A 0.12 L/kWh figure lands alongside a 2.5 billion gallon 2025 disclosure. The cross vendor table is the most useful part of the story, and the part most likely to mislead.
Amazon is now on the record with a water-efficiency number that, on its face, makes the company look like the sober sibling among the Big Four hyperscalers. The disclosure, reported by The Register, puts Amazon's datacenter fleet at 0.12 liters of water per kilowatt-hour in 2025, with a global draw of roughly 2.5 billion gallons for the year. The figure matters less as a brag and more as a yardstick: it is the kind of per-kWh number that local officials reviewing a permit, a procurement team vetting a vendor, or a journalist tracking Big Tech water risk can actually put to work.
Liters per kilowatt-hour measures how many liters of water a datacenter pulls to deliver one kilowatt-hour of IT load. Most of that water does not touch a chip. It cycles through cooling towers and evaporative systems that keep servers from overheating, which is why the metric is denominated in cooling water per unit of compute rather than in raw consumption. The lower the number, the less water a facility burns for each unit of useful work.
Amazon is also the only one of the four that has published a per-kWh figure of its own in a single comparable format, and it used that opening to rank itself against the others. According to the Amazon disclosure as summarized by The Register, the company places Microsoft at 0.27 L/kWh in 2025, Meta at 0.19 L/kWh in 2024, and Google at 1.15 L/kWh. Read in isolation, those numbers would make Amazon roughly twice as efficient as Meta, more than twice as efficient as Microsoft, and almost an order of magnitude ahead of Google. The table is also where the story gets genuinely useful, because it is also where it stops being clean.
The first thing to flag is the year mismatch. Amazon is reporting 2025, Microsoft is being cited for 2025, Meta for 2024, and Google for an ambiguous "same year" reference inside Amazon's own framing. Water efficiency moves quickly when a fleet is being expanded, retrofitted, or shifted geographically, which means a one-year gap can carry more weight than the headline ratio suggests. The second caveat is that the 0.12 figure is Amazon's own measurement of its own fleet, and the comparative figures are Amazon's reading of its rivals' public disclosures, not rival-issued statements. As of The Register's filing, Microsoft, Meta, and Google had not yet responded to comment requests, so the table is, for now, one company's case for why its number is the lowest.
There is also a methodological layer that does not show up in a single L/kWh column. Water usage effectiveness definitions vary across hyperscalers, and the boundary between operational water and replenishment water, between on-site cooling and indirect grid water, is not standardized. A reader who wants to use the table to compare vendors is, in practice, comparing four different reporting choices, not four like-for-like measurements. Amazon is publishing the most legible number. That is a real contribution, but it is not the same as a neutral benchmark.
The 2.5 billion gallons is the figure that the lawn-and-garden comparison invites. Amazon has contextualized the total against U.S. residential outdoor water use on a 350 million population baseline, the kind of comparison that makes a datacenter's draw look either trivial or enormous depending on the denominator. The Register notes that Amazon's 2025 framing leans into that contrast, and it is worth reading it as a rhetorical move rather than a neutral benchmark. A gallon of datacenter cooling water is not the same as a gallon on a lawn, and the two draw on different aquifers, different stress contexts, and different governance. The per-kWh number is the more useful unit for cross-vendor comparison precisely because it normalizes for fleet size and shape.
The "water positive by 2030" pledge sits behind both numbers. Amazon says it is 75 percent of the way there, four years ahead of schedule, on a target the company announced in 2022 to return more water to communities than it consumes through rainwater capture, wastewater reuse, and replenishment projects. "Water positive" is Amazon's own definition of its own progress, a self-measured metric on a self-set timeline, and the original 2022 methodology and scope may have shifted in subsequent sustainability reports. That is not an accusation. It is a flag for anyone who wants to weigh the 75 percent claim against the 0.12 L/kWh claim and the 2.5 billion-gallon total.
The disclosure also lands in a year when community pushback against datacenters has become measurable, not anecdotal. Surveys on U.S. NIMBY sentiment toward new builds, the long-running fight in The Dalles, Oregon, where a single town hosts a meaningful share of a hyperscaler's water and electricity draw, and broader concerns about noise, land use, and grid stress have moved local permitting from a formality into a contested arena. A 0.12 L/kWh figure that procurement teams can benchmark, and a 2.5 billion-gallon total that local officials can divide by acres and townships, are now part of the public record. Whether the number is the right number is a separate question. Whether it can be checked against the other three hyperscalers in the same column, with the same units and the same year, is a question the next round of disclosures will have to answer.
The thing worth watching is not the next press release. It is whether Microsoft, Meta, and Google publish their own 2025 L/kWh numbers in a comparable format, on a comparable timeline, with a comparable definition of what counts. Until then, Amazon's 0.12 is the only number on the leaderboard that the company itself put there. The comparison is real, but the comparator set is not yet complete.