# AI Data-Centre Water Tracker: an open, reproducible referee for data-centre water burden

**NM AI Research** (ORCID 0009-0003-4213-7769) · 8 August 2026 · v0.5 working note
Repository: `sites.csv` + `reproduce.py` + `SOURCES.md` (stdlib Python, no dependencies) · CC BY 4.0

*AI disclosure: parts of this text were artificially generated with AI assistance and reviewed by the author. The models, and the conflicts they create, are named in the Conflict of interest and scope section.*

## Summary

Ren et al. (arXiv 2606.21760, June 2026) define a Water Consumption Impact index (WCI) for AI data centres and apply it to ten sites, but release no code or data and leave two channels open: the coupling between a data centre and the hydropower on its grid, and the off-site relocation of water that "closed-loop" cooling produces. This note ships the open, reproducible version. It rebuilds the WCI from public inputs and checks each recomputed value against the published one; it adds a hydropower-coupling flag asserted on primary reservoir data; and it adds a relocation ledger that quantifies how much of a site's water footprint moves off-site into the electricity supply. The finding that motivates the exercise: closed-loop cooling does not remove water, it relocates the great majority of the footprint to the grid, roughly 92 to 95 per cent on the modelled inputs used here, with the direction holding across every plausible variation of them. Layer 1b adds a second reporting year. Recomputing the same index on the operators' FY2025 figures, with capacity and peaking held constant, moves five of the six primary-verified sites up by 15.6 to 33.3 per cent in a single year. The sixth is flat on total withdrawal and conceals the largest change in the set: its draw on potable municipal supply rose more than sixfold while its reclaimed supply fell.

Figures are reported with their ranges where they are ranges: the off-site relocation shares are bands, and each per-site WCI is a single reproduced value sitting on an (r, PF) decomposition a reader can vary. Every input is named and motive-tagged. The contribution is the open recomputation and the source ledger, not automated data extraction.

## Scope and standing of evidence

The index operates at three altitudes, and the note claims "per-site" only for what is measured per-site. Every row and layer is graded once, here, and the grading is not restated in the sections that follow.

| Rows or layer | Rests on | Standing |
|---|---|---|
| Six Google rows: Council Bluffs, The Dalles, Douglas, Midlothian, Henderson, Mayes | operator environmental reports, FY2024 and FY2025, both EY-assured | measured, primary |
| Microsoft Wisconsin | City of Racine release letter, records request PR2025-343 | planning, primary document |
| Google Botetourt, Meta Lebanon | supply contract and local reporting | planning, secondary |
| xAI Memphis | utility and advocacy sources, carried as a band | range, secondary |
| Layer 2, hydropower coupling | USBR 24-Month Study, USGS 14105700 | measured, primary; a reservoir-level flag |
| Layer 3, relocation ledger | Siddik et al. (2021) EWIF bands, class WUE values | modelled, a grid-average bound |

The direct WCI (Layer 1) is per-site: the site's own withdrawal against its host utility's water capacity. The off-site footprint (Layer 3) is estimated from a regional electricity-water intensity factor and is never a metered per-facility litre count. The hydropower coupling (Layer 2) is a cluster or reservoir-level flag carrying a directional price implication, not a dollar forecast.

## Method (Layer 1)

The index is

    WCI = C_peak / K = (W / K) x r x PF

where W is the average daily water withdrawal (ML/d), K is the host utility's maximum deliverable water capacity (ML/d, not electrical capacity), r is the consumptive ratio (fraction evaporated, dimensionless), and PF is the peaking factor (peak day over average day, dimensionless). W/K is water on water, so the WCI is the dimensionless peak consumptive draw as a fraction of the utility's deliverable water capacity.

`reproduce.py` recomputes the WCI for each row of `sites.csv` and flags any row whose recomputed value disagrees with the source's cited value beyond a 5 per cent tolerance. A flag is an instruction to verify against a primary source, not a number to trust. The seed is Ren et al.'s Table 6, inherited from the framework paper rather than author-selected, which removes site cherry-picking as a bias vector.

## Results

Across the ten seed sites the reproduced WCI spans 0.18 to 1.86, and eight of ten reproduce within 5 per cent. Verification against operator environmental reports and utility open-records upgraded six sites to primary, caught three as non-operational planning figures, and held one as a range.

| Site | WCI FY2024 | WCI FY2025 | Status | Off-site share (Layer 3) | Hydropower coupling (Layer 2) |
|---|---|---|---|---|---|
| Google, Council Bluffs IA | 1.86 | 2.48 | primary | 50–67% | water-only |
| Google, The Dalles OR | 0.18 | 0.24 | primary | 77–91% | water + hydro (Columbia / The Dalles) |
| Google, Douglas Co. GA | 0.42 | 0.42 | primary; potable draw +526% under a flat total | 55–70% | water-only |
| Google, Midlothian TX | 0.31 | 0.38 | primary | 35–50% | water-only |
| Google, Henderson NV | 0.35 | 0.41 | primary | 57–80% | water + hydro (Lake Mead / Hoover) |
| Google, Mayes Co. OK | 0.42 (cited 0.44) | 0.54 | primary; 6% flag (rounding) | 35–50% | water-only |
| xAI, Memphis TN | 0.58 | n/a | range, framework-secondary | 57–70% | water-only |
| Microsoft, Wisconsin | 0.26 (cited 0.34) | n/a | planning, not operational; 22% flag | 92–95% | water-only |
| Meta, Lebanon IN | 0.77 | n/a | planning, not operational | 92–95% | water-only |
| Google, Botetourt Co. VA | 1.43 | n/a | planning, not operational | 55–67% | water-only |

Three of Ren et al.'s ten "sites" are non-operational planning figures presented as measured, across three different operators. Botetourt is a contracted 2 MGD reservation for a campus not online until roughly 2028; Lebanon began construction in February 2026 and is a closed-loop design whose circulating water figures are pipeline capacity for the whole LEAP district (2 to 25 MGD phased to 2031), not Meta's metered use. Both are excluded from any measured-WCI reading. Wisconsin is the third, and the City of Racine's own release letter establishes it: public records request PR2025-343, dated 17 September 2025, describes the material as "estimated water use data" for a campus "currently under construction", and Wisconsin Examiner reports the same figures as "the Racine projections". The withdrawal in Ren's table, 0.087 ML/d, is the full-buildout projection of 8.442 million gallons a year divided by 365; the consumptive ratio and the peaking factor are both assigned defaults rather than site measurements. Its components also give 0.26 against the cited 0.34. Memphis is a moving multi-site target: the reproduced withdrawal sits inside the real range (Memphis Light, Gas and Water committed 1.3 MGD; current aquifer draw around 0.81 MGD; up to 3.7 MGD requested across two sites), so it is reported as a band, not a point.

## Layer 1b: the second reporting year

Layer 1 stays on FY2024, because that is the year Ren et al. drew on and reproducing their published values requires their inputs. Layer 1b recomputes the same index on the operators' FY2025 reports, holding the host utility capacity K and the peaking factor PF fixed, so the only quantities that move are withdrawal and consumptive ratio.

| Site | WCI FY2024 | WCI FY2025 | Change |
|---|---|---|---|
| Google, Council Bluffs IA | 1.859 | 2.479 | +33.3% |
| Google, Mayes Co. OK | 0.415 | 0.539 | +29.8% |
| Google, The Dalles OR | 0.182 | 0.237 | +29.8% |
| Google, Midlothian TX | 0.313 | 0.378 | +20.9% |
| Google, Henderson NV | 0.353 | 0.408 | +15.6% |
| Google, Douglas Co. GA | 0.418 | 0.419 | +0.2% |

Five of the six rise between 15.6 and 33.3 per cent in one reporting year. Across the seven-row measured set the range widens from 0.18 to 1.86 in FY2024 to 0.24 to 2.48 in FY2025, a top-of-range increase of a third. Six of those seven rows are re-based; xAI Memphis publishes no FY2025 figure and is held at its FY2024 value, which sits inside the range and moves neither bound. Because the capacity denominator is held constant, this is growth in draw against the same utility rather than a re-estimate of the utility.

Douglas County is the exception, and it carries the method point. Its total withdrawal is flat, down 0.8 per cent, so its WCI does not move. Google reports the source split for that site, and underneath the flat total the potable municipal draw rose from 7.4 to 46.3 million gallons, an increase of 526 per cent, while reclaimed wastewater fell 9.7 per cent. An index built on total withdrawal cannot see a sixfold rise in draw on drinking-water supply. Where an operator publishes the split it belongs in its own column, because substitution away from reclaimed supply is a different signal from total volume, and it is the one a municipal system feels. Eemshaven reports the same style of split and is the candidate second instance.

## Layer 2: the hydropower coupling

For a site drawing on a reservoir-fed grid, one exogenous drought reaches it through two channels at once: it lowers the reservoir that supplies cooling water, and it cuts the hydropower on the same reservoir, which raises the regional power price. The causal arrow runs from drought to both; the data centre is the exposed party, not the cause. Data-centre water use remains a small fraction of agriculture, and this note does not claim otherwise.

A site is flagged as double-coupled only when it draws from a reservoir that also feeds hydropower and that reservoir carries a primary drought or streamflow signal.

- **Henderson NV, strong.** The USBR June 2026 24-Month Study (Most Probable) projects Lake Mead at about 1,037 feet at year-end 2026, a Level-1 shortage roughly 192 feet below full pool. Hoover Dam's effective generating capacity is head-dependent, so a lower reservoir is simultaneously a water shock and a power-price shock. The site draws Colorado River water.
- **The Dalles OR, real but weaker.** The Columbia at The Dalles (USGS 14105700) has run near 54 per cent of normal in mid-2026, feeding a 1.8 GW dam marketed by the Bonneville Power Administration. The Columbia is a far more resilient system than the Colorado with no minimum-power-pool threat, so this is a directional mechanism rather than acute stress.

A note on a superseded figure: the earlier projection that Glen Canyon Dam would breach its 3,490-foot minimum-power-pool by about August 2026 is retired. The current study's Most Probable run holds Lake Powell near 3,504 feet at year-end 2026, above the line, with Reclamation committed to avoid a decline below 3,500 feet, despite water-year 2026 inflow at 34 per cent of average.

## Layer 3: the relocation ledger

"Closed-loop equals near-zero water" is a claim about on-site cooling only. Per unit of IT energy, the off-site share of the total footprint is

    off-site share = (PUE x EWIF) / (WUE + PUE x EWIF)

which is independent of the site's absolute energy use, so no per-site megawatt figure is required. WUE is the on-site cooling water intensity (about 1.8 L/kWh for evaporative, about 0.2 for closed-loop), EWIF is the grid electricity-water intensity (a grid-average bound), and PUE is fixed at 1.2. The EWIF band per grid is drawn from Siddik, Shehabi and Marston (2021), whose US mean is about 5.1 L/kWh including hydropower reservoir evaporation, with a range from 0 to roughly 85 L/kWh by region.

Evaporative sites relocate roughly 35 to 91 per cent of their water footprint off-site; closed-loop sites relocate roughly 92 to 95 per cent, rounded from a computed 91.5 to 94.7: the on-site number falls toward zero, but the footprint moves to the grid, it does not vanish. The direction holds across every plausible variation of these inputs and the precise band does not, and the difference matters. Raising the closed-loop WUE from 0.2 to 0.3 L/kWh, a shift well inside the uncertainty of a class estimate, moves the share to 88 to 92 per cent; taking the low EWIF from 1.8 to 1.0 gives 86 per cent. What no plausible perturbation removes is the conclusion itself: even at five times the assumed WUE the off-site share is still 68 to 78 per cent. Read the band as a consequence of two estimated inputs and the majority-relocation finding as the result that survives them. The Dalles sits at 77 to 91 per cent off-site, a consequence of the Columbia grid's high water intensity from hydropower reservoir evaporation, which ties the relocation ledger back to the Layer 2 coupling.

The empirical anchor is Morrison, Strubell et al. (2025, ICLR), who report a 20-million to 13-billion-parameter model series consuming 2.769 million litres of water even on an "extremely water-efficient" data centre, with model development at about half of training. Their finding that training power fluctuates between roughly 15 and 85 per cent of maximum draw is direct support for the peaking factor in the WCI.

## Layer 4: water quality, a flagged channel only

Layers 1 to 3 are quantity channels. A distinct quality channel is flagged but not quantified: construction-phase discharge fouling municipal water before a data centre is operational. In July 2026, Meta's construction entity in Cheyenne, Wyoming discharged the pathogen *Cupriavidus gilardii* into the city's reclaimed-water system during a cooling-loop fill-and-flush; the Board of Public Utilities suspended industrial data-centre wastewater connections area-wide while it cleaned up. The bacterium's origin is unconfirmed. This is kept strictly as a single-case qualitative flag, not an index term and not evidence of a systemic pattern; a second case would be required before it became a layer.

## Honest ceilings

- **Bands, not points.** r, PF and EWIF are ranges; the decomposition is the honest output. Peer-reviewed support: Zhang et al. (2025, NeurIPS, "PCAM") show that deterministic footprint accounting understates spatial and temporal grid variance badly enough that a probabilistic treatment cuts error from about 108 to about 7 per cent.
- **The measured set is small and concentrated.** With the three planning rows excluded, seven sites remain and six of them are Google. The measured layer is close to a single-operator dataset, and its behaviour across two reporting years reflects one company's disclosure practice as much as it reflects the sector.
- **Reproducibility is not neutrality.** Open recomputation gives calculation transparency; it does not by itself remove seed-selection or imputation bias. That is why the seed set is Ren et al.'s Table 6 rather than author-chosen, and why every estimate is named and motive-tagged.
- **Static-bounds boundary.** This tracker uses grid-average intensities and annual reservoir studies. It does not attempt real-time macro-grid simulation; a dynamic per-balancing-authority EWIF (via the LBL Water IMPACT Tool) is the natural next version.
- **Reproducible means open recomputation, not scraping.** `sites.csv` is manually curated from cited primaries, because operator environmental reports are non-standardised and shift boundaries year on year. The script recomputes and self-checks; anyone can change an input and re-run.
- **Motive-neutral.** Developers' "we are a fraction of agriculture" counter is correct and stays in; the vendor "closed-loop is near-zero water" framing is corrected by Layer 3. The accurate counter is left standing and the overstated one is checked against the same method.

## Positioning against prior work

A literature sweep confirms the gap and narrows it. The nearest neighbour, Morrison, Strubell et al. (2025), is a training-lifecycle telemetry study on the authors' own data centre, a complement and a modern water anchor rather than a per-site public referee of operators' facilities, and it does not address the hydropower coupling. The premise that water is under-reported and off-site matters is now well established, so the contribution here is narrow and specific: an open, reproducible, per-site referee of operators' facilities from public data, plus the two couplings the framework leaves open. The sweep also names an unaddressed gap, the economic impact of regional water pricing on data-centre operating expenditure, which is exactly where the Layer 2 price-transmission channel sits.

## Reproduce

    python3 reproduce.py            # print the table, the flags, and the relocation ledger
    python3 reproduce.py --tol 0.03 # tighten the reproduction tolerance

The exit code is non-zero if any row fails the reproduction check. On the seed as shipped it is always non-zero: two of Ren et al.'s ten rows do not reproduce within tolerance, Mayes at 6 per cent and Wisconsin at 22 per cent, and those flags are part of the finding rather than a fault to be cleared.

## Verification

The document behind each row, in the order of the standing table above. The reasoning for each is in Results; this section names the sources only.

| Rows or layer | Document |
|---|---|
| Six Google rows | Google 2025 Environmental Report (FY2024) and 2026 Environmental Report (FY2025), both independently assured by EY; utility open-records for Mayes and Henderson |
| Microsoft Wisconsin | City of Racine release letter for public records request PR2025-343, 17 September 2025. The released data table itself is unpublished, so its digits come from Wisconsin Public Radio and Wisconsin Examiner, both 17 September 2025 |
| Google Botetourt | Western Virginia Water Authority agreement, October 2025, contracted 2 MGD rising to a potential 8 MGD; reported by Cardinal News, Virginia Business and a Roanoke Rambler records request |
| Meta Lebanon | Daily Journal and Indiana Capital Chronicle, 11 to 12 February 2026, for the construction start; Citizens-Lebanon supply ramp for the LEAP district, 2 MGD in 2027 to 25 MGD by 2031, for the circulating figures |
| xAI Memphis | Protect Our Aquifer; Memphis Light, Gas and Water |
| Layer 2 | USBR Lower and Upper Colorado 24-Month Study, June 2026; USGS gauge 14105700 |
| Layer 3 | Siddik, Shehabi and Marston (2021), Environmental Research Letters 16 064017, doi:10.1088/1748-9326/abfba1, for the EWIF bands |

Everything in the last four rows is established by reporting, contract documents or a published model rather than by an operator's assured disclosure, which is why none of them enters the measured set. The WUE class values, the peaking factors and the consumptive ratios assigned to non-Google operators are estimates rather than site measurements. Each is named in `SOURCES.md`, and each row carries its standing in the `motive_tier` column of `sites.csv`.

## Conflict of interest and scope

The research question, method, sourcing decisions and analytical judgements are
the author's. Anthropic Opus 4.8-5.0 assisted with data retrieval, calculation,
literature search and drafting across v0 through v0.4. OpenAI GPT-5.6 Sol
assisted with the v0.5 disclosure and bundle revision. Parts of this text were
artificially generated and were reviewed by the author before publication. Each
assisting model is not always a neutral party to the subject matter.

Anthropic states that it signed an agreement to use all of the compute capacity at
SpaceX's Colossus 1 data centre (Anthropic, *"Higher usage limits for Claude and a
compute deal with SpaceX"*, 6 May 2026,
[anthropic.com/news/higher-limits-spacex](https://www.anthropic.com/news/higher-limits-spacex)).
OpenAI and Oracle state that they entered an agreement to develop an additional
4.5 GW of Stargate data-centre capacity in the United States (OpenAI, *"Stargate
advances with 4.5 GW partnership with Oracle"*, 22 July 2025,
[openai.com/index/stargate-advances-with-partnership-with-oracle](https://openai.com/index/stargate-advances-with-partnership-with-oracle/)). The author therefore treats neither lab as neutral on the
buildout. All operators are scored the same way.

Guarantee: every measured claim is traced to the primary document named in the
Verification note, and reproduce.py regenerates both tables from sites.csv.
Modelled or estimated inputs are identified as such in the Verification note
rather than presented as measurements. A reader can check this without
trusting either party.

What the author cannot guarantee: a language model's output can be wrong in ways
that survive review. In prose the error is locally plausible and consistent in
tone with what surrounds it; in code it simply runs, and a wrong constant or a
mis-set filter still returns a clean number. Several methods have been deployed to mitigate
this, including explicit instructions, internal red-teaming and cross-lab
blindspot checks, but the author does not claim the review is exhaustive.
Corrections are logged against the DOI when surfaced.

No warranty is offered beyond the terms of the CC BY 4.0 licence. Independent
analysis and open-science documentation only, not investment advice.

## Sources

Primary and public throughout. Operator environmental reports (Google 2025 Environmental Report, FY2024, and Google 2026 Environmental Report, FY2025, both EY-assured; utility open-records for Mayes and Henderson; City of Racine public records request PR2025-343 release letter, 17 September 2025, for Wisconsin); USGS 14105700 (Columbia at The Dalles); USBR Lower and Upper Colorado 24-Month Study (June 2026); Siddik, Shehabi and Marston (2021, Environ. Res. Lett. 16 064017) for EWIF; Morrison, Strubell et al. (2025, ICLR, arXiv 2503.05804) and Zhang et al. (2025, NeurIPS) for the empirical and methodological anchors; Ren et al. (arXiv 2606.21760) as the framework anchor and seed, cited not copied. Full motive-tiered ledger in `SOURCES.md`.

## Version history

*Corrections are logged rather than silently applied; no finding has been reversed. The full trail is in `SOURCES.md`.*

- **v0.5 (8 August 2026).** Disclosure and presentation only. The first-exposure wording points to the conflict-of-interest section without naming a vendor. The full disclosure names both assisting labs and links each infrastructure interest to the lab's own announcement. `reproduce.py` names both assisting labs. No figure, table, finding, input or code behaviour changed.
- **v0.4 (8 August 2026).** No figure, table, finding or input changed. The AI-assistance disclosure moves to the front of the note, the repository README and the source ledger, and states that parts of the text were artificially generated, which is the wording the EU AI Act Article 50(4) transparency duty uses. The full note at the end is unchanged in substance. One wording change in Layer 3: the sensitivity result is now stated directly rather than summarised in a single adjective.
- **v0.3 (7 August 2026).** Bundle alignment. No figure, table or finding changed. The README was rebuilt against this note, the evidence standing of every row is now stated once in a single table, and the Layer 3 band is reported with its sensitivity rather than as a result that holds regardless of its inputs.
- **v0.2 (7 August 2026).** Added Layer 1b, the FY2025 re-base with capacity and peaking held constant. Re-tagged Microsoft Wisconsin as planning rather than measured, taking the planning rows to three across three operators.
- **v0 (published 12 July 2026, note dated 11 July, record 10.5281/zenodo.21318961).** First public version: the WCI reproduced and opened, the hydropower coupling, and the relocation ledger.