Skip to content

Water-use methodology

AI inference consumes water as well as energy: water evaporated in datacenter cooling, and water consumed generating the electricity the inference draws. This page specifies how Venturi estimates per-query water use. It complements the energy methodology — water is derived downstream of whichever energy tier resolved, so if energy is null, water is null too.

Consumption, not withdrawal

The figures here are water consumption — water that is evaporated or otherwise removed from the local water cycle — not water withdrawal (water drawn and largely returned). Consumption is the figure that matters for environmental impact and is the one Venturi reports.

The two-term formula

Water use has two distinct components, and Venturi reports both:

water_site   = (energy / PUE) × WUE_site      (on-site cooling)
water_source =  energy        × WUE_source     (off-site electricity generation)
water_combined = water_site + water_source
  • water_site is the water evaporated cooling the datacenter that runs the inference (an on-site, Scope-1-style term).
  • water_source is the water consumed off-site generating the electricity the inference drew (a Scope-2-style term).
  • water_combined is their sum, and is the headline per-query figure.

Water is reported per query in millilitres (mL) with the site, source, and combined decomposition all preserved.

WUE is Water Usage Effectiveness, in litres per kWh of IT energy.

The PUE asymmetry — read it carefully

The two terms treat PUE differently, and mixing them up is the easy mistake this page exists to prevent:

  • The site term divides energy by PUE. The intent is to remove datacenter overhead and apply the cooling water-intensity to the IT energy itself.
  • The source term uses the full energy (the energy that was actually drawn from the grid, overhead included), because the grid had to generate all of it.

Applying PUE to the wrong term, or applying it twice, materially distorts the result. Venturi applies it exactly as written above, and a gate enforces the two-term shape so it cannot regress.

Null, never zero

Water follows the same honest-unknown discipline as energy and carbon. If any required term is unknown — energy is null, or a provider's WUE_site or WUE_source is unknown — the corresponding water figure is reported as null, never 0. A zero would falsely imply no water was consumed.

Note that a single-term water model (cooling only, with the source term set to zero) is the degenerate case of the formula above. Venturi uses the full two-term model and does not silently treat an unknown source term as zero.

Provider water-intensity reference

Venturi ingests provider and region Water Usage Effectiveness as versioned reference data. The site term reflects datacenter cooling; the source term reflects the generation mix powering that datacenter. Representative provider values (litres per kWh):

Provider context WUE_site (L/kWh) WUE_source (L/kWh)
Google 0.30 1.10
OpenAI / Azure-hosted 0.30 4.35
Anthropic / AWS-hosted 0.18 5.11–5.12
xAI 0.36 3.14
DeepSeek (native) 1.20 6.02

Regional generation water-intensity varies widely even within one provider — for example, the source term for an AWS region ranges from roughly 2.3 L/kWh in US-East (N. Virginia) to roughly 9.5 L/kWh in US-West (Oregon). Venturi keeps these as region-level reference tables and resolves the most specific value available, mirroring the carbon resolution chain in the energy methodology.

These values are reference inputs reimplemented from published research and provider sustainability disclosures; see licenses for attribution.

Where to go next