Pith. sign in

REVIEW 3 major objections 6 minor 21 references

U.S. hyperscale data centers use about 300 billion litres of water a year, mostly from the electricity grid, and the two water pathways create different hotspots that need different fixes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-12 14:53 UTC pith:2PVGXZGY

load-bearing objection Solid facility-resolved co-map of cooling vs electricity water for 472 U.S. hyperscale sites; the split-geography and concentration claims hold, even with uniform WUE and contested hydro attribution. the 3 major comments →

arxiv 2607.02531 v1 pith:2PVGXZGY submitted 2026-06-05 cs.CY cs.AI

The Hidden Water Geography of U.S. Hyperscale Data Centers in the AI Era

Gianluca Guidi , Francesca Dominici This is my paper
classification cs.CY cs.AI
keywords hyperscale data centerswater footprintScope 1 cooling waterScope 2 electricity-related waterbalancing authoritieshydrologic basinswater stressAI infrastructure
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Data-center water is usually reported as one number, but it is consumed in two places: on site for cooling, and off site when power plants generate the electricity that runs the facility. Mapping both pathways for 472 U.S. hyperscale sites shows that total operational water use is about 300 billion litres a year under baseline assumptions, with electricity-related water accounting for roughly three-quarters of that total. Direct cooling burdens cluster in stressed western and south-central river basins, while electricity-related burdens cluster in a few eastern grid regions that still rely heavily on fossil generation. Just three of twenty-four hosting balancing authorities account for 59 percent of the electricity-related water. Keeping the pathways separate shows which decisions matter where: cooling design and water sourcing at the site, and electricity planning and procurement at the regional grid level.

Core claim

For 472 U.S. hyperscale data centers, baseline operational water consumption is approximately 300 GL per year (range 205–451 across scenarios), of which electricity-related water is about 226 GL (three-quarters of the total) and on-site cooling water is about 74 GL. The two pathways produce different hotspot maps and different concentration patterns: cooling hotspots sit in stressed western and south-central basins, while electricity-related water is dominated by a few eastern balancing authorities, with three of twenty-four hosting authorities accounting for 59 percent of that pathway.

What carries the argument

Facility-resolved dual-pathway accounting: each site’s electricity demand is converted into Scope 1 cooling water via scenario WUE and PUE, and into Scope 2 electricity-related water via the balancing authority’s generation mix and technology-specific water factors; the two volumes are then mapped separately to hydrologic basins and balancing authorities.

Load-bearing premise

Every facility is given the same scenario values for utilization, power-usage effectiveness, and water-usage effectiveness, so the map of cooling-water burden mainly reflects where capacity is sited rather than measured differences in cooling technology at individual sites.

What would settle it

Replace the uniform scenario WUE values with facility-level measured WUE for a large share of the inventory and re-run the basin hotspot map; if the western and south-central cooling hotspots and the national Scope 1 total change substantially, the central geographic claim would not hold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The manuscript estimates annual operational water consumption for 472 U.S. hyperscale data centers by separating on-site cooling water (Scope 1, via scenario WUE applied to IT electricity) from electricity-related water (Scope 2, via eGRID balancing-authority generation mixes and technology-specific consumption factors). Under baseline assumptions the national total is about 300 GL yr−1 (range 205–451 across scenarios), of which roughly three-quarters is Scope 2. Mapping Scope 1 to HydroBASINS basins with Aqueduct stress and Scope 2 to balancing authorities shows divergent hotspot geographies—western/south-central stressed basins for cooling versus a small set of often eastern, fossil-heavy grid regions for electricity—and high Scope 2 concentration (top 3 of 24 hosting BAs account for 59%). The authors argue that pathway separation identifies different decision makers and levers: local cooling design and water sourcing versus regional electricity planning and procurement. Sensitivities (efficiency, high-load, no-hydro) and SI robustness figures support stability of the geographic split.

Significance. If the dual-pathway geography holds, the paper supplies a concrete, decision-relevant framing that single national water footprints obscure: cooling risk is local and basin-specific, while electricity-related water is concentrated in a few grid regions where generation mix and procurement matter most. Strengths include a facility-resolved national scope, transparent accounting equations (S1–S12), explicit scenario windows and a contested-hydro sensitivity, pathway-separated hotspot and leverage maps, and a public analysis repository (with the acknowledged commercial constraint on raw facility coordinates). The concentration statistic and the claim that the two pathways do not co-locate are falsifiable against alternative inventories and factor sets and are useful for utilities, PUCs, and permitting authorities. The work is a solid geospatial accounting contribution rather than a causal or engineering-feasibility study.

major comments (3)
  1. [S1.4; Fig. 2A; Fig. 4A; Limitations] Materials and Methods S1.4 and Limitations: all facilities receive the same scenario WUE (baseline 0.8 L kWh_IT−1), so Scope 1 spatial pattern is almost entirely nameplate-capacity distribution. The Limitations paragraph states this correctly, but main-text language around Fig. 2A and Fig. 4A still reads as “cooling burden hotspots.” Please revise those captions and the corresponding prose to state explicitly that Scope 1 maps show siting/capacity exposure under uniform WUE, not measured site technology differences, and note how this bounds interpretation of basin ranks.
  2. [S1.5; Eq. (S8); Implications] S1.5 and Eq. (S8): Scope 2 uses annual average BA generation shares. Hyperscale loads are high-capacity-factor and may be served by a different marginal mix than the annual average, which can change water intensity especially in hydro- and coal-heavy BAs. A short discussion (or bounding calculation) of average vs. marginal attribution would strengthen the claim that the eastern fossil-heavy concentration is robust to dispatch assumptions, not only to the no-hydro factor set.
  3. [Data availability; Tables S3–S4; Fig. 3] Data and materials availability / Tables S3–S4: raw facility identifiers and coordinates are withheld. Concentration (top 3/24 BAs = 59%) and basin ranks are central results. Please add SI tables listing all hosting BAs and all hosting basins with MW, Scope 1, Scope 2, and totals under each scenario (not only top-15), so the concentration curves and rank-stability claims in Figs. 3, S3, and S4 can be independently recomputed without the proprietary inventory.
minor comments (6)
  1. [Abstract; Table S2] Abstract and main text: “approximately 300 GL yr−1 (range 205–451)” — Table S2 reports 204.91–450.60; align the rounded range consistently (205–451 vs 205–451 is fine if stated as rounded).
  2. [Eq. (S2)] Eq. (S2): the 10−3 conversion is described as litres to cubic metres, but national results are reported in GL; a one-line unit chain (L → m3 → GL) would reduce reader friction.
  3. [Fig. 2; Fig. 4] Fig. 2 and Fig. 4 hatching for stress is useful but dense in grayscale; ensure the print version remains legible or add a non-hatch redundant encoding.
  4. [Main Text, external check paragraph] Partial external check (Google/Microsoft/Meta ~25 GL direct water) is helpful; state the approximate capacity share used for scaling so readers can reproduce the consistency claim.
  5. [S1.2; Ref. [19]] Reference [19] is a companion arXiv on carbon/energy for the same inventory; briefly clarify independence of the water pipeline from that paper’s electricity totals to avoid double-counting concerns.
  6. [Table S4] Typo/consistency: “TV A” vs “TVA” in Table S4; standardize BA abbreviations with eGRID codes.

Circularity Check

0 steps flagged

No significant circularity: open geospatial accounting with external water factors, eGRID mixes, and scenario parameters; self-citation supplies inventory data only.

full rationale

The paper is a facility-resolved national water-accounting exercise, not a first-principles derivation that collapses into its own fitted target. Scope 1 is W^(1)_i = 10^{-3} WUE_i E_IT with E_IT = E_fac/PUE and E_fac = P_i × 8760 × u; Scope 2 is W^(2)_i = 10^{-3} I_grid_r(i) E_fac with I_grid_r = Σ s_r,f w_f. Utilization, PUE, and WUE are scenario windows drawn from external literature (ISO/IEC 30134-9, Shehabi et al., Lei & Masanet, etc.), not fitted to the paper’s own water totals. Technology water factors w_f come from Macknick et al. and Meldrum et al.; generation shares from eGRID; stress from Aqueduct; basins from HydroBASINS. The only author-overlapping citation that is load-bearing for inputs is [19], which supplies the hyperscale facility inventory and nameplate power—data infrastructure, not a uniqueness theorem or a quantity being re-predicted. Hotspot definitions (above-median burden + stress/fossil screens) and concentration statistics (top 3 of 24 BAs = 59% of Scope 2) are direct aggregations of those computed volumes; they are not forced by construction from a prior fit. Scenario and no-hydro sensitivities rescale totals but do not redefine the pathway split. No self-definitional loop, fitted-input-as-prediction, uniqueness import, or renamed known result is present. Score 0 is appropriate.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central geographic claim rests on standard water-footprint accounting, published intensity factors, public grid and hydrology layers, and a set of scenario parameters chosen from literature ranges rather than fitted to the water totals themselves. No new physical entities are postulated. The main free parameters are the shared operating assumptions (u, PUE, WUE) and the technology water factors, especially the contested hydropower value.

free parameters (5)
  • annual utilization u = 0.66 (baseline)
    Scenario window 0.55 / 0.66 / 0.85; baseline 0.66 multiplies nameplate MW into annual electricity and therefore scales both Scope 1 and Scope 2.
  • PUE = 1.25 (baseline)
    Scenario window 1.15 / 1.25 / 1.40; used only to recover IT electricity for WUE-based Scope 1.
  • WUE = 0.8 L/kWh_IT (baseline)
    Scenario window 0.2 / 0.8 / 1.5 L kWh_IT^-1 applied uniformly to all facilities; primary driver of absolute Scope 1 magnitude.
  • hydropower water-consumption factor = 8.0 L/kWh (baseline)
    Baseline 8.0 L kWh^-1 from Macknick et al.; contested attribution of reservoir evaporation; no-hydro sensitivity sets it to zero and cuts Scope 2 by ~43%.
  • thermoelectric and other fuel water factors = literature set (Table S1)
    Coal 1.9, gas 0.7, nuclear 2.5, wind 0.0, solar 0.1, other 1.0 L kWh^-1 taken from literature reviews and held fixed.
axioms (5)
  • domain assumption Location-based (not market-based) attribution: each facility inherits the average physical water intensity of its balancing-authority generation mix.
    Stated in S1.5; deliberately ignores contractual PPAs so that physical grid water is measured.
  • domain assumption Aqueduct baseline water stress ≥ 3 (high to extremely high) defines stress-qualified basins and high-stress BA overlays.
    Hotspot definitions in S1.7 and main-text Fig. 2 rely on this external threshold.
  • domain assumption HydroBASINS level-6 polygons and eGRID balancing-authority polygons correctly assign facilities to hydrologic and grid units.
    Geospatial joins in S1.2; facilities missing valid joins are dropped for the affected layer.
  • domain assumption Operational water boundary excludes embodied and supply-chain water; annual averages suffice for national geography.
    Explicit boundary statement in S1.1 and Limitations; seasonal drought coincidence is out of scope.
  • ad hoc to paper Nameplate MW times utilization yields annual facility electricity; PUE is not used to scale facility electricity.
    S1.3 facility-load interpretation; consistent with the companion inventory paper but is a modeling choice.

pith-pipeline@v1.1.0-grok45 · 19260 in / 3309 out tokens · 30998 ms · 2026-07-12T14:53:46.365811+00:00 · methodology

0 comments
read the original abstract

Water use by data centers is routinely reported as a single footprint, but water is consumed through two physically distinct pathways: at the site for cooling and in the power system that generates electricity. We mapped both pathways for 472 U.S. hyperscale facilities by linking facility locations to electricity regions, hydrologic basins, and water-stress data. Under baseline assumptions, operational water consumption totals approximately 300 GL yr^-1 (range 205-451 across scenarios), with electricity-related water contributing three-quarters of the total. The two pathways produce different hotspot geographies: direct cooling burdens concentrate in stressed western and south-central basins, whereas electricity-related burdens concentrate in a few eastern grid regions with fossil-heavy supply. Just 3 of 24 hosting balancing authorities account for 59% of electricity-related water. Separating pathways identifies which decisions matter where: cooling design and water sourcing locally, electricity planning and procurement regionally

Figures

Figures reproduced from arXiv: 2607.02531 by Francesca Dominici, Gianluca Guidi.

Figure 1
Figure 1. Figure 1: Study design. Facility inventory, grid generation mix, and hydrologic stress data are linked to each hyperscale facility. Facility electricity demand is translated into two water pathways: Scope 1, consumed at the site and mapped to hydrologic basins, and Scope 2, consumed in the power system and mapped to balancing authorities. Pathway-separated estimates support hotspot maps, concentration analysis, and … view at source ↗
Figure 2
Figure 2. Figure 2: Direct cooling and electricity-related water produce different hotspot maps. A, Scope 1 basin classes combine direct cooling-water burden (above-median among hosting basins) with basin water stress (Aqueduct score ≥ 3 in the basin or a touching neighbor). B, Scope 2 balancing-authority classes combine electricity-related water burden (above-median among hosting BAs), fossil-heavy supply (coal + gas >50%), … view at source ↗
Figure 3
Figure 3. Figure 3: Electricity-related water is more spatially concentrated than direct cool￾ing water. A, Scope 1 concentration across hydrologic basins hosting hyperscale facilities. B, Scope 2 concentration across balancing authorities hosting hyperscale facilities. Bars rank regions by their share of total baseline pathway-specific water; maps show the leading regions. Scope 2 reaches half the national total in far fewer… view at source ↗
Figure 4
Figure 4. Figure 4: Different actions would reduce water use most in different places. A, Potential Scope 1 reduction by hydrologic basin under a lower-water operating benchmark. Large values indicate both high current direct cooling water and a large gap to the bench￾mark. Hatching marks high water stress (≥ 3). Audience: local water managers, facility op￾erators, permitting authorities. Action: less water-intensive cooling,… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 4 canonical work pages

  1. [1]

    Data centre water consumption.npj Clean Water, 4(1):11, 2021

    David Mytton. Data centre water consumption.npj Clean Water, 4(1):11, 2021. doi: 10.1038/s41545-021-00101-w. URLhttps://doi.org/10.1038/s41545-021-00101-w

    doi:10.1038/s41545-021-00101-w 2021

  2. [2]

    ISO/IEC 30134-9: Information technology — data centres — key performance indicators — part 9: Water usage effectiveness (WUE), 2022

    International Organization for Standardization and International Electrotechnical Com- mission. ISO/IEC 30134-9: Information technology — data centres — key performance indicators — part 9: Water usage effectiveness (WUE), 2022. Standard defining WUE for data centers

    2022

  3. [3]

    Environmental Protection Agency

    U.S. Environmental Protection Agency. WaterSense at work: Best management practices for commercial and institutional facilities. https://www.epa.gov/watersense/watersense-work, 2012. Cooling-tower and water-management guidance for facilities

    2012

  4. [4]

    Hal- lett

    Jordan Macknick, Robin Newmark, Garvin Heath, and Kimberly C. Hal- lett. Operational water consumption and withdrawal factors for electricity generating technologies: a review of existing literature.Environmental Re- search Letters, 7(4):045802, 2012. doi: 10.1088/1748-9326/7/4/045802. URL https://doi.org/10.1088/1748-9326/7/4/045802

    work page doi:10.1088/1748-9326/7/4/045802 2012

  5. [5]

    Life cycle water use for electricity generation: a review and harmonization of literature estimates.Environmental Research Letters, 8(1):015031, 2013

    Jordan Meldrum, Syndi Nettles-Anderson, Garvin Heath, and Jordan Macknick. Life cycle water use for electricity generation: a review and harmonization of literature estimates.Environmental Research Letters, 8(1):015031, 2013. doi: 10.1088/1748- 9326/8/1/015031

    doi:10.1088/1748- 2013

  6. [6]

    Islam, and Shaolei Ren

    Pengfei Li, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren. Making AI less “thirsty”: Uncovering and addressing the secret water footprint of AI models.arXiv preprint arXiv:2304.03271, 2023. URLhttps://arxiv.org/abs/2304.03271

    Pith/arXiv arXiv 2023

  7. [7]

    Islam, and Shaolei Ren

    Pengfei Li, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren. Making AI less “thirsty”: Uncovering and addressing the secret water footprint of AI models.Commu- nications of the ACM, 2025. doi: 10.1145/3724499. 7

    doi:10.1145/3724499 2025

  8. [8]

    The water footprint of data centers

    Branislav Risti´ c, Kaveh Madani, and Zen Makuch. The water footprint of data centers. Sustainability, 7(8):11260–11284, 2015. doi: 10.3390/su70811260

    work page doi:10.3390/su70811260 2015

  9. [9]

    United states data center energy usage report

    Arman Shehabi, Sarah Smith, Dale Sartor, Richard Brown, Magnus Herrlin, Jonathan Koomey, Eric Masanet, Nathaniel Horner, Inˆ es Azevedo, and William Lintner. United states data center energy usage report. Technical Report LBNL-1005775, Lawrence Berkeley National Laboratory, 2016

    2016

  10. [10]

    Aqueduct water risk atlas

    World Resources Institute. Aqueduct water risk atlas. https://www.wri.org/aqueduct, 2023. Aqueduct 4.0 baseline water-stress indi- cators

    2023

  11. [11]

    Samantha Kuzma, Marc F. P. Bierkens, Sridevi Lakshman, Tianyi Luo, Liz Saccoccia, Edwin H. Sutanudjaja, and Rens Van Beek. Aqueduct 4.0: Updated decision-relevant global water risk indicators. Technical note, World Resources Institute, Washington, DC, 2023. URLhttps://doi.org/10.46830/writn.23.00061

    work page doi:10.46830/writn.23.00061 2023

  12. [12]

    Climate- and technology-specific PUE and WUE estima- tions for U.S

    Nuoa Lei and Eric Masanet. Climate- and technology-specific PUE and WUE estima- tions for U.S. data centers using open data.Resources, Conservation and Recycling, 182:106323, 2022. doi: 10.1016/j.resconrec.2022.106323

    doi:10.1016/j.resconrec.2022.106323 2022

  13. [13]

    Environmental Protection Agency

    U.S. Environmental Protection Agency. eGRID: Emissions & generation resource in- tegrated database.https://www.epa.gov/egrid, 2024. Power-plant emissions and generation data used for grid-based attribution

    2024

  14. [14]

    Mayer, William B

    Peter W. Mayer, William B. DeOreo, Eva M. Opitz, Jack C. Kiefer, William Y. Davis, Benedykt Dziegielewski, and John Olaf Nelson. Residential end uses of water.American Water Works Association Research Foundation, 1999

    1999

  15. [15]

    Mekonnen and Arjen Y

    Mesfin M. Mekonnen and Arjen Y. Hoekstra. The blue water footprint of electricity from hydropower.Hydrology and Earth System Sciences, 16(1):179–187, 2012. doi: 10.5194/hess-16-179-2012

    work page doi:10.5194/hess-16-179-2012 2012

  16. [16]

    2024 environmental report

    Google. 2024 environmental report. Technical report, Alphabet Inc., 2024. URL https://sustainability.google/reports/google-2024-environmental-report/

    2024

  17. [17]

    2024 environmental sustainability report

    Microsoft. 2024 environmental sustainability report. Tech- nical report, Microsoft Corporation, 2024. URL https://www.microsoft.com/en-us/corporate-responsibility/sustainability/report

    2024

  18. [18]

    2024 sustainability report

    Meta. 2024 sustainability report. Technical report, Meta Platforms Inc., 2024. URL https://sustainability.fb.com/report/

    2024

  19. [19]

    Bargagli-Stoffi

    Gianluca Guidi, Francesca Dominici, Tiziano Squartini, Callaway Sprinkle, Jonathan Gilmour, Kevin Butler, Eric Bell, Scott Delaney, and Falco J. Bargagli-Stoffi. Assessing the carbon emissions and energy consumption of u.s. hyperscale data centers, 2026. URLhttps://arxiv.org/abs/2606.05420. 8

    Pith/arXiv arXiv 2026

  20. [20]

    Global river hydrography and network rout- ing: baseline data and new approaches to study the world’s large river systems

    Bernhard Lehner and G¨ unther Grill. Global river hydrography and network rout- ing: baseline data and new approaches to study the world’s large river systems. Hydrological Processes, 27(15):2171–2186, 2013. doi: 10.1002/hyp.9740. URL https://doi.org/10.1002/hyp.9740. HydroBASINS / HydroSHEDS lineage; used for basin polygon identifiers

    doi:10.1002/hyp.9740 2013

  21. [21]

    Koomey, Eric R

    Arman Shehabi, Sarah Josephine Smith, Alex Hubbard, Alexander Newkirk, Nuoa Lei, Md Abu Bakar Siddik, Billie Holecek, Jonathan G. Koomey, Eric R. Masanet, and Dale A. Sartor. 2024 United States Data Center Energy Usage Report. Tech- nical Report LBNL-2001637, Lawrence Berkeley National Laboratory, 2024. URL https://doi.org/10.71468/P1WC7Q. Prepared to mee...

    doi:10.71468/p1wc7q 2024