Everybody publishes a number. Nobody publishes one you can check.
Five questions asked of the largest data centre operators, each answerable from public documents. Four of them separate the field. The fifth does not separate anyone, because on the fifth nobody scores at all, and that is the whole reason this page exists.
Discloses
Is any water figure published at all?
Around 47 per cent of operators measured their water use in 2025, and seven of the top twenty publish no WUE. This column fails first and fails often.
Basis stated
Is it consumption or withdrawal, and does the operator say which?
AWS publishes 0.12 L/kWh on a withdrawal basis and Microsoft 0.27 L/kWh; the bases differ, so the numbers cannot be compared. An unstated basis makes a good-looking figure meaningless.
Site level
Can a community see the facility next to them, or only a fleet average?
A global average tells a council nothing about the building in their township. More than a dozen shareholders have now formally asked Amazon, Microsoft and Google for site-specific figures.
Measured
Is the figure metered, or calculated by the claimant?
The return side is rarely metered anywhere, so consumption is usually withdrawal minus an assumption.
Checkable
Could a third party verify it without asking the operator?
Nobody assessed so far scores here. That is the opportunity, stated as a column.
Open until 30 September 2026
This is version 0.1 and it is preliminary: assessed from published secondary sources on 1 September 2026, with every row to be re-checked against the operator’s own report before v1.0. If a cell is wrong we would rather be corrected than be right by luck.
The assessment
Read the last column first. Every operator here publishes something, several publish carefully, and not one publishes a record anybody outside the company could verify.
v0.2 · rebuilt from primary sources
v0.1 was assessed from secondary coverage. Re-reading every operator’s own report moved six of seven rows, every one of them in the operator’s favour. Two were wrong enough to be indefensible. The corrections are recorded below, and every row now names the document it was read from.
Nothing moved on the last column. Not one operator publishes a record a third party could check without asking them for it — including the two carrying limited third-party assurance, which is an auditor’s opinion rather than something a council could verify itself.
yes 2 · partial 1 · no or not established 0 · A 9–10, B 7–8, C 5–6, D 3–4, F 0–2
| Operator | 1. Discloses | 2. Basis stated | 3. Site level | 4. Measured | 5. Checkable | Grade |
|---|---|---|---|---|---|---|
| CyrusOneColocation | yes | yes | part | yes | no | B |
| MetaHyperscaler | yes | yes | yes | part | no | B |
| Vantage Data CentersColocation | yes | yes | part | part | no | C |
| Digital RealtyColocation | yes | yes | no | yes | no | C |
| MicrosoftHyperscaler | yes | yes | no | part | no | C |
| GoogleHyperscaler | yes | yes | part | ? | no | C |
| Amazon Web ServicesHyperscaler | yes | yes | no | ? | no | D |
What changed from v0.1, and why
Vantage Data Centers · F → C. v0.1 recorded no published water figure. The 2023 ESG report discloses 680,848 m3 withdrawn at 71 per cent declared coverage.
CyrusOne · not graded → B. v0.1 could not establish an operator-wide figure. The 2026 report separates withdrawal, consumption, discharge and restoration at 93 per cent declared coverage.
Meta · D → B. v0.1 marked site level partial. Meta publishes water withdrawal by named facility across five years.
Digital Realty · C → C. Grade unchanged, measurement mark corrected: the 2025 figure carries a signed limited-assurance statement.
Microsoft · D → C. v0.1 recorded measurement as not established. Microsoft states a primary-data basis with estimates disclosed where data is unavailable.
Amazon Web Services · D → D. Grade unchanged, basis mark corrected: WUE is defined explicitly as litres withdrawn per kWh, and scope covers leased and shared sites.
Equinix · C → withdrawn. Could not be verified against primary reporting. Removed rather than carried forward unchecked.
Primary reporting could not be retrieved for verification. Row withdrawn rather than carried over unchecked.
CyrusOne · B
Reports 2025 withdrawal of 1,376.7 ML, consumption of 1,077.7 ML and discharge of 299.0 ML as separate quantities, with 101.6 ML restored, against a declared scope of 93 per cent of colocation capacity. Separating the three flows is what almost nobody else does, and declaring the coverage is the second thing almost nobody else does. A per-facility water-risk heatmap gets it partway to site level without publishing per-facility volumes. Design WUE 0.002 L/kWh. Still nothing a third party could check.
Read from: CyrusOne 2026 Sustainability Report, Water section · confidence H
Meta · B
Publishes water withdrawal by named facility across five years, which is the thing communities actually ask for and which no other operator here does. Total 5,637 ML for 2024, data centres 4,145 ML, also broken out by source. The 1,019 ML withdrawn for construction is excluded from the headline and then disclosed separately rather than left out, which is the correct way to handle an exclusion. Not independently checkable.
Read from: Meta 2025 Environmental Data Index, §3.1 Water Withdrawal · confidence H · claims register C-025
Vantage Data Centers · C
Discloses 680,848 cubic metres withdrawn under SASB IF-RE-140a.2, with withdrawal data covering 71 per cent of total floor area and 15.3 per cent of it in high or extremely high baseline water stress. Names the stressed sites rather than aggregating them away, and states plainly that legacy sites carry higher WUE than the closed-loop design. Coverage is declared but the basis of measurement is not. Not checkable.
Read from: Vantage Data Centers 2023 ESG Report, SASB Index pp. 50-51 · confidence H
Digital Realty · C
Reports CY2025 total water consumption of 1,539,657 kGal, carrying a signed limited-assurance statement from Cventure covering the sustainability metrics. Assurance is the strongest measurement evidence in this table. It is also an auditor telling you nothing appeared materially misstated, which is not the same as a record you could check yourself, and the portfolio view still shows a community nothing about its own facility.
Read from: Digital Realty 2025 Impact Report, Table 2 and Cventure assurance statement p. 95 · confidence H · claims register C-030
Microsoft · C
FY25 global average WUE of 0.27 L/kWh, a 25 per cent reduction since baseline on owned datacentres, and all designs since August 2024 committed to zero-water closed loop. States that water withdrawal and consumption use primary data, with estimates where primary data is unavailable, which is a coverage declaration rather than a silence. Aggregate throughout: no site-level breakdown, so a community cannot see its own facility.
Read from: Microsoft 2026 Environmental Sustainability Report · confidence H · claims register C-003, C-005, C-027
Google · C
Publishes absolute consumption and states the basis, with consumption up 34 per cent from 2024 to 2025 and 7.7 billion gallons replenished against roughly 78 per cent of freshwater consumption. Reporting covers owned and operated sites; the treatment of third-party operations is a scope question we have not been able to close. Whether the figures are metered or modelled is not established.
Read from: Google 2026 Environmental Report · confidence M · claims register C-004, C-026
Amazon Web Services · D
Reports 9.4 billion litres withdrawn in 2025 across global leased, owned and shared data centres, and defines WUE explicitly as litres withdrawn per kWh of IT load, so the withdrawal basis is stated in the metric rather than left to be inferred. It is a withdrawal figure, not the consumption quantity Microsoft's WUE measures, and the two are still routinely compared as though they were the same. No site-level figures.
Read from: Amazon 2025 Sustainability Report, water section · confidence H · claims register C-003
How this is scored, so you can disagree with it properly
- Every cell traces to a source recorded in our public claims register, with the date it was accessed and a confidence mark. Nothing here was inferred from a company’s reputation.
- UNKNOWN is not a bad grade, it is an honest one. Where we could not establish a fact, the cell says so and the row is not graded on it. CyrusOne is ungraded for exactly this reason.
- An absence is recorded as an absence. Vantage is marked as publishing no water figure in the ESG reporting reviewed, which is a claim about what we found, not about what exists. Confidence is marked low and right of reply is open.
- v0.1 is assessed from published secondary reporting. Before v1.0 every row is re-checked against the operator’s own primary report, and any correction received is applied and credited.
- Grades measure checkability, not virtue. An operator using very little water badly documented scores worse than a heavy user who publishes clearly. That is deliberate: this index is about whether a number can be trusted, not whether it is small.
The fifth column is not a gap in our research. No operator publishes a record a third party can verify, because until recently there was no way to. There is now, and you can run it yourself →
Questions
Which data centre operators disclose their water use?
Most large operators publish something. Google reports absolute consumption, Microsoft and AWS publish WUE, Meta reports volume for owned sites only, and Equinix and Digital Realty both publish portfolio WUE. Around 47 per cent of operators measured water use at all in 2025, and seven of the top twenty publish no WUE figure.
Can you compare data centre WUE figures between operators?
Not reliably. AWS publishes 0.12 litres per kilowatt-hour on a withdrawal basis and Microsoft publishes 0.27, and those measure different quantities. Voluntary reports also use different scopes, and they do not consistently include the indirect water embedded in electricity.
Does any data centre operator publish a water figure that can be independently verified?
Not among the operators assessed here. Every published figure rests on the operator's own calculation, and none is accompanied by a record a third party could check without asking the operator. That is the finding this index exists to make visible.
What is the difference between water withdrawal and water consumption?
Withdrawal is what a facility takes. Consumption is what it takes minus what it returns. Because the return side is rarely metered anywhere, published consumption is usually a withdrawal figure with an assumption subtracted from it.
How is The Water Line scored?
Five criteria: whether any figure is disclosed, whether the basis is stated, whether it is site level or a fleet average, whether it is metered or calculated, and whether a third party could check it. Every cell carries its source. Anything not established is marked UNKNOWN rather than guessed, and version 0.1 is preliminary with right of reply open.