Data Centers' Resource Bill: Water, Power, and the Energy Bottleneck

The UAE: power, water, and emissions

A preprint study of rapid data-center expansion in the United Arab Emirates finds that expansion could consume up to 11.5 percent of national electricity generation by 2040, a share that makes data centers a national-scale energy planning question rather than a facilities question. m:…7101

The study quantifies the trade-offs: dry-cooling configurations reduce water demand by 98.2 percent but incur an 11 to 13 percent electricity penalty, and system-wide greenhouse gas emissions increase by 2.27 to 19.9 million tonnes of CO2 equivalent annually when data-center demand is met by natural-gas-dominated grids. m:…7101

The implication the numbers carry is that cooling choices are energy choices: saving water costs power, and meeting the resulting power demand with gas raises emissions, so the three resources have to be planned together rather than optimized one at a time. m:…7101

Meta's water: measuring beats scaling rules

A second preprint, studying water withdrawal at US Meta data centers, finds that electricity-scaling rules for estimating annual water withdrawal consistently overestimate actual usage, and that persistence models provide lower error in annual estimates than those scaling rules. m:…8279

The study also finds that the withdrawal-to-electricity ratio for the studied data centers declined annually between 2017 and 2023, which means water use per unit of electricity fell over that period, a trend that scaling rules assuming a fixed ratio would miss. m:…8279

For planners, the finding is methodological: better measurement beats a convenient rule of thumb, and the UAE study's wide emissions range shows why, since grid mix and cooling choice move the outcome more than any single headline number. m:…8279 m:…7101

The bottleneck warning

Naveen Rao, speaking on the Twenty Minute VC podcast, argued that energy supply is the primary bottleneck for AI scaling, with current infrastructure growth rates insufficient to meet demand, and noted that building one gigawatt of energy infrastructure costs approximately 50 billion dollars. m:…6264

Rao added that China has surpassed the United States in total energy capacity as of 2023, which frames the bottleneck as a competitive question as well as an engineering one: the country or company that can add power fastest sets the pace for AI expansion. m:…6264

Together, the preprints and the podcast argument suggest the buildout's binding constraint is physical: water and power have to be measured, cooled, and generated, and the cost of a gigawatt gives a scale for what closing the gap would require. m:…7101 m:…8279 m:…6264

Measure before you scale

The energy studies agree on method even where their settings differ: the UAE study's wide emissions range under different grid and cooling choices and the Meta study's finding that persistence models beat electricity-scaling rules both argue for measuring actual water and power use rather than assuming fixed ratios. m:…7101 m:…8279

Rao's gigawatt price tag turns that measurement advice into a budget: at approximately 50 billion dollars per gigawatt, and with energy as the primary bottleneck, the decision to expand data-center capacity is a decision to finance power infrastructure on a scale comparable to the compute itself. m:…6264

Cooling is a strategy choice

Cooling deserves treatment as a strategy choice rather than a facilities detail: dry cooling cuts water demand by nearly all of it while adding an electricity penalty, and meeting that added demand with gas raises emissions, so a data-center operator choosing a cooling system is simultaneously choosing a water profile, a power profile, and an emissions profile for the life of the facility. m:…7101