Amazon strategised about keeping water use secret (www.source-material.org)

A leaked 2022 internal Amazon memo shows the company debated keeping the true scale of its data-centre water use secret, advising AWS to report only “primary” water consumption (7.7 billion gallons projected by 2030) while excluding “secondary” water used to generate the electricity that powers servers. The document framed broader disclosure as a “one-way door” that could invite reputational headlines and double the campaign scope and budget; AWS’ Water Positive pledge aimed to cut primary use to 4.9 billion gallons by 2030 largely through replenishment “offsets” (about $109M, half of which the memo says would have been spent anyway). Amazon says the memo is obsolete and that AWS has improved efficiency—claiming a 40% reduction in water per kilowatt since 2021—but declined to release fuller secondary or scope-3 figures (the company estimated scope-3 comprises roughly 90% of Amazon’s total water footprint). For the AI/ML community, the leak matters because rapid AI-driven compute growth is driving massive data-centre expansion—often in drought-prone regions—and accurate water accounting affects infrastructure planning, regulatory risk, and the true environmental cost of training large models. Technically, the case highlights how measurement choices (primary vs. secondary vs. scope‑3) and reliance on offsets can materially understate water impacts; it also underscores industry efforts to shape methodologies. Researchers, cloud customers, and policymakers should press for standardized, transparent water-footprint metrics tied to energy and supply‑chain emissions to make sustainable model deployment decisions meaningful rather than cosmetic.