Specify the rejection stage, not the cooling technology
“Closed-loop liquid cooling” in a specification says nothing about water until the rejection stage is named.
INS-01Water and cooling · Analysis
Closed-loop cooling and the four things people mean by data center water
A closed coolant loop consumes nothing. Water is consumed at one stage only — heat rejection — and the metric the industry reports can only see part of it. Here is the mapping, and the worked calculation.
PUBLISHED LAST VERIFIED BY JOSEF ELIMELECHREVIEWED PODOS AI ENGINEERING
- 12×
- Source water hidden by the site boundary, 2023 US fleet
- 4.52 L/kWh
- Source water in US data-center electricity, 2023
- 2.3×
- How far WUE overstates the saving in the worked case
The short answer
A closed coolant loop consumes no water. Neither does liquid cooling as such. Water is consumed at one stage only — heat rejection — and the metric the industry reports, water usage effectiveness, measures only the water crossing the site boundary. In the 2023 US fleet that boundary hid roughly twelve times more water than it disclosed: 66 billion litres consumed on site against nearly 800 billion litres consumed at the power plants supplying it.[1] Any water decision made on site WUE alone is being made on about eight percent of the evidence.
First principles
Four different quantities, one word
Most arguments about data center water are people comparing four unlike quantities.
1. Facility water — withdrawal versus consumption
What the site meter records. Withdrawal is what is taken from the source; consumption is the part that evaporates and does not return to the basin. A once-through system withdraws enormous volumes and consumes little; a tower withdraws modestly and consumes nearly all of it. WUE is defined on consumption per unit of IT energy[2] — and was proposed in 2011 as a companion to PUE, never as a substitute for it.[3]
2. Source water — the litres inside the kilowatt-hour
Thermal plants evaporate water to condense steam. LBNL puts the 2023 average at 4.52 litres consumed per kilowatt-hour of electricity used by US data centers, against 4.35 L/kWh for US electricity overall.[1] A 2026 analysis of 472 US hyperscale facilities attributes about three-quarters of total operational water to generation rather than cooling.[8]
3. The closed coolant loop
A sealed circuit carrying a fixed charge of treated fluid between cold plates and a heat exchanger. Microsoft describes exactly this for its zero-water design: filled during construction, then circulated without a fresh water supply.[4] A closed loop has no consumption term. Make-up water is a maintenance event, not an operating input.
4. The heat-rejection method
The last stage, where heat leaves for the atmosphere. Towers buy low approach temperatures with water; dry coolers buy zero water with fan and compressor energy. This is the only stage that consumes anything, and it is chosen almost independently of everything upstream of it.
So “does liquid cooling use water?” is a malformed question. The real question is which rejection stage the loop terminates in — a siting decision, not a cooling-technology one. The same direct-to-chip architecture can be the highest-water or the zero-water option on the same drawing.
Table 1 · Original analysis
Mapping architecture to actual water consumption
The last column is the point: in every row, site WUE is either measuring the wrong stage, averaging away the hours that matter, or silent on the term that dominates.
| # | Architecture | Coolant loop | Heat rejection | On-site water | Where WUE misleads |
|---|---|---|---|---|---|
| W-01 | Air-cooled room, air-cooled chiller or DX | None — air is the only fluid touching IT | Air-cooled condenser | Effectively zero; humidification only | Looks best on paper. LBNL states the tradeoff plainly: air-cooled chillers use no water and use more energy, which moves the water to the power plant. |
| W-02 | Air-cooled room, water-cooled chiller, open tower | Closed chilled-water loop, open condenser loop | Evaporative cooling tower | High — evaporation plus blowdown | LBNL's worst case, and the one WUE captures well. The error is generalising from it to every water-cooled design. |
| W-03 | Airside economiser with adiabatic assist | None or partial | Outside air, evaporative pre-cool in wet mode | Low annual average, spiky in hot hours | An annual average hides that consumption concentrates in the heatwave hours when the basin is most stressed. |
| W-04 | Direct-to-chip liquid, facility loop to open tower | Closed technology loop, fixed coolant charge | Evaporative cooling tower | High — set entirely by the tower | Read as proof that liquid cooling consumes water. The sealed loop consumes none; the tower consumes all of it. |
| W-05 | Direct-to-chip liquid, facility loop to dry cooler | Closed technology loop, fixed coolant charge | Dry cooler or air-cooled chiller | Zero in operation after the initial fill | WUE falls to about zero — true and incomplete. The fan and compressor energy replacing evaporation raises PUE, and therefore source water. |
| W-06 | Direct-to-chip liquid, dry cooler with adiabatic assist | Closed technology loop, fixed coolant charge | Dry cooler, wet mode on hot days | Low annually, weather-driven, not zero | Called dry in procurement documents. LBNL ranks adiabatically assisted dry coolers highest among the otherwise low-water designs. |
| W-07 | Any of the above, with heat exported for reuse | Unchanged | Partly displaced by the heat off-taker | Falls in proportion to heat exported | WUE has no term for exported heat, so displacing rejection load earns no credit in the metric that frames the water-permit conversation. |
Original analysis
Worked calculation: what a zero-WUE site actually saves
The rankings in W-01, W-02 and W-06 come from LBNL's cooling-system modelling, which also warns directly that a low site WUE is not necessarily a good one.[1]
Microsoft's zero-water design is the clearest published version of this tradeoff: it avoids more than 125 million litres per year per datacenter, and the same disclosure notes a nominal increase in annual energy usage against its evaporative designs.[4] Google has put a number on the same direction, reporting that water-cooled data centers use about 10% less energy than many air-cooled ones.[6] That energy difference is the one worth pricing. Here it is, for one megawatt of IT load.
Stated assumptions
- 1 MW IT load, 8,760 h/yr — 8.76 GWh of IT energy.
- Case A, evaporative rejection: annualised PUE 1.20, site WUE 0.8 L/kWh of IT energy — the baseline of the 2026 hyperscale study, whose scenario range is 0.2–1.5 L/kWh.[8]
- Case B, closed loop to dry cooler: annualised PUE 1.30, site WUE 0.
- Source-water intensity 4.52 L/kWh, LBNL's 2023 US data-center-weighted average.[1]
- Same IT load, same year. Water embodied in construction, fluids and hardware is excluded.
| Annual result, 1 MW IT | A — evaporative | B — closed loop, dry cooler |
|---|---|---|
| Site water consumed | 7.0 million litres | 0 |
| Facility electricity | 10,512 MWh | 11,388 MWh |
| Source water in that electricity | 47.5 million litres | 51.5 million litres |
| Total water consumed | 54.5 million litres | 51.5 million litres |
| Reported site WUE | 0.8 L/kWh | 0 |
Reading the result
The saving was relocated, not eliminated
WUE says the dry-cooled site saved 7.0 million litres. It saved 3.0 million. The metric overstates the benefit by a factor of about 2.3, because most of the apparent saving was not eliminated — it was relocated to a cooling tower at a power station upstream.
Setting the site water saved equal to the source water added gives a break-even grid intensity that needs only two numbers an operator already has:
w* = WUE(evaporative) ÷ ΔPUE
Above a grid intensity of w*, going dry consumes more total water than staying wet. Holding WUE at 0.8 L/kWh and varying only the PUE penalty:
Figure 1 · Original analysis
Where the break-even crosses the
Each bar is the grid water intensity at which dry rejection stops saving water. The dashed line is LBNL's 2023 US data-center-weighted average. A bar below that line is a configuration that reports a WUE of zero while consuming more total water than the evaporative design it replaced.
16.0
8.0
5.3
4.0
Table 2 · The numbers behind Figure 1
Break-even grid intensity, row by row
At a 0.20 penalty — plausible for a hot site where compressors run most of the year — the dry-cooled configuration consumes 55.4 million litres against the evaporative site's 54.5, while reporting a WUE of zero. The ranking has inverted and the metric cannot see it. The limit case is the general rule: WUE is exactly right only when the grid's water intensity is zero, and flatters dry rejection in proportion to how thirsty the local generation mix is.
| ΔPUE of going dry | Break-even grid intensity w* | Reading |
|---|---|---|
| 0.05 | 16.0 L/kWh | Dry rejection wins on total water on any realistic grid. |
| 0.10 | 8.0 L/kWh | Dry rejection still wins on an average US grid, but by roughly a third of what WUE implies. |
| 0.15 | 5.3 L/kWh | Margin nearly gone; a hydro-heavy or nuclear-heavy balancing authority can flip it. |
| 0.20 | 4.0 L/kWh | Below the US average. The dry-cooled site consumes more total water while reporting a WUE of zero. |
A closed loop has no consumption term. Make-up water is a maintenance event, not an operating input.
8%
Share of the evidence a site-WUE decision rests on
Practice
What this means for operators
Carry PUE and WUE together or carry neither
Low-WUE designs are usually the higher-energy ones, so a scorecard with one and not the other will reliably pick the wrong plant.[1]Both hyperscalers that publish the pair report them together for exactly this reason — Microsoft's FY25 disclosure gives a global PUE of 1.17 alongside a WUE of 0.27 L/kWh.[5]
Get your balancing authority's water intensity before choosing
The break-even test needs one regional number, and LBNL's county map shows wide variation around the 4.52 L/kWh national average.[1]
Litres are not fungible
WUE counts reclaimed effluent identically to potable groundwater in a stressed basin, and the 2026 study finds direct and electricity-related burdens landing in entirely different basins and grid regions.[8] Google frames its own siting as balancing watershed condition against carbon rather than optimising a single number.[7]
Price the hot hours, not the annual average
Adiabatic designs consume water exactly when the basin is most stressed; an hourly wet-mode profile shows that, an annual WUE does not.
Expect to report it
Article 12 of the EU energy efficiency recast already requires data centre reporting from 500 kW of installed IT power.[9]
Honest limits
What this does not prove
This is a framework, not a verdict, and several inputs are softer than they look.
- It does not prove dry cooling is worse. In the base case it still wins — by 3.0 million litres rather than the 7.0 the metric claims. The inversion needs a large PUE penalty and a water-intensive grid together.
- The source-water intensity is a modelled average built from regional generation mixes, not a meter reading.[1]It does not track a site's power purchase agreements or behind-the-meter generation, either of which moves its real intensity. On dedicated wind the term approaches zero and the calculation collapses in favour of dry rejection.
- The PUE penalty is assumed. It depends on climate, facility water temperature, chiller selection and load profile; Microsoft characterises its own only as nominal.[4]
- The WUE input is a scenario value spanning 0.2–1.5 L/kWh,[8]while LBNL's 2023 US fleet average sits just over 0.36 L/kWh.[1] A lower true WUE lowers the threshold — the direction is stable, the magnitude is not.
- Consumption is not scarcity. Nothing here weights a litre by basin stress, recharge rate or potability, so identical totals can carry very different local consequences.
- Operational only. Water embodied in construction, in the semiconductors and in the cooling plant sits outside every figure here.
In the product
How this shapes a factory-built unit
On a real project these land in three places: the thermal enclosure fixes how much heat must leave and at what temperature, site power readiness establishes which grid you are actually drawing from, and heat recovery decides whether any of the load can be sold rather than rejected. Terms are defined in the AI infrastructure glossary; the upstream choice is covered in liquid cooling versus air cooling.
A modular unit forces the distinction to be explicit, because the coolant loop and the rejection stage cross the factory boundary at different points. Each PODOS Pod is designed as a standardized 1 MW building block and designed for 128 GPUs, with its closed technology loop specified and filled as part of the enclosure. What that loop terminates in stays a site decision — taken with the local grid's water intensity and the local basin in front of you, alongside the power architecture that ultimately sets the source-water term.
Key takeaways
- Closed loops consume no water. Heat rejection does. Everything else in the chain is neutral.
- Site WUE measures one stage of a two-stage problem — and in the 2023 US fleet, the smaller stage by roughly twelve to one.
- The break-even test is w* = WUE ÷ ΔPUE. Above that grid intensity, going dry raises total water consumption while reporting zero.
- WUE is exactly correct only on a zero-water grid; its error grows with the water intensity of the mix behind the meter.
- Report PUE and WUE as a pair, name the rejection stage in the specification, and get the regional water intensity before either.
Sources
- [1] 2024 United States Data Center Energy Usage Report (Shehabi et al.) — Lawrence Berkeley National Laboratory, December 2024
- [2] ISO/IEC 30134-9:2022 — Data centres key performance indicators, Part 9: Water usage effectiveness (WUE) — ISO/IEC, 2022
- [3] WP#35 — Water Usage Effectiveness (WUE): A Green Grid Data Center Sustainability Metric — The Green Grid, 1 March 2011
- [4] Sustainable by design: next-generation datacenters consume zero water for cooling — Microsoft, 9 December 2024
- [5] Measuring energy and water efficiency for Microsoft datacenters (FY24–FY25 PUE and WUE) — Microsoft, accessed 2026-08-31
- [6] Our commitment to climate-conscious data center cooling — Google, 21 November 2022
- [7] Advancing responsible water use at our data centers — Google, accessed 2026-08-31
- [8] The Hidden Water Geography of U.S. Hyperscale Data Centers in the AI Era (Guidi & Dominici), arXiv:2607.02531 — arXiv, June 2026
- [9] Directive (EU) 2023/1791 on energy efficiency (recast), Article 12 — data centre reporting — European Union (EUR-Lex), September 2023
Review schedule — 90 days. Next verification due 2026-11-29.
Bring your grid, your basin, and your rejection stage
Engineering will run the break-even against your balancing authority's water intensity before anything is specified.