The efficiency number improved. The water number in the same report did not.
Berkeley Lab projects average US data center PUE falling to 1.15-1.35 by 2028 while average site WUE rises to 0.45-0.48 L/kWh, and names the shift to hyperscale and colocation as a cause of both. Most of the water is not on the site meter at all.
Data centers are getting more power-efficient, and the number that says so is real. The Berkeley Lab report projecting it also projects a water number moving the other way, and names the same causes for both. That is not a contradiction but a trade, made in the cooling plant. The efficiency figure in a siting announcement is one side of a ledger; the other side is rarely on the page.
Two numbers, one report
Both metrics hang off the same denominator. “Power Usage Effectiveness (PUE) is defined as the total electricity demand of the data center divided by the electricity demand of the IT equipment.” A PUE of 1.4 means 40% more electricity enters the building than reaches the computers. “Water Usage Effectiveness is similarly defined as the total water consumption of the data center divided by the electricity demand of the IT equipment.” It comes out in liters per kilowatt-hour.
The report is careful about which water. It calls the on-site figure WUE (site), “primarily associated with cooling infrastructure”, and separates it from “the water consumption from the electricity generation”, which it calls WUE (source).
Here is PUE. “The resulting annual average PUE falls from 1.6 in 2014 to 1.4 in 2023,” and “This decline is primarily due to the shift towards larger data centers (hyperscale, colocation) that have a lower PUE.” It keeps falling: “by 2028, the average PUE falls to between 1.15 and 1.35,” “driven again by the shift into more energy efficient hyperscale and colocation facilities, combined with the increase in liquid-cooled AI servers.”
Here is WUE, from the next page. “The shift toward hyperscale and colocation data centers results in an increase in the overall average WUE”, which “stays just over 0.36 L/kWh through 2023”. After that “the average WUE rises slightly, reaching between 0.45 and 0.48 L/kWh,” which the report attributes to “increasing WUEs in the hyperscale and colocation data centers, along with the increased water consumption of liquid-cooled systems.”
Put those side by side. The same shift into hyperscale and colocation is named as the reason the power number falls and as the reason the water number rises. Liquid cooling sits on both sides too.
The trade is made in the cooling plant
Evaporating water carries heat away without a compressor doing the work, so the more of the job you hand to evaporation, the less electricity you spend and the more water you consume. That is how one report moves both numbers at once.
It gets concrete about the case that matters. “Liquid IT cooling is an emerging technology for cooling dense IT equipment (e.g., AI).” For that equipment “the base case is a waterside economizer”, and of that system the report says: “While highly energy-efficient, this system consumes substantial amounts of water, like all evaporative cooling systems.”
Efficient and thirsty, in one sentence.
The bigger number is not on the site’s meter
WUE (site) counts water the facility itself consumes. Direct consumption by US data centers was “21.2 billion liters” in 2014, and by 2023 hyperscale and colocation accounted for 84% of “the 66-billion-liter total”.
The water consumed generating the electricity those buildings draw is a separate line. “The total indirect water footprint of U.S. data centers is nearly 800 billion liters,” about twelve times the on-site figure. Per unit of electricity that is “4.52 L/kWh of indirect water consumption”, against a national average across all uses of “4.35 L/kWh”.
So data centers are not unusually thirsty per kilowatt-hour. They are unusual in how many kilowatt-hours they buy.
That ratio sets a break-even, and this arithmetic is mine. At 4.52 liters per kilowatt-hour, a cooling change that spends more than about 0.22 extra kilowatt-hours to save one liter on site raises total water use while lowering the reported site WUE. I have not found published figures for where real systems fall.
Volume and scarcity point different ways
Volume is not risk. A 2021 paper in Environmental Research Letters, co-authored by the Berkeley Lab report’s lead author, resolved US data center water use down to individual subbasins using 2018 data. On volume it agrees with the direction but not the size: “Only one-fourth of the volumetric water footprint of data centers resulted from onsite water use.” Three to one, where the 2023 figures give twelve to one. Take the direction, not the ratio.
Then it weights each liter by how scarce water is where it was taken. A water scarcity footprint (WSF) “indicates the pressure exerted by consumptive water use on available freshwater within a river basin”, and on that measure “more than 40% of the WSF is attributed to direct water consumption.” Direct use, the paper concludes, “is skewed toward water stressed subbasins compared to its indirect water consumption, which is distributed more broadly geographically.”
A quarter of the volume, more than 40% of the scarcity. The on-site number is the small one and the sharp one.
What this changes in a model
The two numbers do not share an address. WUE (site) belongs to one building on one parcel. WUE (source) belongs to generators across whatever balancing authority serves it; the report notes that “Each county is assigned a balancing authority based on its geographic location”. The grid connection I wrote about earlier turns out to carry a water liability alongside the power one.
So treat a published PUE as half a disclosure and ask for two more things: the site WUE beside it, and the balancing authority. The site number is the smaller share of volume but carries a disproportionate share of the scarcity, so it is the one exposed when a basin runs short. The grid number is most of the volume and moves with the generation mix wherever the site sits. A model carrying only one of them is understating something.
Where I could be wrong
The indirect figure excludes exactly the thing large buyers do. The report says its method “does not incorporate any power purchase agreements between individual data center facilities and their electricity providers”, nor on-site generation behind the meter, and that the omission “could significantly affect water consumption and emissions estimates, depending on the electricity source.” It is the strongest argument against the indirect numbers, and it comes from the source. It has a limit: the 2021 paper finds that “Purchasing renewable energy certificates from electricity providers does not necessarily reduce the water or carbon footprints of a data center.” Its remedy is direct connection, not a certificate.
These numbers are simulated, not metered. “For this report, these metrics are simulated using established thermodynamics-based models”, then paired with assumptions about where data centers sit. The report’s own sensitivity check moves the hyperscale median WUE “from 0.32 to 0.40, which is still within the range of uncertainty shown in Figure 4.5.” Inside the band, then, but on the wetter side.
The scarcity split is built on 2018 data. Siting has moved since, and I do not know which way.
These are national averages. A single facility can sit anywhere in the distribution, and the 2028 figures are scenario ranges. The report varies “equipment shipments and operational practices, as well as variations in cooling energy use” together, so no single driver owns a range.
Rising WUE may be composition rather than choice. More of the fleet is hyperscale and colocation, and the report says that shift raises average WUE, so the trade can be real industry-wide without describing any operator’s decision.
Volume is not cost. I have no water price series; the argument rests on siting and scarcity, not price.
The efficiency number is improving and the water number is rising, and both are true because they are the same decision seen from two sides. A disclosure that reports one of them is not wrong. It is half-priced.