Methodology

What we model

For each data center we publish a modeled annual electricity consumption (GWh/yr) and annual water withdrawal (US gallons/yr). The model is intentionally simple — a deterministic function of facility IT load, PUE, water-use effectiveness (WUE), and utilization. We chose simplicity over apparent precision because hyperscale operators do not publish the inputs a more complex model would need.

Formula

kwh_per_year = it_load_mw × 1000 × utilization × pue × 8760
gwh_per_year = kwh_per_year / 1,000,000

gallons_per_year = kwh_per_year × wue × climate_factor × 0.264172  # L→US gal

• it_load_mw — IT-load capacity in megawatts. Sourced from operator disclosure where possible; back-derived from disclosed annual MWh when the operator publishes that but not capacity (see Calibration below).
• pue — Power Usage Effectiveness. Use the operator's design_pue when published; otherwise the tenant-type default below.
• utilization — Fraction of peak IT load actually drawn, fleet-averaged. Default 0.6.
• wue — Water Use Effectiveness, L/kWh. Use the operator's reported_wue when published; otherwise the cooling-type default below.
• climate_factor — Multiplier on evaporative WUE for local wet-bulb temperature. Currently 1.0 for all facilities. NOAA/PRISM integration is a tracked follow-up.

Default assumptions

When an operator does not publish a value, we fall back to these tenant-type or cooling-type defaults. They are intentionally conservative and round-number where possible so the source of any displayed estimate is easy to audit.

Uncertainty bands

Every facility carries a confidence label that drives the displayed uncertainty band on its modeled values:

Calibration vs operator disclosures

We tune the model's defaults so that the modeled annual MWh round-trips to operator- disclosed annual MWh for every facility where the operator publishes both. The table below is regenerated on every build from data/facilities/ and gated in CI — any facility outside ±5% blocks merges.

Rows stamped with an earlier year (e.g. Eagle Mountain 2022) reflect the most recent publicly disclosed figure; the operating campus may have grown since. We refresh `reported_annual_mwh` on each curation pass when a newer disclosure is available.

4 of 338 facilities currently carry an operator- disclosed annual MWh figure. 5 have a sufficient it_load_mw to compute an estimate (the others render on the map but without a modeled energy/water number until a load value is sourced).

Where this falls short

• Utilization is fleet-averaged, not site-specific. Some campuses run closer to 80% utilization on AI training workloads; others coast at 40%.
• Climate factor is uniform. A facility in Phoenix and a facility in Quincy with the same nominal WUE will produce different actual water consumption — our model currently treats them the same.
• Backup-generator and on-site solar capacity ≠ IT load. We exclude both from load calculations even when they're widely reported.