Does electricity scaling improve annual water-withdrawal estimates? Evidence from US Meta data centers

Jaivime Evaristo 

Planners often need to estimate data-center water use before detailed operating data exist, and a common shortcut multiplies electricity use by a historical water intensity. The logic is appealing: computing produces heat, and removing heat often takes water. Yet whether this shortcut beats the simplest alternative, persistence (assuming that each facility withdraws what it withdrew the year before), remains an open question. Here we test three electricity-scaling rules against persistence, using operator-reported withdrawal and electricity for 15 US Meta data-center sites in 2016-2024, cross-checked across eleven report versions. In 64 year-to-year comparisons at 13 sites, persistence had the lowest pooled error (a weighted absolute percentage error of 30.9%, against 41.6-56.2%), even though the scaling rules were given each target year’s observed electricity. All three scaling rules overestimated withdrawal on average. The ranking was not fixed, however: scaling by a site’s own previous-year ratio had lower error in three of eight target years. And in an exploratory comparison, more history made the pooled rules worse, not better: because the withdrawal-to-electricity ratio fell every year from 2017 to 2023, histories that began in 2016 carried outdated ratios forward and raised errors on 2021-2024 targets by 5.4-7.0 percentage points. These retrospective results come from one operator and establish neither a universal ranking nor forecasting skill. But they set a simple bar for anyone converting electricity projections into water demand: test the conversion against last year’s withdrawal, year by year as well as on average.