Local resources · resource ledger

Water

Water use is real and intensely site-specific; prompt-level slogans are not measurements.

Contested>10,000×observed workload-level water-use variation

What the evidence supports

LBNL’s review found more than 10,000-fold variation in workload-level water use. GAO says company reporting is insufficient to cleanly separate generative AI from other data-center activity.

Mechanism: Water can be consumed directly by cooling and indirectly by electricity generation. Climate, cooling design, utilization and grid mix dominate the result.
Who pays—or gains: Water-stressed communities bear opportunity cost when new demand competes with existing users; operators can reduce it through siting, cooling and reclaimed water.
Binding constraint: Transparent facility-level reporting and the local physical availability of water—not a universal liters-per-prompt constant.
Strongest caveat: A global average cannot resolve whether one project burdens one watershed.
What would change the grade: Grade individual campuses when permits, source water, consumptive use and seasonal operations are disclosed.

Evidence file

Primary and first-party sources

Test the claims behind this market