Data center cooling load and water use estimator
An IT load in, the heat it makes and the water it takes to move that heat out as a range for all four cooling designs at once, with the energy each one costs alongside it. The design sets the water number far more than the size of the site does.
The cooling load is the IT load
There is no conversion factor and people keep looking for one. Essentially every watt an IT load draws leaves the room as heat: a server does no mechanical work, stores no energy and emits a negligible amount of light and sound. So 100 kW of servers is 100 kW of heat, and a data center cooling load calculator that multiplies by something is either adding the electrical losses around the IT gear or is wrong.
The losses are worth adding, which is why they are shown as a second, slightly larger range. UPS conversion losses, PDU and transformer losses and lighting land inside the cooled envelope and have to be carried out with everything else: commonly another 5 to 10%. The rest of the facility's overhead, which is mostly the cooling plant itself, sits outside and does not add to what the plant has to remove.
Tons of refrigeration, and why the unit exists
One ton of refrigeration is 12,000 BTU per hour, which is the rate of heat removal that would melt one short ton of ice over 24 hours. It is a nineteenth-century unit from the ice trade and it survived because the people who size chillers have always used it.
That matters in a specific and avoidable way: the networking side of a build talks in kilowatts per rack, and the mechanical side quotes plant in tons. A room specified at 300 kW and a chiller quoted at 85 tons are the same number, and nobody in the meeting says so. To convert kW to tons of refrigeration, divide by 3.517. To go the other way, multiply.
A rough sense of scale: one ton is about 3.5 kW, so a single 7 kW rack is two tons, and a megawatt hall is roughly 284 tons before any redundancy.
Water usage effectiveness, the number that compares sites
Water usage effectiveness (WUE) is annual water consumed in litres divided by the IT equipment's energy use in kilowatt-hours. It is the water equivalent of PUE and it is the only figure that lets two facilities of different sizes be compared at all: a gallons-per-day total mostly tells you how big a site is.
| WUE (L/kWh) | What it implies |
|---|---|
| 1.8–1.9 | Around the reported industry average: evaporative cooling doing the work |
| 0.2–0.5 | Hybrid, or evaporative used only on hot days |
| under 0.05 | Best-in-class: closed-loop or air-cooled |
| 0 | Filled once and left alone |
The gap between the average and best-in-class is about thirtyfold. Nothing else in a data center's design moves a number that far, which is why this page makes you choose a cooling design before it will tell you anything, which is why it then shows you the other three anyway.
The trade the water number hides
Evaporative cooling is not a mistake somebody made. It is used because evaporation moves heat extremely efficiently, which means less mechanical cooling, which means a lower PUE. Closed-loop and air-cooled designs cut the water to nearly nothing and raise the electricity bill to do it.
So the two outputs move in opposite directions, and a facility can be criticised on either one depending on which number somebody chose to publish. That is why an indicative PUE sits next to every WUE in the comparison above. The PUE figures are ranges from published surveys, not a quote for a specific plant: treat them as the direction of the trade rather than as a design input. To take a PUE you already have and turn it into facility load, annual energy and a bill, use the data center power calculator.
There is a second-order version of the same trade that this page does not model: the electricity itself has a water footprint at the power station, and a thermoelectric grid consumes water per kWh generated. An air-cooled site can move water consumption off its own meter and onto somebody else's. Counting that properly needs a grid mix, and the boundary disagreement it causes is one reason published figures vary so widely.
Consumed, not withdrawn
Everything above is water consumed: evaporated or otherwise removed from the local supply, not coming back. Water withdrawn is what a site takes from a source, some of which is returned, often warmer. The two get quoted in the same sentence constantly, and the larger, more alarming figure is usually withdrawal.
A closed-loop facility can withdraw a great deal once, at fill, and consume almost nothing for years afterwards. An evaporative facility consumes most of what it withdraws, continuously. If you are handed a number for a real site and it does not say which one it is, it cannot be interpreted.
What this cannot know
These are estimates built from published survey and disclosure ranges, not measurements, and a real facility will sit somewhere in the band for reasons this page has no access to:
- Hours, not averages. The estimate assumes the IT load runs continuously at the figure entered. Real halls ramp, and an evaporative plant only consumes water on the hours it runs.
- Local weather, not a climate label. Wet-bulb temperature is what actually drives evaporative water use, and three buckets cannot stand in for a weather file.
- Water chemistry. Cycles of concentration, or how many times the water goes round before it is dumped as blowdown, can change consumption by a third on identical hardware.
- Redundancy and set points. N+1 plant, supply temperature and free-cooling hours all move the energy side of the trade.
Used as a data center water usage calculator for a sanity check, an order of magnitude, or a first pass at a capacity conversation, the ranges are the right kind of answer. Used as a design figure, they are not: that needs a weather file and a plant model.
Read the argument in full
The reasoning behind every range here, the sources, and why a gallons-per-day headline is nearly always uninterpretable are set out in how much water a data center uses. If you only take one thing from it: ask which cooling design before you react to the number.
Nothing is sent to us
The arithmetic runs in your browser. No load figure, no site, nothing typed here reaches a Kodura server, because there is no server involved.
Seeing it against real projects
The designs above stop being abstract when there is a site attached. Data centers in Minnesota runs this estimator's WUE bands across every facility in the state, and works Minnesota's 100-million-gallon water review backwards into the IT load that actually trips it, which turns out to depend almost entirely on the cooling design rather than the size.