Microsoft Just Redesigned the Data Center. The Thirstiest Thing Is the Break Room.

As GPUs run hotter, employee break room water usage is surpassing the water needed for AI compute. Is this the new data center flashpoint?

By: Jason Bak

Take a 100MW AI data center and put it in a Southwest desert. Cool it using Microsoft's new design standard, which is a sealed liquid loop with dry air cooling. Then add 70 people to run it, plus around 30 if power is being generated on site.

The AI cooling system will use about 8,500 gallons of water a year.

The 100 people will use about 900,000.

You read that right: The people out-consume the AI compute by more than 100 to 1. Two truckloads of water per year for the AI compute, plus three truckloads per week for the humans.

A 10x increase in capacity to 1GW would still leave the AI cooling consuming less water than three pecan trees.

Unlike pecan trees, though, the two truckloads per year for AI compute cooling don’t come from local rivers, aquifers or water utilities. In practice, it’s easier to have the water distilled and blended with food-grade anti-bacterial additives at specialty facilities. Most of these are in Texas and Louisiana.

The restroom water, however, can come from any of a variety of local sources. But for the sake of three truckloads per week, it doesn’t make sense to dig a well or build infrastructure to connect local utilities or rivers. Not to mention the permitting involved.

As more and more employment is created by the data centers, the break room is rapidly becoming the industry’s #1 use of water.

What Changed?

For decades, server rooms had to be fed air at about 75°F, which meant serious cooling. Especially in a 100°F desert. Today’s processors, however, can run at over 170°F internally, which is hotter than the hottest desert day ever recorded, and far warmer than the warmest desert night.

With today’s chips tolerating higher operating temperatures, the solution becomes intuitive: Blow hot air from the chips out of the building and then pump in fresh air at relatively cooler desert temperatures.

Not so easy: A broom-closet-sized processor rack pulls the same power as 75 top-of-the-line Dyson hairdryers, all blasting at full power with the closet door closed. Multiply by thousands of racks. The GPUs would melt if they didn’t have an auto-shut off function (just like hairdryers do when they’re used to dry the insides of shoes).

To get enough ventilation, the chips would literally have to run outdoors in a windstorm.

In comes cold plate cooling. It works by attaching metal plates to the hottest parts of each processor and then running water through fine channels inside each plate. Water holds 3,500x more heat than air by volume, so hot water can actually absorb more heat from the chip than cooler air. Today’s Nvidia processors increasingly ship with cold plates machined directly on them, and can be cooled with water as hot as 113°F.

The loop is sealed tight, so essentially none of this water is ever lost. Which raises the obvious question:

If the Loop Is Sealed, Why Do Data Centers Still Use So Much Water?

Two truckloads of water per year is about the same amount as a pecan orchard the size of two parking spots. So why do data centers have such a bad rap?

It all depends on how that water coolant is chilled:

The most common method is to spray additional water from outside the closed loop system over the hot sealed pipes and let it evaporate. Evaporating water pulls astonishing amounts of heat along with it.

Think of stepping out of a pool in the desert: Even though the air might be warmer than the water, you still feel cold. This is because water pulls heat out of surfaces, which makes your skin feel cold until you dry off.

You feel particularly cold in the desert because the dry air evaporates water off your skin faster, bringing the heat with it. Exit the same pool in a humid place like Miami, and you won’t feel as cold.

This heat transfer works just as well on pipes. It takes place in the familiar-looking towers that stand over data centers. They might look like smokestacks, but what’s rising is pure water vapor, not smoke.

Counterintuitively, the hot, dry air is what makes hot deserts a great place to keep processors cool. It’s why so many data centers have located in these arid climates. When water was less of a concern than power, that was a trade-off everyone accepted.

The desert-tax for conventional systems is that desert groundwater arrives so loaded with silica that it can only be reused about two-and-a-half times before being dumped. This amounts to an extra 40-50% of water, discarded as mineral-heavy waste. In a hard-water desert, a 100 MW system can therefore reach about 620,000,000 gallons of water a year, which is the equivalent of twelve tanker trucks arriving every hour, day and night, forever. Versus the 8,500 gallons used per year by Microsoft’s standard design.

Water-Free Cooling: Microsoft’s New Standard

Instead of evaporating massive amounts of water, the water coolant can be cooled water-free using a technology that’s been sitting in your driveway the whole time – your car’s radiator.

Car engines run even hotter than GPUs, and they can’t possibly carry around large water tanks to spray the engine whenever it’s running. So their radiators run water-based coolant through a pipe that touches the hottest parts of the engine (the “water jacket”), much like the cold plates that touch the GPUs. The radiator coolant then flows through thin metal coils that are positioned just inside the front grill where outside air flows in. Since the radiator coils are so thin, the water coolant inside gets maximum exposure to the outside air temperature. The faster the car moves, the more air flows through the grill, thereby exposing the tubes to more outside air. Even on a 110°F day, the air is far cooler than a 200°F engine.

Microsoft has committed to build all its GPU cooling systems to work the same way: The coolant water that runs next to the hottest parts of the chip is pumped outside into a heat-rejection field, which is a series of radiator-style coils that have similarly-thin profiles to maximize contact with the outside air. Instead of spraying water on these small coils, outside air simply flows through them.

This method doesn’t cool anywhere near as much - think of how much a fan cools you on a hot day versus a dip in the pool. No comparison, unless maybe it’s a really big fan.

However with today’s GPUs able to run at hotter temperatures, outside air in the low 100s is cool enough to keep them under their limits. The radiator doesn’t fight the 113°F supply temperature. It rejects heat from the return side of the loop, which comes off the chips at roughly 130°F. So with a typical 10–20°F approach, the fans keep up even when afternoon air pushes past 110°F. Our own sites’ design afternoons (the hottest hour engineers actually size for) run in the mid-90s, leaving a wide margin, and the rare 105°F records are exactly what the oversizing margin is for. And desert temperatures routinely fall into the 70s on warm summer nights, and below freezing during high desert winters. Not to say that splashing a bit of water on the cooler won’t help if there’s any around (more on this later), but it’s not necessary.

Best of all is that NVIDIA is now publishing reference designs that hold this 113°F tolerance while roughly doubling each rack’s power. They overtly specify that no chillers will be needed at all in most climates, the desert Southwest included.

The drawback? Power. Dry cooling takes on the order of 10–15% more electricity than evaporative systems at the height of summer (published comparisons run as high as 35%, but those assume chillers in the mix, which newer heat-tolerant, warm-water loops don’t require).

So a data center trades a slightly higher power bill for essentially zero water usage, except the water needed for employee restrooms (which is a wash - pun intended - between evaporative and waterless cooling systems).

Solar Supplies the Additional Power for Free, Without Additional Panels

Naturally, water coolant doesn’t get chilled as much by hot air as it does by cold air. As a result, on hotter days, fans need to push air through the outdoor coils faster and pumps need to circulate the coolant water faster. Bottom line – power consumption spikes on hot days.

But hot days are exactly the times when the sun is brightest. Solar panels therefore produce the most power when the cooling system needs it the most. On these bright days, it’s more power than our batteries can absorb.

So electricity that would otherwise go to waste is used to run the extra fan and pump speeds.

Adding to the efficiency is our east-west solar racking, which smooths out power yield throughout the day, thereby extending the hours where there’s excess power for the cooling systems (see my recent article on 5B and Jurchen Technology GmbH east-west racking: https://www.linkedin.com/pulse/meta-wants-beam-solar-from-space-real-race-unfolding-ofo2c).

However there’s a second water bill that almost nobody counts, and solar deletes that one too.

Solar Eliminates Even More Water

If a conventional evaporative cooling system is powered from a grid that generates via nuclear, gas or coal, add another 175 to 610 million gallons per year (gas at the low end, nuclear at the high end) on top of the 620 million the data center cooling itself evaporates. Tens of thousands of additional tanker trucks a year.

Here’s why: Coal, gas and nuclear plants don’t generate electricity directly from fuel. They boil water into steam, spin a turbine with it, and then have to condense that steam back into water to do it again. Condensing is done in cooling towers, and those towers evaporate water for exactly the same reason data centers do. A coal plant consumes roughly half a gallon of water for every kilowatt-hour it sends. Nuclear is closer to seven-tenths, and the most efficient gas plants are around two-tenths.

Solar panels boil nothing, spin nothing and condense nothing. There is no steam cycle, so there is no cooling tower and no evaporation. So a grid-connected data center that eliminates evaporation from its own roof is still, invisibly, evaporating water at a power plant a hundred miles away. Not to mention the carbon that’s produced by coal and gas.

It’s worth noting that fuel cells sit closer to solar. Bloom Energy-type units convert gas to electricity chemically - no boiler, no steam, no cooling tower. They also use essentially no water during operation beyond a one-time fill. The chemistry does make water vapor in the exhaust, but it’s as a product of the fuel cell process, not any water system intake. Smog-forming pollution falls sharply: Oracle puts the nitrogen-oxide reduction at about 92% against the turbine plan it replaces, and carbon dioxide about 20% lower. The fuel’s water footprint remains, however it’s shifted upstream and spent at the wellhead rather than at the data center.

Classic Radiator Coolant, but Food-Safe

Car radiator coolant typically consists of near-distilled water and ethylene glycol. This solution prevents microbial growth, and protects the coolant from freezing or boiling as the seasons change.

However, environmentalists and vehicle fleets in ecologically-sensitive places, such as national parks and marinas, often switch their coolant from ethylene glycol to propylene glycol. Propylene glycol is a biodegradable food preservative, found mostly in baked goods, including some products sold at health food stores like Whole Foods and Trader Joe’s. It avoids environmental contamination as well as those toxic stains people find in their driveways when radiators leak.

NVIDIA’s newest reference design specifies exactly this. Propylene glycol doesn’t protect from boiling or freezing quite as well as ethylene glycol, but it’s more than enough given the climates in most of the continental United States. This includes the freezing winter temperatures in the high deserts of Nevada, Utah, Arizona, New Mexico and Colorado.

Propylene glycol moves heat almost as well, and it’s a little thicker, which takes slightly more energy to pump.  In our designs, that extra pumping energy comes from surplus solar. For everyone else, the premium hardly matters anyway because the whole supply is only two truckloads per year. And going green is well worth it.

Two Summers

One desert quirk: After searing, bone-dry weather in May and June, the Southwest’s monsoons from mid-July into September bring darker, cloudier days, which are less friendly to solar. However, when those storm clouds roll in, temperatures can quickly drop 20-30 degrees, which cools the heat rejection apparatus right down without as much energy. Further, when the rain comes down, any water that touches this apparatus gives it an evaporative boost, chilling it even further. So the radiator’s worst days are the bone-dry scorchers in June, which thankfully take place during the longest solar days of the year. Enough additional sunlight hours to power the radiators overtime.

The monsoon moisture is a story of its own, which I’ll detail in my forthcoming piece.

Are Closed Loops Actually Closed?

The two truckloads of coolant water that the closed-loop system “consumes” each year is actually deliberate drainage for maintenance purposes. Sure there’s some natural leakage, but the vast majority is due to intentional coolant removal to swap pipes, plates or pumps. The additive solution also wears out and needs to get replaced every five to ten years. That’s what the ~8,500 gallons a year of top-up actually is - a maintenance number, not a consumption number.

Also, at the start, approximately 600,000 gallons are needed to fill the full facility loop. This is the equivalent of about 100 trucks, one time. It pales in comparison to the roughly three trucks every week for the employee bathrooms, and the twelve trucks every hour that an evaporative system would need (on top of the same three trucks per week), forever.

Why Not Just Take the Break Room Water From Local Rivers?

Legally, the ~100 trucks needed for the one-time fill and the 2 trucks needed for the annual maintenance refill could be sourced from an existing water rights holder (e.g. nearby sod farm or pecan orchard). They fall under the one-time exemption for construction and the threshold for annual operating exemptions. Still, it makes more sense to truck in pre-mixed coolant water than to attempt to distill and mix well water on site.

Water for the employee restrooms, on the other hand, can come from rivers and aquifers. The amounts needed, however, exceed the permitting thresholds.

AI developers therefore face a lengthy permitting and public consultation process for water that’s solely used to let employees use the restrooms.

In practice, instead of an extended wait for what’s essentially a restroom permit, it makes sense to truck in this water. As an added bonus, it saves the wear and tear on plumbing systems that salty, hard, mineral-loaded river water can cause.

Vintage, Not Virtue: Why Water-Heavy Evaporative Cooling is Still Popular

It isn’t that anyone’s foolish or has zero regard for the environment. It’s simply the vintage: A campus designed for 2016 chips made good decisions with good information; a campus drawn today has more heat-tolerant chips and different environmental dynamics. The advantage of building now isn’t intelligence, it’s timing. What matters to anyone underwriting this sector is that the asymmetry runs one way: Designing a sealed loop from the ground up is cheap. Retrofitting one into a building plumbed for a cooling tower is not. In other words, it’s better to build green from the ground up.

To be sure, water use at conventional evaporative data centers will fall as they replace legacy chips with modern, liquid-cooled hardware. But only if they also replace the evaporative equipment on the roof. Hotter-running chips permit dry cooling; they don’t deliver it by themselves.

Microsoft, the second-largest cloud provider on Earth, is building this zero-water design right now at its Phoenix-area campuses in Arizona and in Mt. Pleasant, Wisconsin. The first sites are expected to come online in late 2027. These aren’t pilot programs: Since August 2024, every new Microsoft data center has been designed this way. In their words, “without requiring a fresh water supply.” They’ve accepted the slightly higher power bill to get there, and without on-site solar, that bill lands in hard dollars, even though they’re powered by off-site solar.

While other hyperscalers are rapidly converging on this design, only Microsoft has publicly committed to zero-water evaporation as the default design for every new company-owned site, and has publicly committed to limiting WUE (water usage effectiveness) for its new sites to “near zero.”

When the number-two cloud on the planet makes waterless cooling its default and shows its math, it’s safe to assume that new data center builds will ease public concern about water usage as early as this time next year. And with GPUs having an obsolescence cycle of 3-5 years, legacy facilities will refresh onto heat-tolerant chips within a few years. These facility refresh cycles are the natural moment to swap the evaporative roof too, since, as noted above, the chips alone don’t save the water.

Google e has similarly been running liquid-cooled chips in production since 2018, when its third-generation AI processors outgrew what air cooling could handle. GPUs still couldn’t run as hot back then, so the cold plates continued to rely on evaporative cooling.

Google still deserves the pioneer’s credit for the cold plate side of the equation. When Sundar Pichai introduced those processors at I/O 2018, he noted that “for the first time, we’ve had to introduce liquid cooling in our data centers”. Google has since run liquid cooling at gigawatt scale across more than two thousand TPU pods at ~99.999% uptime since 2020, all while operating one of the most energy-efficient data center fleets on Earth.

Meta broke ground at Los Lunas in 2016, when none of today’s heat-tolerant chips existed. Air cooling was correct. Evaporation was correct - brilliant, even. They’ve since invested over $2.5 billion in New Mexico, and a pilot run across half the facility proved they could reduce indoor humidity from 20% to just 13%, which cut water use in the trial half by roughly 40% over nine months. Like many facilities of that vintage, Los Lunas chilled the interior by bringing in outside air, which needed to be humidified – another requirement that’s essentially eliminated by closed-loop liquid cooling, whose sealed halls simply recirculate the same air. This water reduction should further increase in their newer AI buildings, which are now all liquid-cooled and built with the new heat-tolerant GPUs. The company has committed to restoring more water than they consume by 2030.

Oracle’s Project Jupiter switched to sealed, closed-loop cooling design, together with a shift from gas turbines to as much as 2.45 gigawatts of Bloom fuel cells. Oracle says converting their natural gas to electricity chemically instead of by burning cuts smog-forming nitrogen oxides by about 92%. Like the closed-loop cooling system, the fuel cells need a one-time startup fill of water of their own. However the fuel cells are less picky, so Oracle has arranged for non-potable well water from an existing agricultural rights holder. Not a drop of the cooling or fuel cells touches the public system. The employees’ water, as at any workplace, comes from the regional utility.

If employee water use is now the biggest data center guzzler, how can it be reduced?

Even though employee restroom usage at data centers is no different than in any other industry, its water is still worth trying to conserve. We start where any office building would: Low-flow fixtures, sensible landscaping, no thirsty lawns. Then one extra step for the desert, where the cemented hardpan caliche makes ordinary septic leach fields impractical: A self-contained packaged treatment plant, the kind highway rest stops use, with its tanks shaded by — wait for it — solar panels.  The treated water then goes back to toilets, dust control or, if there’s any left over, a splash on the heat-rejection coils.

Why the Environment Finally Lines Up with Economics

I’ve been a renewables developer and company-builder my whole life. For most of it, I had to answer to investors who said the economics of going green couldn’t match a conventional, polluting facility. It’s a dream come true that, for data centers, green now quite simply delivers the best economics, so long as it’s designed that way from the start.

To recap, together with my last thought piece: Going green reduces time to power because the negligible amount of water used by closed-loop cooling avoids community harm along with the lengthy wait for water permits (and without the uncertainty of success). Similarly, on-site “behind the meter” power generation skips the wait for grid interconnections, which are now averaging 4-6 years nationally. On-site solar is the lowest-cost build per kW, and it’s completely pollution-free. It’s the absolute fastest to power because panels can be lit up in small modules as they’re built, without having to wait for the full buildout (as is the case with gas turbines). Ditto for the waterless, radiator-style coolers, as they can be built and commissioned in smaller increments than water-heavy cooling towers. Solar also brings the “free” additional energy needed by closed-loop cooling right when it’s needed (and some shade for the wastewater systems to boot). All told, this combination accelerates revenue by years, enormous in present-value terms. Communities, meanwhile, get silent, non-polluting, light-free solar farms, and data centers that don’t guzzle water beyond what their employees need.

Going green isn’t just about ideology anymore – it’s a competitive advantage. That’s why we’re leveraging our decades of green solution development and building green from the ground up.

Jason Bak develops solar-powered compute infrastructure in Texas and New Mexico.

#DataCenters #AIInfrastructure #LiquidCooling #NewMexico #DataCenterCooling