US datacenters tripled their water use in a decade, and that was before the AI boom

US datacenters tripled their water footprint in the decade that ended just before the current AI buildout. The figure comes from The Register's reporting on a new analysis of US datacenter water use, and it lands as a surprise precisely because the water conversation has spent the last three years focused on the hyperscale AI campuses in Phoenix, Quincy, and central Texas. The 3x growth is older than that story, and it came from a fleet the AI debate does not see.

A 3x curve in a decade is the kind of growth that takes a city with it. Put in volume terms, the additional water the data-economy footprint now draws is on the order of a mid-sized American municipality's annual consumption, enough to fill hundreds of Olympic swimming pools in a single day. Per user, the cost is still modest: a handful of gallons a day for each American's slice of the data economy, less than a single load of laundry. The aggregate is the problem, and the aggregate is what utilities and water authorities are now staring at.

The phrase "datacenter water footprint" hides more than it shows. The direct count is the cooling towers, the chilled-water loops, and the adiabatic systems on site, water that enters the building and leaves as vapor or blowdown. The indirect count is larger. It includes the water burned, boiled, or evaporated at the power plant that feeds the datacenter, because electricity is water-cooled at almost every fossil and nuclear plant on the grid. Add the water that fabs use to make the silicon, and the number grows again. The Register's analysis uses the broader definition. Most AI-water coverage does not, which is one reason the headline 3x figure feels larger than the public debate.

And the 3x is the pre-AI baseline. The dataset the article leans on stops before the 2023-2026 generative-AI buildout, the period when Microsoft, Google, Meta, and Amazon all announced multi-gigawatt campus expansions. The current footprint is plausibly worse. The next decade will not be a linear continuation of the prior one, because the workload mix has shifted from search, email, and streaming toward training runs, inference, and agentic traffic, each of which pulls more electricity per query than the workloads they replace.

Most of the 3x growth, the article reports, came from the in-house and enterprise datacenter stock, not the hyperscalers whose names fill the AI headlines. Enterprise datacenters are the private server rooms, on-prem installations, and corporate IT buildings that house most of the world's stored data, largely outside public view. They are the older, less efficient, less visible fleet that has been the workhorse of the data economy since before cloud computing went mainstream. The implication is direct. The AI-water story is a second act on top of a trend that was already steep, and most of the actors responsible for the trend are not the ones in the news.

The Register's piece is a teaser for a larger study, and the underlying baseline year, methodology, and the share of growth attributable to the enterprise fleet are not visible in the published excerpt. The 3x headline number should be read as the wire is reporting it, accurate at the order-of-magnitude level, with the precise split between on-prem and hyperscale growth pending the full report.

The water conversation a year from now will also have to cover the older fleet, the in-house and enterprise datacenters that drove most of the 3x curve. Whether the utilities, water authorities, and corporate sustainability teams planning for the next decade treat that older fleet as a target, or keep tracking only the new hyperscale load, is the open question.