AI infrastructure water consumption: what the Google 8.1B disclosure and EU 2023/1791 tell procurement

Bottom line. Google reported 8.1 billion gallons of data-centre water consumption in 2024, a 33% year-over-year increase from 6.1 billion in 2023 (Google 2024 Environmental Report). Microsoft reported 6.4 million cubic metres in 2022 at a Water Usage Effectiveness of 0.30 litres per kilowatt-hour, a 39% improvement from 0.49 the prior year (Microsoft Environmental Sustainability Report). The EU Energy Efficiency Directive 2023/1791 made WUE and water-consumption reporting mandatory for data centres above 500 kilowatts of IT power demand starting 15 September 2024 (Directive (EU) 2023/1791). Microsoft launched zero-water evaporation cooling as the standard for new builds in August 2024. AI infrastructure water consumption is a procurement-deck variable now, with vendor disclosure postures already differentiating Cohort A and Cohort B in the same shape the AM-007 security-disclosure analysis frames.

The procurement-deck question on data-centre water consumption shifted in 2024. Three changes drove the shift, each separately documented and each with a clean operational read.

First, the major hyperscalers’ published numbers reached a scale at which the operational footprint became unavoidable in cloud and co-location procurement decisions. Google’s 2024 disclosure of 8.1 billion gallons against 6.1 billion the prior year is the most-cited specific figure, but Microsoft’s parallel disclosure and the corresponding rise across the rest of the hyperscaler population establish the cohort-level signal.

Second, the EU Energy Efficiency Directive 2023/1791 made data-centre water-consumption reporting a mandatory regulatory line for facilities above 500 kilowatts of IT power demand, with the first reporting cycle requiring 2023-data submission by 15 September 2024. Vendor-discretionary disclosure became regulated disclosure for in-scope deployments overnight.

Third, the cooling-technology stack matured. Microsoft’s zero-water evaporation cooling launched in August 2024 as the standard for new Microsoft data centres, and immersion cooling at PUE in the 1.02-1.03 range became a production-grade procurement option from named vendors. The procurement-deck question stopped being whether water consumption can be reduced and started being which cooling class the vendor commits to and when.

This piece reads the three changes at the procurement level rather than at the sustainability-policy level. The operative thesis: data-centre water consumption is now a vendor-disclosure variable at the same level of procurement-decision weight as the AM-007 security-disclosure variable, with a parallel Cohort A vs Cohort B split running across the same hyperscaler population.

What the published hyperscaler numbers actually say

Google’s 2024 Environmental Report puts total data-centre water consumption at approximately 8.1 billion gallons for calendar year 2024 (Google sustainability reports). The figure is the published sum of freshwater withdrawals across the operated estate, including evaporative-cooling losses and other consumptive uses. The 33% year-over-year increase from approximately 6.1 billion gallons in 2023 reflects both the AI-workload growth and the underlying capacity expansion. The Council Bluffs, Iowa facility consumed nearly 1 billion gallons on its own, the highest of any individual Google facility.

Google Chief Sustainability Officer Kate Brandt has publicly framed the trajectory as compatible with the firm’s clean-energy goals, noting that “we’re showing it’s possible to advance the two great transformations of our time, the AI revolution and clean energy growth, hand in hand”. The framing is the vendor-side communications posture; the 8.1 billion gallon figure is the operational data point against which the framing reads.

Microsoft’s parallel disclosure in the corresponding sustainability reporting puts 2022 consumption at 6.4 million cubic metres, a 34% year-over-year increase driven primarily by AI-workload expansion. The Water Usage Effectiveness figure of 0.30 litres per kilowatt-hour represents a 39% improvement over the 0.49 baseline reported in 2021. The two figures together describe the same operational shape Google’s reporting describes: efficiency improving at the per-kilowatt-hour metric level while absolute consumption rises with workload growth.

Microsoft also discloses, distinctively, that 42% of its data-centre water consumption comes from areas experiencing water stress under the World Resources Institute Aqueduct definition. Google’s parallel figure puts 15% of freshwater withdrawals in high water-scarcity regions. The water-stress geographic distribution is procurement-relevant for the same reason as the absolute consumption figure: a vendor’s water-stress geographic exposure is part of the operational footprint the procuring enterprise inherits when it buys cloud capacity in those regions.

UC Riverside’s Shaolei Ren has published the most-cited per-workload research on AI water consumption. Ren’s analysis puts GPT-3 training at approximately 700,000 litres of direct water consumption with total usage including indirect supply-chain consumption reaching approximately 5.4 million litres; the per-conversation figure for ChatGPT in the original Ren framing is approximately 500 millilitres for a 25-50 question conversation. Ren’s published commentary on operational implications notes that data-centre water consumption is highly time-of-day-sensitive: “we don’t want to water our lawns during noon, so let’s not water our AI at noon either”. The time-of-day dimension is procurement-relevant because cooling-load timing intersects with regional water-availability cycles.

What the EU Directive 2023/1791 actually changed

Directive (EU) 2023/1791, the recast Energy Efficiency Directive, includes data-centre-specific mandatory reporting requirements in Annex VII. The reporting threshold is 500 kilowatts of installed IT power demand; data centres at or above that threshold must report annual data covering water consumption, Water Usage Effectiveness, energy consumption, Power Usage Effectiveness, and waste-heat metrics.

The first reporting cycle covered calendar year 2023 data, with submission required by 15 September 2024 and annual submissions due each 15 May thereafter. Member State enforcement mechanisms attach to the reporting requirement; the operational consequence is that an EU-based or EU-serving cloud or co-location vendor that does not report under the Directive 2023/1791 schedule is exposed to Member State enforcement separately from any procuring-enterprise contractual question.

The procurement-deck shift from this is straightforward. Vendor-discretionary disclosure became regulated disclosure for in-scope deployments. Procurement contracts with EU-based or EU-serving vendors should reference the Directive 2023/1791 reporting line as a required transparency point, and procurement teams evaluating vendors that decline to commit to the reporting line are evaluating vendors with a documented compliance gap rather than a marketing position.

Member State implementation varies. The Directive sets the reporting framework; Member States transpose the framework into national legislation with the operational specifics. Procurement teams should reference both the Directive and the relevant Member State transposition, with the Directive serving as the floor and the Member State transposition adjusting up where local regulators apply tighter operational rules.

Adjacent regulatory regimes operate in parallel. Singapore mandates data centres achieve Water Usage Effectiveness at or below 2.0 cubic metres per megawatt-hour, with facilities consuming over 60,000 cubic metres annually subject to mandatory water-efficiency management practices. China requires water consumption ratios below 2.5 litres per kilowatt-hour for government-procurement eligibility, alongside the broader push to bring average Power Usage Effectiveness below 1.5 by 2025. India’s regulatory landscape carries the underlying water-stress challenge (17% of global population on 4% of freshwater resources) without an analogous mandatory-disclosure regime yet in place. The procurement-deck question for any deployment touching multiple jurisdictions is which framework’s floor applies and whether the vendor commits to the highest-floor jurisdiction’s standard or the lowest-floor jurisdiction’s.

The cooling-technology landscape in 2026

Two cooling-technology classes have matured to procurement-grade availability.

Closed-loop water-efficient cooling. Microsoft’s zero-water evaporation cooling launched in August 2024 as the standard for all new Microsoft data centres. The design eliminates evaporative water loss as a variable rather than reducing it; the cooling fluid recirculates in a closed loop, and water consumption per kilowatt-hour drops substantially toward zero for the cooling load specifically (some non-cooling water uses remain). Microsoft’s reported figures put facility-level savings on the order of 125 million litres annually per facility against the legacy evaporative-cooling baseline. The technology is now Microsoft’s procurement default for new builds, which sets the cohort-level standard for new hyperscaler capacity.

Immersion cooling. Vendors operating immersion-cooling deployments at production scale report Power Usage Effectiveness in the 1.02-1.03 range against the legacy air-cooled industry baseline closer to 1.5. Cooling-power consumption reductions in the 90% range against air cooling are documented across multiple vendor deployments, with named cases including the University of Texas Lonestar6 supercomputer using GRC’s ICEraQ Series 10 systems at densities of approximately 70 kilowatts per rack. Construction-cost reductions in the range of 50% against legacy designs follow from eliminated raised floors, reduced mechanical infrastructure, and smaller facility footprints.

The procurement-decision question for any cloud or co-location procurement in 2026 has shifted from “what is the vendor’s PUE” to “which cooling class the vendor commits to for new capacity, what timeline the vendor commits to for converting existing capacity, and whether the vendor’s reported WUE figures are converging on the published cooling-class baselines or diverging from them”. A vendor with WUE of 0.30 L/kWh against an industry-leading 0.10 baseline at full immersion-cooling deployment is a different procurement proposition than a vendor whose reported WUE is 0.30 against a 1.5 industry-laggard baseline; the WUE figure alone does not differentiate, the cooling-class commitment does.

How this maps onto the AM-007 vendor-response-split framework

The same Cohort A vs Cohort B framing AM-007 applies to security disclosure transfers cleanly to sustainability disclosure with one cross-domain shift.

Cohort A vendors publish annual environmental reports including absolute water-consumption figures, WUE metrics, and water-stress regional breakdowns, with year-over-year deltas disclosed and Board-level sustainability accountability named. Google, Microsoft, AWS (within the Amazon corporate disclosure), and most major hyperscalers sit in Cohort A on this axis at the corporate level; granularity below the corporate level varies. The procuring enterprise reading Cohort A disclosure has procurement-grade evidence to anchor the cloud or co-location decision against, including the AI-workload-attributable portion of the disclosed numbers where the vendor reports at that granularity.

Cohort B vendors decline to publish at this granularity, citing competitive sensitivity or operational complexity. Some smaller co-location operators sit in Cohort B by default; some specialised AI-infrastructure providers position the non-disclosure as an operational efficiency. The procurement-deck consequence is symmetric to AM-007: the deploying enterprise inherits the residual transparency burden when buying from Cohort B, including the burden of synthesising water-consumption estimates from indirect signals (kilowatt-hour delivery × industry-baseline WUE) when the vendor will not report directly.

Under the EU Directive 2023/1791, Cohort B becomes a non-compliance posture for in-scope deployments rather than a discretionary product-positioning choice. This is the cross-domain shift from the AM-007 security-disclosure analogue: in the security domain, “intended functionality” classification is a defensible (if procurement-relevant) product position; in the sustainability domain for in-scope EU deployments, non-disclosure is a Member State enforcement risk. Procurement teams evaluating both axes can apply consistent decision logic: prefer Cohort A across both, treat Cohort B as a residual-burden inheritance in security and as a compliance-risk inheritance in sustainability.

What the data implies for Q2-Q4 2026 procurement

The hyperscaler absolute-consumption trajectory is not bending downward in the public 2024 numbers. The Google 33% year-over-year and Microsoft 34% year-over-year increases reflect AI-workload growth outrunning per-kilowatt-hour efficiency improvements at the WUE metric level. The procurement-deck implication is that absolute water consumption from AI infrastructure will continue to rise at the cohort scale through 2026 even as the named-vendor WUE figures continue to improve.

Two procurement-decision moves follow from that operational read.

First, contract-level commitments to cooling-class are procurement-deck defensible in 2026 in a way they were not in 2023. A procurement contract that requires the vendor to commit to closed-loop or immersion cooling for new capacity, with documented WUE targets and reporting cadence, is technically achievable on the available technology stack. Vendor pushback that the commitment is technically infeasible reads against the Microsoft August 2024 zero-water-evaporation rollout as the operational counterexample.

Second, water-stress geographic exposure becomes a procurement-decision variable separately from absolute consumption. A deployment that lands in a Microsoft 42%-water-stress region produces a different operational footprint than one that lands in an 8%-water-stress region, regardless of absolute consumption per workload. Procurement teams that scope cloud and co-location decisions against the vendor’s water-stress regional disclosure rather than the corporate-aggregate figure are reading the procurement-relevant data; teams that read only the corporate aggregate are missing a meaningful variable.

The five questions for the procurement committee on this axis add to the AM-140 procurement-committee six and to the AM-007 cross-agent five. They focus specifically on the vendor-side sustainability-disclosure posture.

1. Does the vendor publish annual data-centre water consumption, WUE, and water-stress regional breakdown at granularity sufficient to map the procuring enterprise’s deployment to specific facilities or regions?
2. For in-scope EU deployments, does the vendor commit to Directive 2023/1791 reporting on the published schedule, with submission evidence available to the procuring enterprise?
3. What cooling class does the vendor commit to for new capacity in the deployment region, with a documented WUE target?
4. What is the timeline the vendor commits to for converting existing legacy-cooling capacity in the deployment region, and the contract-level remedy if the timeline slips?
5. What is the deployment’s water-stress geographic exposure, measured by the share of the in-scope vendor capacity sitting in regions classified as high-stress under the World Resources Institute Aqueduct water-risk atlas or an equivalent third-party reference?

A vendor that cannot answer all five in writing for the deployment region is in Cohort B on this axis regardless of corporate-level Cohort A posture. A vendor that can answer all five is in Cohort A, the residual deployment-layer transparency work is bounded, and the procurement contract has the documentation surface that the EU Directive 2023/1791 audit cycle and the procuring enterprise’s own sustainability reporting both require.

Holding-up note

The primary claim of this piece (that AI infrastructure water consumption has moved from sustainability-footnote to procurement-deck variable, anchored on the Google 8.1B / Microsoft 0.30 WUE / EU Directive 2023/1791 trio, with a Cohort A vs Cohort B vendor-disclosure split parallel to the AM-007 security analysis) is on a 60-day review cadence. Four kinds of evidence would move the verdict.

A major hyperscaler reporting a year-over-year decrease in absolute water consumption against rising AI workload would weaken the central trajectory framing and require the piece to be re-read against the inflection. EU Member State enforcement action under Directive 2023/1791 setting reporting precedent would strengthen the regulatory-line procurement-deck weight materially. A published cooling-technology benchmark showing sub-1.05 PUE at production scale beyond the current immersion-cooling pilots would update the cooling-class procurement question with a third class to evaluate against. A major regulatory regime (US, UK, China) establishing a parallel mandatory reporting framework would extend the EU 2023/1791 procurement-deck shift to a global-procurement variable rather than an EU-specific one.

If any land, the Holding-up record for AM-008 captures what changed, dated. Original claim stays visible. Nothing is quietly removed.