdrawal confused with consumption.
  • !On-site water confused with off-site water.
  • !Per request numbers without technical context.
  • !Undeclared scopes (site, season, load, cooling, electricity mix).

France: water becomes a public issue

ARCEP measures 681,000 m3 of water withdrawn by data centers in 2023, almost exclusively potable, with an annual increase of +19% (after +17% in 2022). ARCEP estimates nearly 6 million m3 including indirect water related to electricity.

These volumes are low compared to agriculture at the national level. They can become structuring locally, in the wrong place, at the wrong time.

Electricity and water meet in planning: RTE projects a rise to 23 to 28 TWh by 2035 in certain scenarios, or around 4% of French electricity consumption.

Europe: data center reporting is being structured

The EU is setting up a European database and annual reporting with KPIs including water (WUE), via delegated regulation (EU) 2024/1364.

This does not solve everything. It changes the norm: water becomes a tracked, compared, discussed variable.

The energy signal that accelerates everything

Google and NextEra announced the restart of Duane Arnold (615 MW) with a 25-year purchase contract, target 2029. Other nuclear plant restart projects are underway in the US (Palisades, Three Mile Island).

The OSIEN reading is simple.

  • →Demand is structural, not cyclical.
  • →The water-energy-territory trade-off becomes an infrastructure issue.

OSIEN position: from debate to evidence

OSIEN takes an open science approach applied to the environmental impact of digital. On water and AI, our roadmap:

  • ✓Strict vocabulary (withdrawal vs consumption, on-site vs off-site).
  • ✓A minimal disclosure format for AI services (metrics, assumptions, scope).
  • ✓A database of documented, comparable cases.
  • ✓An inventory of real levers and trade-offs (water vs energy, performance vs sobriety, location vs network).

What we are looking for now

  • Operators ready to share metrics in a comparable format.
  • Local authorities who want to equip project review and acceptability.
  • Researchers and engineers to audit assumptions and consolidate methods.

Follow OSIEN. Join the Reef. Contribute to the commons.

Minimal metrics to require

Before believing a number on the water footprint of an AI service, check that these elements are declared:

PUE (Power Usage Effectiveness)

Ratio between total site energy and IT equipment energy. A PUE of 1.2 means 20% overhead for cooling and infrastructure. French average: 1.6. Best sites: 1.1.

WUE (Water Usage Effectiveness)

Liters of water consumed per kWh of IT energy. Allows comparing sites. Typical value: 0.5 to 2 L/kWh depending on climate and cooling technology.

Declared scope

On-site only, or including off-site water (electricity production)? Without this precision, comparisons are meaningless.

Water source

Potable water, surface water, recycled water, desalinated seawater? Local impact depends on the source.

Local water stress

Is the site in a risk zone? The reference indicator is Baseline Water Stress (WRI Aqueduct).

Withdrawal vs consumption

Withdrawal: water removed from the environment. Consumption: water not returned (evaporated, integrated). The difference can be 1 to 10 depending on technologies.

If this data is not provided, the published number is not verifiable.

References

Follow OSIEN. Join the Reef. Contribute to the commons.

JoinContact