Dataset Viewer
Auto-converted to Parquet Duplicate
operator
stringclasses
4 values
location
stringclasses
10 values
W_d_ML_d
float64
0.09
14.6
r
float64
0.58
0.83
PF
float64
2.21
30
K_ML_d
float64
7.6
52
stated_WCI
float64
0.18
1.86
grid_region
stringclasses
10 values
reservoir
stringclasses
2 values
dam
stringclasses
2 values
double_coupled_candidate
stringclasses
2 values
source_tag
stringclasses
8 values
motive_tier
stringclasses
3 values
cooling
stringclasses
2 values
WUE_L_per_kWh
float64
0.2
1.8
EWIF_lo
float64
0.8
5
EWIF_hi
float64
1.5
15
W_d_2025_ML_d
float64
2.73
18.1
r_2025
float64
0.6
0.84
vintage_note
stringclasses
6 values
range_flag
stringclasses
1 value
Meta
Lebanon, IN
4.81
0.77
6.3
30.3
0.77
MISO-Indiana
null
null
null
ren_seed_NOT_operational_closedloop
planning_not_measured
closed_loop
0.2
1.8
3
null
null
null
null
Google
Council Bluffs, IA
14.62
0.716
6.5
36.6
1.86
MISO-Iowa
null
null
null
google_fy2024_verified
primary
evaporative
1.8
1.5
3
18.112
0.7707
Google 2026 Env Report, FY2025: 1,746.4 M gal withdrawn / 1,346.0 consumed
null
Google
Mayes Co., OK
11.49
0.752
2.5
52
0.44
SPP-Oklahoma
null
null
null
utility_openrecords_2024_verified
primary
evaporative
1.8
0.8
1.5
14.485
0.7743
Google 2026 Env Report, FY2025: 1,396.7 / 1,081.4
null
Google
The Dalles, OR
4.78
0.784
2.21
45.4
0.18
Columbia-River
Columbia-River
The-Dalles-Dam
Y_primary
google_fy2024_verified
primary
evaporative
1.8
5
15
6.786
0.7168
Google 2026 Env Report, FY2025: 654.3 / 469.0
null
Google
Douglas Co., GA
4.6
0.826
2.5
22.7
0.41
Southern-Co-GA
null
null
null
google_fy2024_reclaimed_verified
primary
evaporative
1.8
1.8
3.5
4.569
0.8332
Google 2026 Env Report, FY2025: 440.6 / 367.1
null
Microsoft
Wisconsin
0.087
0.77
30
7.6
0.34
MISO-Wisconsin
null
null
null
ren_table6
planning_not_measured
closed_loop
0.2
1.8
3
null
null
null
null
Google
Botetourt Co., VA
7.57
0.77
2.5
10.2
1.45
PJM-Dominion-VA
null
null
null
contracted_2MGD_NOT_operational
planning_not_measured
evaporative
1.8
1.8
3
null
null
null
null
xAI
Memphis, TN
3.79
0.77
4.5
22.7
0.57
TVA-Tennessee
null
null
null
ren_seed_W_in_primary_range_multisite
framework_secondary
evaporative
1.8
2
3.5
null
null
null
0.81-7.0 MGD across Colossus 1 and 2; the static row is Ren's single-site value and is indicative only
Google
Midlothian, TX
2.29
0.825
2.5
15.1
0.31
ERCOT-Texas
null
null
null
google_fy2024_verified
primary
evaporative
1.8
0.8
1.5
2.734
0.8354
Google 2026 Env Report, FY2025: 263.6 / 220.2
null
Google
Henderson, NV
3.73
0.576
2.5
15.2
0.36
Colorado-River
Lake-Mead
Hoover-Dam
Y_source_confirmed
google_fy2024_verified_city_corroborated
primary
evaporative
1.8
2
6
4.166
0.596
Google 2026 Env Report, FY2025: 401.7 / 239.4
null

AI Data-Centre Water Tracker

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.

An open, reproducible referee for the water burden of AI data centres. It rebuilds the Water Consumption Impact index for ten sites from public data, self-checks each value against its source, and adds the two channels the original framework leaves open: the hydropower coupling, and the off-site relocation of water that closed-loop cooling produces. Figures are bands and decompositions, every input named and tagged by the standing of its source.

Files

  • sites.csv (10 rows): the per-site inputs and rebuilt index, FY2024 and FY2025. Columns include operator, location, W_d_ML_d (daily water, ML/day), r, PF, K_ML_d, stated_WCI, grid_region, reservoir, dam, double_coupled_candidate, source_tag, motive_tier, cooling, WUE_L_per_kWh, EWIF_lo, EWIF_hi, W_d_2025_ML_d, r_2025.
  • reproduce.py: standard-library reproducer that rebuilds the index from sites.csv. Exit code is 1 whenever a row is flagged, which on the seed as shipped is always: two of the ten rows do not reproduce within tolerance, and those flags are part of the finding.
  • SOURCES.md: the source ledger. Per-row provenance, the inputs that are estimates rather than site measurements, the Layer 2 and Layer 3 sources, and the known limits.
  • LICENSE: Creative Commons Attribution 4.0 International.

Findings

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. The direction holds and the band does not: raising the closed-loop WUE from 0.2 to 0.3 L/kWh gives 88 to 92 per cent, though even at five times the assumed WUE the share is still 68 to 78 per cent.

Three of the ten seed sites, across three different operators, are non-operational planning figures presented as measured, and are excluded from the measured set.

Recomputing the index on the operators' FY2025 reports, holding capacity and peaking constant, moves five of the six primary-verified sites up by 15.6 to 33.3 per cent in a single reporting year. The sixth is flat on total withdrawal while its draw on potable municipal supply rose more than sixfold.

Method and standing of evidence

Each site's index value is rebuilt from public data and checked against its source, then decomposed into on-site use and the off-site relocation that closed-loop cooling moves to the grid. The six Google rows are traced to the operator's own EY-assured environmental reports. The remaining four rest on records releases, supply contracts and reporting, and are excluded from the measured set for that reason. The relocation ledger is modelled from a published grid-average intensity factor, not metered per facility. SOURCES.md grades every row and names every estimate.

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). 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). The author therefore treats neither lab as neutral on the buildout. All operators are scored the same way.

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 rather than presented as measurements. A language model's output can be wrong in ways that survive review; in code it simply runs, and a wrong constant or a mis-set filter still returns a clean number. The author does not claim the review is exhaustive, and 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.

Citation

NM AI Research (2026). AI Data-Centre Water Tracker. Zenodo. https://doi.org/10.5281/zenodo.21318960 . Licensed CC BY 4.0.

Downloads last month
79