Every infrastructure argument in this forum eventually arrives at the same shape: someone quotes a headline number, someone else quotes a different headline number, and the disagreement turns out to be about which denominator each of them used. Power had it with capacity factors. HBM had it with wafers versus stacks. Water has it worse than either, because the two numbers that circulate differ by a factor of five and both get called "the industry average".
This piece does three things. It separates the two water numbers properly. It puts the missing 92% back on the table. And it works out what that does to the one decision operators actually make — evaporative or closed loop — in numbers rather than adjectives.
The metric, in one paragraph
Water Usage Effectiveness (WUE) is liters of water consumed per kilowatt-hour of IT equipment energy: WUE = annual site water use ÷ annual IT energy. Lower is better, and zero is achievable but not free — a facility that rejects heat to air consumes no water on site and burns more electricity doing it. Two accounting choices matter more than the number itself. First, consumption versus withdrawal: water that evaporates and never returns to the watershed is consumed; water that is drawn, used and discharged is withdrawn. WUE is usually reported on a consumption basis. Second, site versus source: the tower on your roof is site water; the water consumed at the power plant generating your kilowatt-hours is source or indirect water, and it is usually excluded from the headline WUE entirely. Keep those two axes in mind and the confusing numbers sort themselves out.
Two numbers that were never the same number
The figure that travels — in trade coverage, in vendor decks, in half the explainers — is about 1.8 L/kWh, described as the global or industry average. Facilities that rely on cooling towers with evaporative heat rejection typically land in a 1.6–2.0 L/kWh band, so 1.8 is a fair description of that design family. It is not a fleet average, and treating it as one overstates the sector by roughly 5× against the only large measured dataset available.
Lawrence Berkeley National Laboratory's 2024 United States Data Center Energy Usage Report, prepared for the Department of Energy, puts the US fleet's average site WUE at just over 0.36 L/kWh for 2023, and projects it rising to 0.45–0.48 as denser liquid-cooled AI racks come online. The gap is not a disagreement about physics; it is a difference in population. A fleet average includes every enterprise server room with a split system and every air-cooled colocation hall, and LBNL's own breakdown shows how much the mix has shifted: internal data centers were 64% of direct water consumption in 2014 and were down to 12% in 2023, with hyperscale and colocation now at 84%. The population changed, so the average moved.
Which number you should use depends on the question. Sizing a permit for one new evaporatively cooled campus? 1.6–2.0 is the right band, and you should design to your own cooling vendor's guarantee rather than to any published average. Comparing the sector to agriculture, or modelling a grid? 0.36 is the honest number, and even that is an average over a fleet whose two ends differ by more than an order of magnitude — closed-loop and immersion designs sit at 0–0.2 L/kWh, open towers at 2.0–4.0 in hot climates.
The 12× that almost never makes the slide
Now the part that changes conclusions. The same LBNL report estimates the water consumed generating the electricity US data centers used in 2023 at nearly 800 billion liters — against 66 billion liters consumed on site. Divide the two and the indirect footprint is about 12× the direct one; add them and 92% of a US data center's combined water footprint sits at the power plant, not on its own site.
Per unit of energy, that works out to 4.52 L/kWh of indirect water for the grid mix data centers actually sit on, slightly above the 4.35 L/kWh US grid average — meaning the industry's siting has drifted marginally toward more water-intensive generation, not less. For scale: US data centers drew 176 TWh in 2023, about 4.4% of national electricity, with projections of 325–580 TWh by 2028, or 6.7–12%. Apply the same 4.52 to the midpoint and the indirect footprint alone goes to somewhere north of one and a half trillion liters a year. That number moves with the generation mix, not with the cooling design, which is exactly why it gets left out.
This is also why per-facility "gallons per day per megawatt" figures in trade coverage should be treated with suspicion. They are usually conversions from a WUE figure by an intermediary, and the conversion hides the two choices that matter (consumption vs withdrawal, site vs source) plus the load factor. The arithmetic from WUE is short and auditable; the converted gallons figure is neither.
What the arithmetic does to the closed-loop argument
Here is where the 12× stops being trivia. The standard sustainability move is: replace evaporative cooling with a closed loop, a dry cooler, or direct-to-chip liquid cooling, and cut site water to near zero. That is real — closed loops report WUE around 0.19 L/kWh, and direct liquid cooling systems 0.0–0.1. It is also only half of a two-sided ledger, because dry heat rejection costs energy, and energy costs water at the plant.
Run it on one megawatt. A megawatt running flat out for a year is 8,760 MWh. At WUE 1.8 that is 15.8 million liters of site water; at 0.36 it is 3.15 million. Save the difference — going from an open tower to a closed loop at 0.19 — and you avoid about 1.6 L/kWh on site. But the PUE penalty for dry heat rejection is real and priced in the same units: operators describe accepting a 0.05–0.10 PUE increase, and in hot climates air-cooled plants can run 10–15% higher PUE than evaporative ones, which on a 1.15 baseline is closer to 0.12–0.17. Each 0.1 of extra PUE is 0.1 kWh more purchased per kWh delivered, and at 4.52 L/kWh of grid water that is about 0.45 L/kWh of new indirect consumption.
So the break-even is 1.6 ÷ 4.52 ≈ 0.36 of PUE. A closed loop that costs 0.05–0.17 of PUE is a clear net water win. A design that costs 0.4 — say, replacing an evaporative plant at 1.1 with an air-cooled plant at 1.5 — saves 1.8 L/kWh on site and adds 1.81 L/kWh at the plant. It is a wash, dressed as a virtue.
Three things follow. The grid decides. On a low-water grid the closed loop wins outright; on a coal-heavy one in a dry basin the same design can lose, because the cooling choice is being financed by someone else's watershed. The claim needs a grid attached. "Water-neutral" or "zero-water cooling" without a source-water figure is a site-only statement, and site is 8% of the footprint. The decision is locked in. Cooling and mechanical plant are a small slice of build cost but a 15–20 year commitment, so the PUE-versus-WUE trade is made once, at design, and then defended for two decades.
The regulators arrived before the arithmetic did
Policy has started to price the design choice directly, which is unusual — most efficiency metrics stay voluntary. Oklahoma's SB 259, the Groundwater Modernization Act, takes effect November 1, 2026 and does something no other state statute had done: it prohibits data centers from using groundwater in open-air evaporative cooling, or in any system that consumes groundwater through evaporation without recirculation. To obtain a groundwater permit, an operator must demonstrate low-consumptive cooling — closed loop, dielectric immersion, or equivalent. The mechanism matters more than the stringency: the cooling architecture became a permit condition, which turns a design preference into an engineering deliverable with a date on it.
It is not an isolated case. NCSL's review of 2025–26 legislation counts roughly 100 bills and resolutions across 39 states touching data-center water, most of them on reporting and disclosure rather than bans: California's AB 2469 requires a water supply assessment before approval and, from 2028, a drought contingency plan; AB 2619 folds water reporting into business licensing; Florida's SB 484 bars allocations that would harm local water resources and pushes reclaimed water where feasible. Reporting requirements look soft until you notice what they do to a claim: once a facility files annual numbers with a state, "we use almost no water" becomes a document rather than a slogan — and site-only numbers, filed alone, will invite exactly the question this article is about.
The infrastructure side is where the money is. A March 2026 study from UC Riverside and Caltech, "Small Bottle, Big Pipe", estimated that without new efficiencies, community water systems would need 697 million to 1.45 billion gallons per day of additional peak capacity within four years to serve data-center demand — a range the authors compare to the daily supply of a large city, with the observation that the constraint is not money but the absence of a national reservoir to draw from.
Why a compute buyer should care
Because water is a siting input that turns into a price input, and the transmission channel is unglamorous. A campus that cannot get a groundwater permit for an open tower either moves, or pays for a closed loop and the electricity it costs. Either way the capex and the PUE land in the same model as the GPU depreciation, and per-GPU-hour pricing inherits them. The facilities that clear permitting fastest in water-stressed regions are the ones whose cooling design already satisfies the strictest state rule — which is why direct-to-chip and immersion went from a density solution to a permitting solution.
Two practical checks when you read a provider's sustainability page. First, is the number site or source? A stated WUE with no source-water figure is describing 8% of the footprint. Second, is the PUE cost of that WUE disclosed? A near-zero WUE with an unstated PUE is an incomplete trade, not an achievement. And when the answer to both is "we buy renewable certificates", remember that certificates move the emissions attribute, not the water — the plant still evaporates.
There is a structural point underneath all of this that favours distributed capacity: water permits are granted per site and per watershed, and a 20 MW campus inside an existing industrial envelope inherits an existing water allocation instead of queueing for a new one. That is one of the quieter reasons smaller, already-powered sites can come online faster than greenfield hyperscale (omc.network).
The 2026 reality check
State the soft spots. The 66 billion liter and 21.2 billion liter figures, the ~800 billion liter indirect estimate, the 176 TWh and 325–580 TWh projections, the 0.36 and 0.45–0.48 WUE values, the 4.52 versus 4.35 L/kWh, and the 145–275 billion liter 2028 range all come from one place: LBNL's December 2024 report to DOE. That is a strong source — a national laboratory working from reported operator data and grid mixes — but it is a single modelling effort, and the indirect figure in particular depends on balancing-authority grid assumptions that individual power purchase agreements can invalidate. Reproduced across several outlets, it is still one estimate with one origin, and should be quoted as such. The 12× and 92% are my divisions of those two numbers, not figures from the report.
The 1.6–2.0 band for evaporative facilities and the 0.05–0.17 PUE penalty are ranges assembled from trade and vendor material, not measurements from a controlled study; the PUE cost of dry cooling varies enormously with climate, and a design in Singapore and a design in Ireland are not the same trade. The per-facility gallons-per-day figures circulating in coverage are conversions that even the outlets publishing them flag as unverifiable against a primary source — I have deliberately not used any. The UC Riverside/Caltech peak-capacity range is a research estimate, not an observation. The Oklahoma statute, its November 1 effective date and its permit condition are on the record and corroborated; the NCSL bill count is one organisation's legislative review and covers introduced bills, not enacted law. And the 2021 finding that roughly a fifth of US data centers drew direct water from moderately-to-highly stressed watersheds, with nearly half powered by plants in stressed regions, comes from one peer-reviewed study — cited here as one study, not as consensus.
What survives the caveating is the structure, and the structure is durable: site water is a small fraction of the footprint, the multiplier is set by the generation mix, the closed-loop trade has a calculable break-even, and regulators have started writing the cooling design into permits. Re-run the arithmetic against the current grid mix before you price anything. Do not quote an average without its population.
Comments