The 2% data center water figure comes from a report to Congress

On July 31, the Congressional Research Service published a 43-page report on data centers and water. On page 1, one sentence sets the scene: water consumed directly by data centers amounts to "a relatively small proportion (approximately 2%) of total U.S. water consumption". A small superscript 3 follows.

Footnote 3 points to a single document: Cooling the Cloud, published on October 28, 2025 by the American Water Works Association, the trade body for U.S. water utilities. So we opened it.

What's at the end of the footnote

The matching passage sits in section 3.1, and it runs to a single line: "Water consumption by data centers currently makes up a small proportion of total water consumption in the United States." A small proportion, and nothing more.

AWWA attaches no percentage to that statement. The document does carry one a few pages earlier, but it measures an entirely different quantity: the share of enterprise data centers in the sector's direct water use, projected to 2028. A share of data centers, not a share of America's water.

The CRS's 2% appears nowhere in the source it cites. It is born on page 1 of the report itself.

One detail makes it odder still. In the same paragraph, AWWA cites the 2024 Lawrence Berkeley National Laboratory report, the public reference on the subject: cooling water rising from 5.6 billion gallons in 2014 to 73 billion by 2028. It then adds that projections vary, with one study putting the 2018 total 80% lower. The numbers were on the table. AWWA drew no percentage from them.

A footnote that carries more weight than most

The Congressional Research Service is Congress's in-house research arm. It has no side to take: it writes for members of both parties, and its reports are the base layer staffers build on. This is documentation, not advocacy.

That's what makes the case interesting. A spokesperson shrinking a frame to defend a project, we've seen that in this column before, and the move is easy to spot. Here, nobody stood to gain anything. The number simply showed up.

No denominator produces it

The report supplies the numerator itself, on page 3: about 17 billion gallons for 2023, or 46.6 million a day. That left the denominator. So we went to the USGS series, the national inventory of American water.

Against total 2015 withdrawals, 322 billion gallons a day: 0.014%. Against water delivered by public supply systems, 39 billion a day: 0.12%. Against the agency's most recent net consumption estimate, 100 billion a day: 0.047%.

For "approximately 2%" to hold, the United States would have to consume 2.33 billion gallons a day, all uses combined. In the USGS tables, the closest quantity is withdrawals for livestock: 2.00 billion a day. A country does not consume what its cattle drink.

One caveat on our own math, and it's a serious one: the last national consumption estimate the USGS published covers 1995, after which the agency stopped producing the number. Our 0.047% therefore rests on a thirty-year-old denominator. The withdrawal figures are current.

Someone had already run the numbers

In its March-April issue, months before the CRS report, the Environmental Law Reporter published a study on the resources data centers consume. It takes its denominator from the same USGS series, and its numerator, 17.4 billion gallons for 2023, from a scientific journal rather than from Berkeley.

Published result: "Data centers, therefore, account for roughly 0.015% of total U.S. water usage". Fifteen thousandths of a percent.

The piece is signed by a Yale student, in a journal written for environmental law practitioners, and its denominator mixes withdrawals and consumption in the same way. So it doesn't outrank a report to Congress. What it establishes is more modest, and more awkward: the calculation was published and checkable before the 2% was ever written.

And another outfit chose not to put a number on it

On April 8, MOST Policy Initiative, a nonprofit that briefs Missouri lawmakers, published its own note on data center water. Two pages written for legislators: exactly what the CRS report is for.

It leans on the same Berkeley report, also gives 17.4 billion gallons for 2023, and goes further: it publishes the full 2028 projection range, 38 to 73 billion, where AWWA kept only the upper bound.

Across those two pages, no share of American water at all. The only percentages cover what the note can actually measure: up to 85% of withdrawn water evaporating instead of returning to the system, 78% at Google. Leaving out a national percentage wasn't an oversight. It was the finding.

What the report gets right

This report doesn't deserve a black mark. Across the forty-two pages that follow, the work is clean.

The CRS never claims to be measuring: it writes "One study estimates", points to the LBNL and gives the page numbers. It separates direct water from indirect water, the kind that leaves as steam at the power plant, and states on page one that the latter is out of scope. It repeats that this consumption is hard to quantify.

The gap comes down to one introductory sentence, and its position is what makes it dangerous: it reads fast, it quotes cleanly, and it lands before the caveats.

So where does the 2% come from? All anyone can offer is a hypothesis. You'd have to stack two shifts: fold in the indirect water the sentence excludes in so many words, then swap the country's total consumption for water from public supply systems.

That is neither an excuse nor an established explanation. The question stays open, and we did not put it to the CRS.

The verb changes with every pickup

What happened next is more instructive than the gap itself.

On August 3, Legis1, a legislative intelligence service sold to firms working with Congress, wrote: "The report finds that data centers now account for approximately 2 percent of total U.S. water consumption". The word "direct" is gone, and "estimated" has turned into "finds". The report wasn't estimating anymore, it had found.

On August 24, The Register dissolves the parenthesis: the figure is pushed to the end of the sentence, glued to the denominator. Along the way, the piece credits the report with a "more than 80 percent" indirect water share that the CRS writes nowhere, its own footnote 4 citing the International Energy Agency at 60%.

On a statistics aggregator, one of those pages you open for a ready-made number, the wording today is: "the CRS confirms that data centers now account for roughly 2% of total U.S. water consumption nationally". Confirms.

In four weeks, a sentence with no number became an estimated number, then a found one, then a confirmed one. The number itself never moved. Only the authority grew, and that is the one thing somebody added along the chain.