Decentralised News · Editorial by Heath Muchena · Evidence reviewed 11 November 2026. Original DN analysis and local instrument; field testing is claimed only where expressly reported.
What Matters
A data center’s water figure becomes useful only when its boundary is clear. Withdrawal, consumption, cooling efficiency and basin stress answer different questions. DN’s worksheet records the facility, period, numerator and IT-energy denominator before calculating a ratio. It also separates local water context from electricity-related impacts, without inventing a facility map or assigning a water footprint to an individual AI prompt.
Evidence Block
Evidence reviewed: 11 November 2026. Primary sources are the US Department of Energy, US Geological Survey, Lawrence Berkeley National Laboratory and World Resources Institute. No facility meters, supplier accounts or municipal permits were independently audited. The example is constructed. Aqueduct supplies screening context; the local tool records user-provided evidence and does not query a live map.
An AI service may run in a region that contains several data centers, water suppliers and watersheds. A national average or regional cloud label does not identify the water used by the facility that processed a particular workload. The right starting point is a boundary record: which facility, which water measure, which period and which operational activity?
This matters when a buyer tries to compare two confident statements. One provider may report water withdrawn from a utility. Another may report water consumed in cooling. A third may show an annual efficiency ratio for a global fleet. The figures can all be genuine and still be unsuitable for a direct comparison. A smaller number is not evidence of a better outcome until the definitions align.
Claim versus evidence: stop mixing the water questions
The USGS distinguishes withdrawal from consumptive use and return flow. Withdrawal concerns water removed from a source; consumption concerns the part that is not readily available again in the immediate water environment, including evaporation. The relationship depends on the system and its return flows. Those distinctions should remain visible in a data-center record. Source: USGS water-use concepts.
| Question | Evidence needed | Common comparison error |
|---|---|---|
| How much water entered the facility? | Withdrawal or delivery volume, source and dates | Treating all incoming water as consumed |
| How much water was consumed? | Defined consumption method and return-flow information | Subtracting unrelated discharge figures |
| How much site water was used per unit of IT energy? | Aligned site-water numerator and IT-energy denominator | Using total facility energy without relabeling the ratio |
| How exposed is the water source to local stress? | Basin, indicator, dataset version and seasonal context | Using a country ranking as a facility impact verdict |
DN’s contribution is the connection between these records. We have not calculated a supplier league table. A complete row establishes what the reviewer has documented, not that every environmental question has been answered.
Find the facility and its actual water source
Use a facility identifier that can be connected to the underlying report. If only a corporate total is available, retain it as corporate evidence. Do not silently allocate it to a particular building. If the supplier serves a workload across multiple locations, ask how locations and volumes are attributed. A refusal or inability to disclose is an evidence gap, not a reason to invent a split.
Identify the source separately from the facility address. The physical site, utility service territory and supplying basin may not be the same boundary. A map pin is useful only when the record explains what it represents. An approximate public region may support a broad screening question while remaining insufficient for a facility-level conclusion.
Then identify water type and operational purpose. The evidence might describe potable supply, reclaimed water, another source or a mix. Keep cooling, other site uses and any exclusions visible. Changing the source of water can alter the question being asked even if a headline volume stays unchanged.
A practical supplier request is short: provide the reporting unit, water source, metric definition, start and end dates, and method for assigning any customer share. Ask for a document version and responsible reporting function. The public DN worksheet needs only reference aliases; precise sensitive infrastructure information can remain in an appropriately controlled internal file.
Put the right energy below the water figure
The Department of Energy describes water usage effectiveness as annual site water use divided by annual IT equipment energy, commonly expressed in litres per kilowatt-hour. It treats cooling and water management as connected operational questions. A ratio calculated with total facility energy is a different ratio unless the definition explicitly permits that boundary. Source: DOE data-center water guidance.
Before dividing, align the periods. A calendar-year water report and a trailing-twelve-month energy report do not necessarily cover the same operation. A monthly water figure and an annual energy total are plainly mismatched. Record the mismatch instead of converting it through a guessed seasonal factor.
Also distinguish an observed denominator from allocated IT energy. A tenant may receive a modeled share of a facility total rather than a meter reading. The allocation can be useful if the method is documented, but the resulting ratio should retain that label. Adding decimal places does not turn the allocation into a direct observation.
DN’s agent energy-cost worksheet considers energy boundaries and accepted tasks. This water record adds a different numerator and a local environmental context. Joining the two requires consistent dates and boundaries; it does not justify dividing a facility’s annual water total by an arbitrary count of prompts.
A basin map is context, not a meter
WRI’s Aqueduct 4.0 provides modeled water-risk indicators, including baseline and forward-looking layers. Its screening function helps identify locations for further assessment. A selected indicator needs a stated geography, time frame and version; a risk category is not a measurement of a named facility’s water use. Source: WRI Aqueduct 4.0 data.
Use the map to formulate the next question. If the relevant basin appears exposed under a chosen indicator, ask how the facility sources water, how conditions vary across the operating year and which local restrictions or monitoring records are relevant. The model result and the site evidence should remain separate entries so a reviewer can see whether the second step was completed.
WRI also publishes country and provincial rankings built from Aqueduct information. Those aggregated products answer a broader geographic question than an individual site review. DN does not convert a country position into a facility score. Source: WRI country-ranking dataset.
A facility can therefore have a low reported annual intensity and still require careful local review. Another can report a higher intensity with a different source or seasonal pattern. The evidence record preserves both observations without forcing them into a universal best-to-worst ordering.
Keep indirect water visible
LBNL’s 2024 US data-center report separates direct site water from water associated with electricity generation. Its modeling explains that generation technologies and locations affect the indirect component. National modeling provides useful context, but it does not identify a particular customer’s facility or marginal workload. Source: LBNL 2024 report.
For a procurement comparison, state whether electricity-related water is included. Do the same for equipment manufacturing and other lifecycle activity. A site-only record may be a sensible first step, provided it is labeled site-only. A broader estimate needs its own sources and assumptions; it should not be merged into the meter total without explanation.
This prevents a common wording problem. “Our facility uses less water under this measure” is narrower than “this AI workload has a lower complete water footprint.” The latter requires more evidence. DN recommends writing the claim after the boundary record is assembled, so the wording can match what was actually observed.
Use the DN Data Center Water Boundary Record
DN Data Center Water Boundary Record v1.0
Record the evidence you have and the questions you still need to resolve. Entries stay in this page until you download them. Enter reference aliases only; do not enter personal names, private documents, credentials or wallet secrets. This tool performs no external verification and grants no approval.
Machine-readable record
Schema: dn.data-center-water-boundary.v1. Version 1.0.0. Record creation time is not evidence verification time. Any edit clears the previous output.
The seven areas cover facility boundary, water numerator, IT-energy denominator, water source, basin context, indirect exclusions and workload allocation. For each, record a non-sensitive reference alias, an issue or an unknown. The tool counts documentation states; it does not check the meter, inspect the location or validate the map selection.
The optional ratio accepts litres, IT kilowatt-hours, a water-measure type and a matching-boundary selection. It requires a positive denominator and reported agreement on the period and boundary. The output labels the numerator. A withdrawal-based ratio is not silently relabeled consumption, and an undefined numerator produces no ratio.
Constructed example: enter 2,000,000 litres and 1,000,000 IT kWh for the same reported period and boundary. The result is 2 litres per IT kWh. If the water entry means withdrawal, that is the label retained in the record. The number says nothing about how much returned to the local environment, the supplying basin’s condition or an individual prompt.
If the dates do not match, the ratio is withheld. If the denominator is zero, the tool asks for correction rather than displaying infinity. Any edit clears the previous result. The JSON export preserves the assumptions, documentation gaps and creation timestamp, with no approval or independently reviewed evidence date.
Turn the record into a useful supplier conversation
Start with one proposed claim and one reporting period. Give competing suppliers the same definitions. Ask them to indicate which fields they can support directly, which use estimates and which are unavailable. That comparison shows whether the evidence is commensurable before anyone debates whose headline is smaller.
Prioritize unresolved boundaries over additional precision. A more accurate site volume is valuable, but it cannot correct a denominator from the wrong year. A detailed basin model cannot supply a missing link between the site and its water source. Resolve the connection that changes the interpretation first.
For the broader infrastructure context, DN’s AI infrastructure stack explains the surrounding physical systems. The water record is a separate environmental evidence task; investment exposure to a cooling company does not establish the water performance of a particular data center.
Methodology and limitations
DN’s method is a documentation framework, not a standardized environmental audit. The conditional arithmetic is water litres divided by IT kWh for the entered boundary. The seven evidence areas have no weights, composite rating or hidden penalty. A full set of references can still contain weak or contradictory evidence and requires substantive review.
No facility-level map was produced because verified facility, source and period data were not assembled for this article. No claim is made about water saved by a cooling technology or about causal effects on a community. Basin screening, operating permits, local observations and lifecycle estimates require their own assessment. The worksheet keeps those boundaries available for that later work.
Frequently Asked Questions
Can I compare withdrawal with consumption?
Keep them separate. They describe different parts of the water system. A comparison needs aligned definitions and evidence of the relevant return flows.
Is a low litres-per-kWh value proof of low water risk?
No. Operational intensity and local water context answer different questions. Review the source, location, season and chosen risk indicator separately.
Can this produce water per AI prompt?
No. The tool has no verified workload allocation or marginal-water model. Dividing a broad total by a prompt count would introduce assumptions it does not contain.
Why is there no live map?
A map requires verified location and source relationships. The record identifies the evidence needed to build one without presenting approximate geography as measured facility performance.
Corrections and change log
Send corrections with the metric definition, source document, period and affected passage to press@decentralised.news. 11 November 2026, version 1.0: introduced the boundary record and conditional ratio; separated withdrawal, consumption, basin screening and indirect water. No facility ranking or measured per-prompt result is published.
