How Much Water Does a Data Center Use?
Water has been this community’s most persistent question for over a year. It’s also the one with the widest range of correct answers. Depending on how a facility is designed, a 100 megawatt data center might consume almost no water at all, or well over a million gallons a day.
That’s not me dodging. It’s the actual state of the engineering, and understanding why the range is so wide is what lets you evaluate anything anyone tells you about this project.
What the project has said so far: it will not draw water from or discharge into the Huron River, and will not use local groundwater. Any municipal water required would be purchased through YCUA. The University is reviewing designs to consider closed-loop cooling and reclaimed wastewater. The facility will use direct-to-chip cooling.1 At the September 9 town hall, officials confirmed the cooling design is still an open decision, and no figure for water use was given.
Start with three distinctions
Most confused arguments about data center water come from collapsing these.
Withdrawal is not consumption
Withdrawal is water taken from a source. Consumptive use is water that does not come back — in a data center, almost entirely through evaporation. The rest returns as discharge, usually warmer and more concentrated in minerals.
For evaporative cooling, roughly three-quarters of what is withdrawn is consumed.2 So a facility can accurately say it returns water to the system while still permanently removing most of what it takes.
Direct water use is not the whole footprint
Every data center consumes water twice. Once on site, for cooling. And again at the power plants generating its electricity, where water is used for steam cycles and thermal cooling.
The indirect share is usually the larger one. A federal report estimated the indirect water footprint of U.S. data centers at roughly 211 billion gallons in 2023 — about 1.2 gallons per kilowatt-hour nationally.3
This has a consequence worth noting, because it cuts in the project’s favour on one point and against it on another: a facility that runs on wind and solar has a far smaller indirect water footprint, because those sources consume almost no water. So a genuine clean energy commitment is also a water commitment. But a commitment that covers only the academic portion of the load leaves the larger share unaddressed.
Liquid cooling is not the same as water consumption
This one matters here, and it’s where a well-meaning argument goes wrong — including one I’ve heard from people on my own side.
Direct-to-chip cooling circulates liquid through cold plates mounted on the processors. It is a heat capture method. It says nothing by itself about how much water the facility consumes, because the question is what happens at the other end of that loop — where the captured heat is rejected to the outside world.
- If heat is rejected through dry coolers — essentially large radiators — the loop is closed and water consumption is close to zero. The cost is electricity.
- If heat is rejected through cooling towers, water evaporates continuously and consumption is substantial. The benefit is lower electricity use.
So “direct-to-chip cooling” isn’t an answer to the water question. The heat rejection method is the answer, and that’s the specification nobody has published.
The actual cooling systems, compared
People ask about this more than anything else on the page, so here it is properly.
Think of a data center’s cooling as two jobs. Capture the heat off the equipment, then reject it to the outside world. Those are separate systems, often described with the same vocabulary, and that’s where most of the confusion starts.
Capturing the heat
Air cooling. The traditional approach. Fans push cold air through the racks. It works, it’s simple, and it runs out of headroom as chips get denser — which is why high-performance computing has largely moved on from it.
Direct-to-chip liquid cooling. Cold plates sit directly on the processors and liquid carries the heat away. Far more effective per watt, and it’s what this project has said it will use. The liquid in this loop is sealed and recirculates — it isn’t consumed.
Immersion cooling. Whole servers submerged in a non-conductive fluid. Rare, effective, and still has to reject its heat somewhere.
Notice that none of these three answers the water question. They’re all closed at this stage. The water story happens downstream.
Rejecting the heat — where the water goes
Once-through, or true open loop. Draw cold water from a river or lake, run it through heat exchangers, return it warmer. Consumption is low because nearly all of it goes back. Withdrawal is enormous, and the thermal discharge is its own environmental problem. Power plants still do this; new data centers essentially never do, and it would require discharge permitting nobody has applied for here.
Open-circuit cooling towers. The classic evaporative design, and the one that drives the alarming numbers. Warm water is sprayed through a tower, air moves across it, and some evaporates — carrying heat away as it goes. Extremely efficient thermally. It consumes water permanently, needs continuous make-up, produces mineral-concentrated blowdown that goes to the sanitary sewer, and requires chemical treatment and Legionella management.
Dry coolers, or closed-circuit. Essentially enormous radiators. The working fluid stays sealed in pipes and fans blow ambient air across finned coils. Zero water consumption. The cost is electricity — more fan power, and on hot days you may need mechanical refrigeration to make up the difference, because a dry cooler can only ever get you down toward the outside air temperature.
Adiabatic or hybrid. Runs dry most of the year and adds a fine water spray to pre-cool incoming air only on the hottest days. Water use is a fraction of a full evaporative system because it’s intermittent. This is increasingly the compromise large operators land on, and in a climate like Michigan’s it can run dry for the large majority of the year.
Chillers. Mechanical refrigeration, used alongside any of the above when ambient conditions aren’t sufficient. Air-cooled chillers use no water and a lot of power. Water-cooled chillers are more efficient but need a cooling tower — so they land back in the evaporative column.
| Heat rejection method | Water consumed | Electricity | Discharge to sewer |
|---|---|---|---|
| Once-through (open loop) | Low, but withdrawal is very large | Low | Thermal discharge to surface water |
| Open-circuit cooling tower | High — continuous evaporation | Lowest of the practical options | Yes — blowdown, treated |
| Adiabatic / hybrid | Low — hot days only | Moderate | Minimal |
| Dry cooler (closed-circuit) | None | Highest | None |
| Air-cooled chiller | None | Highest | None |
Why “closed loop” is a slippery phrase
This is the part I’d want people to take away.
“Closed loop” gets used for both systems. A direct-to-chip cooling loop is closed — the liquid touching the processors recirculates and isn’t consumed. That’s true, and it’s accurate to say so.
But a facility can have a perfectly closed chip loop and still reject all that captured heat through an evaporative cooling tower, consuming a million gallons a day. Both statements hold at once. The chip loop being closed tells you nothing about the facility’s water use.
So when you hear closed-loop cooling described, the question to ask is: closed at which stage? If the answer is the whole system through to heat rejection, that’s a real commitment to near-zero water. If it’s only the loop touching the chips, it isn’t an answer at all.
One thing Michigan has going for it
Evaporative cooling is most valuable where it’s hot and dry, because it can chase a temperature well below the ambient air — which is exactly why Arizona facilities lean on it so heavily, and exactly why their water numbers are frightening.
Michigan is the opposite case. For a large share of the year the outside air is cold enough that dry cooling works well on its own, and the penalty for going without evaporation is far smaller than it would be in Phoenix.
Which cuts against one argument before anyone makes it. If the case for evaporative cooling here is energy efficiency, that case is weaker in this climate than almost anywhere else they could have built. A dry or hybrid system in southeast Michigan gives up less than the national averages suggest.
The measurement everyone uses
Water Usage Effectiveness (WUE), defined in ISO/IEC 30134-9, is litres of site water consumed per kilowatt-hour of IT equipment energy. Lower is better, and zero is achievable.4
| Design | Typical WUE (L/kWh) |
|---|---|
| Air-cooled or fully closed-loop | ~0 |
| Best in class (NREL achieves 0.7 alongside a 1.06 PUE) | 0.3–0.7 |
| Large operators using mixed evaporative and non-evaporative | 1.1–1.6 |
| Industry average | 1.8–1.9 |
| Evaporative cooling | 1.5–3.0 |
| Hot-climate monthly peaks (Arizona) | above 9 |
Sources: industry benchmarks and reported operator figures.235
A caution about the headline average. The widely quoted 1.8 to 1.9 L/kWh figure traces to a 2016 Berkeley Lab study.6 Berkeley Lab’s 2024 report projects average WUE nearer 0.45 to 0.48 L/kWh going forward, as liquid cooling and non-evaporative designs spread.7 Those two differ by a factor of four, and anyone citing only one of them is telling you half the story. I’d rather show you both.
What that means at 100 megawatts
These are illustrations, not predictions. I don’t have this project’s cooling specification, because it hasn’t been published and by the University’s own account hasn’t been finalized. What follows shows the range that one decision spans.
At the older industry-average WUE, a 100 MW facility works out to roughly 1.1 million gallons per day.6
A 2026 academic analysis of data center impacts on public water systems runs the calculation more carefully. Using a conservative average WUE of 1.2 L/kWh, a peaking factor of 2.5 for hot days, and a consumptive ratio of 0.75, it estimates that evaporative cooling for a 100 MW IT load requires peak withdrawal capacity of roughly 2.5 million gallons per day — and notes that in hotter climates the requirement can exceed 5 million.2
The peaking factor matters more than the average, and it’s the number utilities actually care about. You size a system for the hottest week, not the annual mean. A facility with a modest average draw can still need a connection several times that size.
And at the other end: a fully closed-loop facility of the same size consumes essentially none.
That’s the whole point. The gap between those outcomes isn’t uncertainty about data centers. It’s a choice that hasn’t been made public.
The trade-off nobody gets to avoid
Water and electricity trade against each other. Evaporative cooling uses the heat of vaporisation to do work that a compressor would otherwise have to do, improving energy efficiency by something like 15 to 30% in suitable climates.5 Eliminate the water and you pay for it in power.
The panel said as much at the town hall: closed-loop cooling uses more electricity but no water.
So the community should be clear-eyed about what it is asking for. A closed-loop commitment protects the watershed and increases the load on the grid — which in turn increases the indirect water consumption at the power plants, unless the electricity is genuinely clean. These questions are connected, and a project that answered all three together would be making a real commitment rather than three separate reassurances.
What is specific to here
Michigan isn’t a water-scarce state, and I’ll say so plainly. Most of the national reporting on data center water comes out of Arizona, Texas and Georgia, where aquifers are stressed and the arithmetic really is alarming. Ypsilanti Township isn’t Phoenix. Anyone telling you this facility will drain the Great Lakes isn’t helping.
The local questions are narrower and more practical.
Municipal capacity. YCUA serves roughly 151,000 water customers and purchases treated water from the Great Lakes Water Authority. A large new industrial customer competes for that capacity and for the infrastructure that delivers it. In April 2026 YCUA voted a twelve-month moratorium on water and sewer service to data centers and AI computing facilities.
Who pays for capacity expansion. If serving this facility requires upsizing mains, pumping or treatment, that is a cost borne by a system whose other customers are households.
Discharge. Evaporative cooling produces blowdown — water concentrated in minerals and treatment chemicals — which goes to the wastewater system. Volume and composition both matter.
The 550,000 gallon threshold. Senate Bill 1046 would require any facility consuming 550,000 gallons per day or more to stop withdrawing from state waters, partner with a community supply, and obtain an EGLE permit — and to hold three properly noticed public hearings in the municipality before applying. Whether this project falls above or below that line is determined entirely by the cooling decision that has not been made.
The bill that would force the answer. House Bill 6251 would require any facility above 25 megawatts of IT load or 50 megawatts of total demand to file an evaluation with the Public Service Commission at least 120 days before construction — including projected water use, the cooling systems and heat rejection methods to be used, and whether closed-loop cooling and waste heat reuse were evaluated. Its companion House Bill 6252 requires a thermal impact assessment covering the same equipment.
Two features make them unusually useful here. Neither depends on a local permit, so neither is defeated by exemption from local zoning. And a developer may not designate the projected total electrical demand or a general description of the cooling systems as confidential — those would be public by statute. The heat rejection question at the top of this page is precisely what HB 6251 would answer.
The questions that would settle this
- What heat rejection method will be used — dry coolers, cooling towers, or a hybrid? This is the question. Everything else follows from it.
- What is the design WUE, and what is the projected consumptive use in gallons per day, at annual average and at peak?
- What peak withdrawal capacity is being requested from YCUA, and what infrastructure must be built to supply it?
- What discharge volume and composition will go to the wastewater system?
- Will actual water use be publicly reported, annually, measured rather than modelled?
- Will the University commit to a binding cap on consumptive use, enforceable by whom, with what consequence if exceeded?
The first question has an answer today. Heat rejection isn’t a detail you settle late — it drives the size of the mechanical yard, the electrical load, and the utility connections. Anybody who has specified direct-to-chip cooling and put a number on its efficiency gain has already thought about where the heat goes.
Until that answer is public, a promise not to draw from the Huron River is compatible with consuming a great deal of water, and a promise to consider closed-loop cooling is compatible with not adopting it.
Sources
- University of Michigan Office of the Vice President for Research, LANL FAQ. research.umich.edu
- “Small Bottle, Big Pipe: Quantifying and Addressing the Impact of Data Centers on Public Water Systems,” preprint, 2026. Peak withdrawal estimates for a 100 MW IT load, consumptive ratio, and peaking factors. arxiv.org
- Environmental and Energy Study Institute, “Data Centers and Water Consumption.” Average WUE and the national indirect water footprint. eesi.org
- ISO/IEC 30134-9 defines Water Usage Effectiveness as litres of site water use per kWh of IT electricity.
- WUE benchmarks and the evaporative cooling energy trade-off, industry technical guidance. Trade source. introl.com
- Data Center Dynamics, summarising Shehabi et al., Lawrence Berkeley National Laboratory 2016: average U.S. data center WUE of 1.8 L/kWh, equating to roughly 1.1 million gallons per day for a 100 MW facility. datacenterdynamics.com
- Berkeley Lab, 2024 United States Data Center Energy Usage Report, projected average WUE of 0.45 to 0.48 L/kWh beyond 2023. full report (PDF)
Last updated September 2026. The per-facility figures above are illustrative and rest on assumptions stated in the sources; they aren’t predictions for this site. If you work in this field and I’ve got something wrong, tell me — I’ll fix it and say that I did.