When Air Is No Longer Enough — Where the Heat from Computing Goes

An Experiment Turning On an Entire Rack and Counting the Heat

There is an experiment where an entire rack was set up in a laboratory and powered on. A rack is a shelf-like frame in a data center where multiple servers, which are computers for computing, are plugged in one above the other. Four servers were plugged into this rack, and each server had eight GPU slots. A GPU is a graphics processing unit, a chip mainly responsible for artificial intelligence calculations.

A 1,000-watt dummy heater was attached to each GPU slot. A dummy heater is a test component made to be plugged into that spot instead of a real GPU to generate that exact amount of heat. The watt is a unit indicating how strongly electricity is being used right now ("Watts and Watt-Hours — Reading Electricity's Two Numbers Like a Speedometer and an Odometer").

Instead of a heat sink—a metal fin that transfers heat into the air—a metal plate through which liquid flows was attached on top of the chips. Water mixed with 25% propylene glycol, commonly used as antifreeze, entered at 32 degrees, and the surface temperature of the simulated GPUs hovered around an average of 64.5 degrees. When the authors matched the electricity that went into the rack and the heat that went out, about 97.1% of the heat emitted by the servers had passed into the facility's chilled water (DOI: 10.1115/ipack2024-140672).

This value is the capture rate, in other words, the share of the heat emitted by the servers that was caught in the cooling circuit without leaking into the room. The denominator is the heat emitted by the servers. In this rack, heat hardly scattered into the room, and 97.1% was carried by the water and passed into the facility's chilled water circuit.

This rack did not perform calculations. The simulated heating elements containing cartridge heaters simply generated heat according to a set wattage. Therefore, what this experiment shows is one thing: liquid actually sweeps away that heat.

Where the electricity used by computing equipment ultimately goes is supported by other evidence. Almost all the energy used by a data center is eventually discharged as heat (DOI: 10.2139/ssrn.7322018, preprint). This is the fact taken as a starting point by this preprint, which rewrites the cooling of artificial intelligence data centers as a matter of material selection. Computing equipment does not make electricity disappear; it changes its form. The electricity that enters becomes heat, and that heat must be discarded somewhere.

Connecting these two facts reveals the axis of this article. Almost all the electricity used by computation becomes heat, and that heat must be carried by something—air or water—and sent out of the building.

That is why a data center always needs two paths. The path that draws in electricity and the path that sends out heat. "Plugging It In Is Harder Than Building It — What an Interconnection Queue Actually Decides" followed why the former path gets blocked. This article follows the latter path.

What Changed Is Not the Total Amount, but the Density

A 2007 paper comparing data center cooling described that era like this. Even in high-performance data centers, the power used by a single rack stayed within 1 to 3 kilowatts. A kilowatt is a thousand watts. Divided by the floor area, the heat fell short of 100 watts per square foot (DOI: 10.1115/ipack2007-33176). The same authors predicted that a single rack would soon pump out 15 kilowatts, and within 2 to 4 years would exceed 30 kilowatts, and on a floor basis, 500 watts per square foot (DOI: 10.1115/ipack2007-33176).

Even those projections look small by today's standards. A testing paper published in 2024 deals with a cold plate, which is a metal plate attached onto a chip to flow liquid through. The site described in this paper is like this. A single accelerator—a chip made to speed up artificial intelligence calculations, like a GPU—is approaching 1 kilowatts, the power allowed for a single server has crossed 10 kilowatts, and the power of an artificial intelligence rack is nearing 100 kilowatts. The heat this paper faced on the test bench climbed up to 300 watts per square centimeter (DOI: 10.1115/ipack2024-141367).

Heat flux is the amount of heat passing through and escaping a slice of area every second. The "watts per square centimeter" from the previous paragraph is precisely this heat flux. Even the same 100 watts coming out of a palm-sized plate feels lukewarm, but coming out of a fingernail-sized chip, that spot burns away. Over the past 20 years, what has changed in data centers is not just the total amount of electricity. The spot where that electricity turns into heat and comes out has narrowed.

The baseline for a single chip has risen along with it. Cooling is designed using the amount of heat that the chip is expected to give off at its maximum as the baseline. This baseline has already surpassed 1 to 2 kilowatts for a single chip (DOI: 10.21203/rs.3.rs-9293095/v1, preprint). The same preprint built a test rack using a distribution unit rated at 160 kilowatts by boiling refrigerant right above the chips, and based on those results, concludes that this method can safely sweep away heat exceeding 160 kilowatts per rack. In this test as well, what generated the heat was a simulated heating element containing a 2.5-kilowatt cartridge heater instead of a real chip.

The reason for laying out these numbers is one. More than how well we cool, the fact that what cooling must face has become a completely different object over 20 years matters more.

Saying that density is the problem means this. Even if we cut the electricity used by the entire building in half, the fact that a single chip gives off 1 kilowatt in an area much smaller than a palm remains exactly the same. The total amount is a problem for power plants and transmission lines, and how to count that total amount is what "How Much Electricity Does AI Use? — Why Every Source Gives a Different Number" investigates. Density is a problem right above that chip, and this article starts there.

What Air Cannot Do

Air cooling is a cooling method that blows cool air into a server to carry away heat. When a fan pushes cool air into the front of a server, the air passes between heat sinks, absorbs heat, and exits the back warmed up. An air conditioner in the room cools that hot air down again. Here, what carries away the heat is only the portion where the air is warmed up.

Air can carry away only a very small amount of heat at a time. How much heat a substance carries away is determined by two things: how dense the substance is, and how much heat it takes to warm it by 1 degree. Air is small in both, so the heat that can be contained in the same volume is overwhelmingly smaller than that of water.

Therefore, there are only two ways to clear away more heat with air. One is to make the incoming air colder to widen the temperature difference, and the other is to push more air faster. Because each of these two paths hits a limit, a ceiling forms on the heat flux that air can handle.

High-performance servers cooled by air demand a low supply air temperature and a high airflow rate together, so when the heat that a single chip is designed to emit exceeds a certain line, air cooling becomes inefficient (DOI: 10.1115/ht2023-107598). This sentence is a premise written as the background of a study by a conference paper advocating for a method of immersing servers in liquid, and it is not the experimental result of that paper. The introductions of several papers in this field begin with the same premise (DOI: 10.54097/dhmrxm38, review; DOI: 10.1115/ipack2024-140672; DOI: 10.1115/ipack2022-96370).

Each of the two paths comes with a bill. To make the air colder, the refrigerator must be run harder. A refrigerator is a machine that makes water or air colder than the outside temperature by repeatedly compressing and releasing refrigerant. To push in more air, the fans must be spun harder. The electricity used by the fans increases steeply as the rotation gets faster, and that electricity ultimately turns into heat and remains in the same room. The noise grows larger as well.

If you handle air well, you can save quite a lot. In a laboratory case where the pathways through which cool air passes were blocked with walls and roofs so that it did not mix with the hot air that had passed through the servers, cooling energy decreased by 22% (DOI: 10.1115/ipack2013-73201). At the time, the estimation that the thermal management system consumes roughly 30% to 40% of the entire data center's energy was cited together, so the value of such improvements was large (DOI: 10.1115/ipack2013-73101).

However, blocking pathways and the heat on top of a chip are on different layers. Blocking pathways prevents the air already in the room from circulating in vain. When the heat flux rises to the point where a single chip emits 1 kilowatt in an area much smaller than the palm of a hand, whether that heat can be accepted by air is a matter of the properties of the substance called air, beyond the problem of operation. The portion that can be squeezed out through management and the portion that is opened only by changing the substance are in different bank accounts.

Carrying Heat Away in a Liquid

In the experimental rack earlier, the reason 32-degree water could hold the surface of the GPU near 64.5 degrees is that liquids differ from air in two ways. A liquid holds much more heat in the same volume, and instead flows much more sluggishly. Thanks to the first property, the amount of flow needed to carry away heat drops sharply, and because of the second property, a pump is needed to push the liquid along.

The methods using liquid split broadly into two. The first is direct-to-chip cooling. Direct-to-chip cooling is a method where a metal plate through which liquid flows is attached directly onto the chip, carrying heat straight into the liquid without passing it to the air. That metal plate is the cold plate we saw earlier. This is the difference between turning on a fan at a hot forehead and pressing a wet towel directly against it. The experimental rack that opened this article used precisely this method, which is why 97.1% of the heat could pass straight into the water (DOI: 10.1115/ipack2024-140672).

The water that went into that rack was 32 degrees. Not ice water, but water a little lower than human body temperature. The difference becomes clear when compared with air cooling, which demands a low supply temperature and a high volume of airflow together. Changing the medium that carries heat changes the very temperature difference needed to do the same job, and this difference leads directly to the question of whether or not to turn on a refrigerator later.

The second is immersion cooling. Immersion cooling is a method where the server is submerged entirely in a non-conductive liquid, allowing the liquid to directly take away the heat coming from all the parts. Instead of fanning each heated part, the entire part is put into a bathtub. Unlike direct-to-chip, which attaches plates only on the chips, in immersion the same liquid deals with everything submerged, whether it is the memory or the power supply.

The two methods differ in the range the liquid touches. Immersion does not require building complex liquid distribution piping and the liquid touches all parts of the server directly, as explained in a conference paper discussing immersion (DOI: 10.1115/ht2023-107598). Conversely, in direct-to-chip, the liquid deals only with the chips attached with plates, and the heat coming from the memory or power supply must still be carried away by air.

Each of the two methods splits again inside. Single-phase cooling is a method where the liquid carries heat while remaining a liquid all the way, and two-phase cooling is a method where the liquid takes away heat as it boils and turns into gas. Two-phase means it goes back and forth between two states: liquid and gas. This is the difference between water in a pot simply getting hot and water boiling while spewing steam. Because the heat it takes to turn a liquid into gas is much greater than the heat it takes to warm the liquid, the boiling side takes away far much more heat. In this article, "boiling immersion" refers to immersion using two-phase cooling.

This shift changes space along with temperature. When the average rack power density exceeds 7 kilowatts per rack, the area actually occupied by IT equipment drops to nearly half of the entire computer room area, with the rest taken up by cooling equipment (DOI: 10.1115/ipack2022-96370). IT equipment refers to equipment that performs calculations and communications, such as servers, storage, and network devices. This is a summary of existing research by a conference paper reviewing boiling immersion.

A metaphor from the same paper shows this space problem at a glance. At a common air-cooling floor density of 3 kilowatts per square meter, if you actually gather the electronic components of a data center that uses a multi-story building entirely, they would fill just the bottom 5 millimeters of the building. The authors estimate that boiling immersion can raise floor density by more than 6 times (DOI: 10.1115/ipack2022-96370).

Actual deployment has also begun. A report that Microsoft is applying boiling immersion to its cloud was presented at a 2022 conference (DOI: 10.1364/ofc.2022.tu2a.4, short presentation abstract).

What the Phrase "Immersion Is the Answer" Erases

Submerging an entire server does not end the problems. One paper directly comparing immersion and air cooling recorded the gains and burdens together within a single abstract.

The gains are substantial. The authors note that immersion reduces energy consumption by roughly 50%, cuts the occupied space by about two-thirds, and requires lower initial investment and operating costs. The latter part of the same abstract is the burden. Demanding maintenance procedures and an increased number of IT equipment failures are the main drawbacks, and cloud operators must make a decision to trade reductions in energy and cost for increases in maintenance and reliability burdens. The final sentence is clearer: retrofitting data centers already built for air cooling into immersion is costly and generally not recommended (arXiv:2205.04257, preprint).

Even after submerging them, the liquid does not work on its own. Among the coolant that enters inside the server, the proportion that actually passes between components is under 50% for both CPU and GPU servers, while the rest flows past the components. A CPU is a central processing unit that handles general calculations. This is the result of a preprint numerically simulating single-phase immersion in a GPU server, and the authors read this ratio as significant room for improvement through the design of the flow path, meaning the path the liquid flows. In the same study, the thermal resistance of the GPU heat sink was roughly 3 times that of the CPU heat sink (DOI: 10.2139/ssrn.7151773, preprint, numerical analysis). Thermal resistance is the degree to which heat is difficult to escape, and the larger it is, the hotter the chip gets for the same amount of heat.

What you submerge also heavily determines performance. The heat transfer coefficient is a value measuring the ability of a liquid to take away heat from a surface. In a study testing single-phase immersion with three liquids, the average heat transfer coefficient was 349.29 watts per square meter-kelvin for deionized water, 194.69 watts for propylene glycol, and 74.44 watts for white mineral oil (DOI: 10.37934/arfmts.111.2.141153). Even with the same device, what you fill it with changed the result by more than 4 times. However, this is a small-scale test run with heat generation between 300 and 500 watts.

On the boiling side, there is one more cliff. Critical heat flux is the limit of heat flux that a boiling surface can handle, and if this line is crossed, the surface is covered with a vapor film and the temperature catastrophically skyrockets (DOI: 10.2139/ssrn.7380728, preprint). This happens when liquid fails to timely refill the spots where bubbles have left the surface.

This limit is not a fixed constant. In a study experimenting with this limit targeting immersion to be used in power semiconductor modules, when pump circulation, coolant, and pressure were adjusted together, the limit of the test heating surface increased by more than double. The authors present these results as design guidelines for high-density immersion (DOI: 10.2139/ssrn.7380728, preprint). In the same experiment, the method of roughening the surface increased the limit for aluminum, but rather lowered it by 20.5% for silicon. The position of the cliff can be moved considerably by design, but above it there is an inherent ceiling determined by the surface and the coolant, and design determines how closely it approaches that ceiling, the authors write.

Still, on the test bench, this cliff is quite close. A chip cooling circuit that operates on its own through repeated boiling and cooling without a pump ran stably up to 1.2 kilowatts, but at 1.3 kilowatts the temperature skyrocketed rapidly and a local dry-out phenomenon appeared (DOI: 10.14293/p2199-8442.1.sop-.prlksr.v1, preprint). However, this is an initial measurement made with a baseline type lacking internal fins in the evaporator, the part where the liquid absorbs heat and boils, and the authors set improving that very fin geometry as their next task.

Direct liquid cooling, which boils refrigerant on top of the chip, is also evaluated as lacking sufficient foundation to widen its adoption. The cooling plate test paper that cited density figures earlier concludes that to widen the adoption of pumped two-phase direct liquid cooling, the ecosystem of various components making up the cooling circuit must develop further (DOI: 10.1115/ipack2024-141367).

There are also limits to packing things densely. In a boiling immersion tank, to keep the server surface below 85 degrees, submerged servers had to be spaced more than 15 millimeters apart at 1,000 watts, 20 millimeters at 1,500 watts, and 30 millimeters above 2,000 watts (DOI: 10.3390/en14051395). This is a result from the simulation side among a study that performed both experiments and simulations. The picture that filling a tank with servers seamlessly is always better stops here.

There is no standard yet either. Even the paper from that academic society advocating for immersion lists a long list of unresolved items. Whether to go with natural convection or forced circulation is the first. Natural convection is the flow where heated liquid rises and circulates on its own, and forced circulation is the flow pushed and circulated by a pump. How to redesign heat sinks designed for air cooling to suit a non-conductive liquid also remains, and studies dealing with the reliability of this cooling method are very sparse (DOI: 10.1115/ht2023-107598).

Specifications are also tailored to air cooling. Semiconductor thermal cycling testing is a test to see if components endure by repeatedly raising and lowering temperatures. The existing standards for this test were made on the premise of air cooling and could not be used as-is for immersion, and the authors revised and used material compatibility standards for insulating oil instead (DOI: 10.1115/ht2023-107598).

The situation is similar on the direct liquid cooling side. Studies on the reliability of cooling water used in direct liquid cooling are limited, and methods for testing reliability are not standardized either. This is the starting point written by a study that stepped up to fill this gap. The same study also takes as a starting point the fact that even with the same base liquid, corrosion inhibitors and additives put in by manufacturers can affect thermal performance. The authors attempt to borrow corrosion test standards for engine coolants to measure differences in reliability by manufacturer (DOI: 10.1115/ipack2023-110576).

Liquid cooling is the answer to thermal problems, but it is not a free answer. Heat is definitely removed. Instead, maintenance procedures, choice of liquid, material compatibility, the boundaries of boiling, and the item of an already built building are written together on the bill. Reliability points in conflicting directions. A comparative paper comparing immersion and air cooling notes that IT equipment failures increased (arXiv:2205.04257, preprint), while a society paper advocating for immersion notes that liquid prevents contamination and harsh environments to increase reliability, while adding that related studies are very sparse (DOI: 10.1115/ht2023-107598). Therefore, choosing immersion is less about picking a single correct answer and more about trading off the gains in energy and cost against the burdens of maintenance and reliability, as the preceding comparative paper notes.

What a Better Metric Misses

The number that appears most frequently in articles covering liquid cooling is a Power Usage Effectiveness sitting stubbornly close to 1. Power Usage Effectiveness is the total electricity drawn by the entire facility, divided by the electricity actually going into the computing equipment within it. It is called PUE for short. If this value is 1.5, it means the building drew 1.5 to put 1 into the server, and the remaining 0.5 was used mostly for cooling and moving it. The smaller the value, the less electricity leaks outside of computation.

Looking at the numbers alone, the transition seems already complete. Simulating a high-density data center with a maximum scale of 1,168 kilowatts, Power Usage Effectiveness drops from 2 with air cooling to 1.01 with boiling immersion (DOI: 10.1007/s10973-025-14006-0, CFD simulation). This value is the result of a computational fluid dynamics simulation calculating the flow of air and liquid by computer, not the measured values of a facility in operation. In a study actually measuring a boiling immersion tank, this value dropped from 1.053 to 1.037 while server power increased (DOI: 10.3390/en14051395). Even this is a partial metric, not the entire facility. Partial Power Usage Effectiveness is a value calculated the same way by isolating just a single cooling system instead of the whole facility.

Power Usage Effectiveness measures the electricity leaking outside the computing equipment. It is true that cooling well makes this value better. However, there are two things this value cannot see.

First, in places where cooling accounts for only a fraction of total electricity and the margin to reduce it by changing operations is also small, cutting cooling electricity significantly during peak hours barely moves this value. A preprint analyzing 8,736 meter records at 15-minute intervals over three months from a commercial data center in Dhaka, Bangladesh, tackles this head-on (DOI: 10.2139/ssrn.7169838, preprint). At this facility, chiller power was better explained by outside temperature than by the IT load—the electricity the computing equipment was using at that moment. Explanatory power is a value indicating how much a factor explains the rise and fall of chiller power, and it was 0.44 for outside temperature and 0.11 for the IT load. The optimal operating policy found by the authors cut cooling electricity by 23.3% during peak hours, but Power Usage Effectiveness improved by only 0.83%.

The authors do not view this as a failure of the model, but call it a structural ceiling. At this facility, the share taken up by cooling is 28.6%, and the maximum margin by which cooling can be reduced by changing operational methods is 5.8%. The upper limit of metric improvement is the product of these two, binding it to 1.65%. Like this facility, if cooling amounts to only a part of the facility's electricity and the cooling that can be reduced by changing operations is only a few percent overall, Power Usage Effectiveness barely moves even if you sharply cut cooling electricity during peak hours. Therefore, the authors consider this value an incomplete target where cooling is a minor load, and peak electricity priced by tariffs a more actionable target.

However, this result comes from a single hot and humid facility. The authors themselves note that hot, humid locations with a tight grid—meaning a strained power grid—are almost absent from existing literature. And they state that this result overturns the load-centric narrative coming from large facilities in temperate regions, which is the explanation that cooling electricity mainly moves following the IT load. The numbers from Dhaka cannot be transferred as-is to other places. What can be transferred is the method of multiplying cooling's share by the reducible margin to view the upper limit.

Second, this metric does not see what happens after the heat leaves the building. Because both numerator and denominator are electricity, how much water was used for cooling does not enter this value at all. A synthesis review in a policy journal reviewing whether Power Usage Effectiveness is still appropriate for artificial intelligence infrastructure summarizes complementary metrics that look at water use, carbon emissions, and equipment utilization together (DOI: 10.38126/jspg280101, review).

To this, this article adds two more. This is an observation of this article, not an argument from that review. At what temperature that heat left, and who receives and uses that heat, also do not enter this value. Why the former matters is seen in the next section.

There is one more thing. This value is a ratio. Even if the electricity used by computing equipment grows while the electricity outside it does not increase along with it by that much, the value improves even if cooling does not improve by a single bit. This is because what this value measures is not how much heat came out, but the proportion of electricity spent to clear away that heat.

So the sentence "We switched to liquid and it became 1.05" means the chillers and fans spun less, not that the heat problem is over. The heat generated by the computing equipment still leaves the building just as before.

Where Does the Waste Heat Go?

Right next to cities that need heating every winter, data centers are scattering heat into the sky with cooling towers. A cooling tower is an outdoor facility that releases heat gathered inside a building into the outside air. It makes you ask why that heat is simply thrown away. Half of the answer is "we can use it," and the other half is "it depends on what temperature it is."

The grade of heat refers to what temperature that heat is, in other words, how hot it is. Even for the same 1 kilowatt, 40-degree heat can only make lukewarm water, while 70-degree heat can be pushed straight into heating pipes.

The heat that data centers put out is mostly on the low-grade side. Data centers produce an abundance of 35 to 45 degree low-temperature waste heat, which is low-temperature heat that is thrown away without being used, but its uses are limited and its economic viability is poor, which blocks its recovery (DOI: 10.2139/ssrn.7396487, preprint). The answer this preprint puts forth is to use a heat pump to raise that heat to 100 to 200 degrees and store it. A heat pump is a machine that uses electricity to raise low-temperature heat to a high temperature. A configuration using carbon dioxide as the working fluid transferred 2.36 times as much heat for 1 unit of electricity in the model. This means that raising the grade takes electricity once again.

Therefore, the choice of cooling method determines the value of the waste heat in advance. District heating is a method where water heated in one place is sent through a pipe network to multiple buildings to provide heat. Immersion cooling liquids, which do not conduct electricity, can produce outlet temperatures of 65 to 75 degrees, and in that case they can be connected directly to the district heating network without a heat pump (DOI: 10.2139/ssrn.7322018, preprint). The main point of the materials-perspective preprint we saw earlier is that choosing a cooling fluid is not only a thermal management design decision, but also an energy system design decision.

The receiving side's situation is also a condition. District heating networks demand different water temperatures depending on the generation of buildings, and later generations run at lower temperatures. When dynamically simulating the case of putting the waste heat of a data center that uses a direct-water cooling circuit and a chilled-water rack cooler together into a district heating network, the answer splits by generation. Compared to the baseline configuration, when putting it into a 5th-generation network running at low temperatures, primary energy consumption (which is energy in its original form before being changed, like coal or gas) and carbon dioxide emissions drop by 88% and the payback period is under 1 year. In a 3rd-generation network that demands high temperatures, a high-temperature heat pump is needed, so the payback period stretches to 14 years (DOI: 10.3390/app16010323, dynamic simulation). Even for the same heat coming from the same data center, the answer splits depending on what temperature the other side of the pipe demands.

The demand side's situation also goes into the calculation. When simulating the waste heat recovery of a liquid-cooled data center by factoring in the real-time ups and downs of heat demand, cooling system electricity drops by 20 to 30% and partial power usage effectiveness drops down to a minimum of 1.13 compared to a liquid-cooling configuration without heat recovery. However, a caveat is attached to the share that can be reused. Considering the temperature constraints of the side receiving the heat, a maximum of 23% of the energy used by the data center is actually reused (DOI: 10.18686/cest630). The core of this sentence is "maximum" and "considering the temperature constraints." Even with those caveats included, the authors conclude that if the recovered heat is sold for hot water, the payback period is short, making a liquid-cooled data center an economically viable heat source under actual operating conditions.

Distance also gnaws away at the value. There is a preprint that calculates a plan on a corridor basis to take heat coming from an incinerator, hook it up to an air conditioner run by heat, and send it to an artificial intelligence data center (arXiv:2607.26324, preprint). A corridor is the pipe path that sends heat from the heat source to the data center. This calculation tallies the share that cools down while passing through the pipe, the share that turns heat into cooling, and the incidental electricity it takes to send and return it together, charting where the zone where profit remains ends as the corridor gets longer. The authors pin down that this framework is not for designing individual projects, but a tool for filtering out candidate sites.

The distance threshold itself is not carried over here. The heat source of that calculation is much hotter heat put out by an incinerator processing 1,500 tons a day, and the direction the heat flows is also the reverse of this article, going from the incinerator into the data center. That kilometer count cannot be carried over as-is to 35 to 45 degree data center waste heat. What remains is a single property. Heat loses value while following along a pipe, and because of that, how close the recipient is becomes a condition.

There is also a picture drawn on a country scale. According to a study simulating Latvia's data center waste heat potential, 51.37 gigawatt-hours of 65-degree heat came out in 2022, and it is projected to grow to 257 gigawatt-hours by 2050, with 201 gigawatt-hours of that expected to be used (DOI: 10.3390/en17020445). A gigawatt-hour is a large unit that measures how much energy something is, just like a watt-hour. The 65 degrees written here is higher than the 35 to 45 degrees we saw earlier. This means that recovery temperatures vary by facility.

A study that investigated the same topic earlier clearly writes down the stumbling block. It is that Latvia's data center distribution is uneven, which limits country-level waste heat utilization (DOI: 10.7250/conect.2023.007). This investigation takes as background the point that data center waste heat tied with heat pumps has been confirmed in various other regions as a cheaper heat source than fossil- or renewable-energy-based heating, and estimates Latvia's waste heat potential while also putting forth recommendations for adoption.

The sentence "You can just use waste heat for heating" is not wrong, but its conditional clause has been chopped off. It is true when you factor in whether the grade of heat matches the recipient's demands, whether you can pay the cost of raising it with a heat pump, whether the recipient is close by, and even whether that recipient's demand goes up and down from moment to moment.

The Same Design Works in One Place and Fails in Another

The exact same type of cooler that rejects heat using only air fails to hit its target temperature on the hottest day in Phoenix, yet forms the core of a design that eliminates chillers altogether in Pyhäjoki, Finland.

Dry cooling is a method that rejects heat to the outside using only air without consuming any water, while evaporative cooling is a method that rejects heat after spraying water into the incoming air to lower its temperature. When calculating both methods for a 10-megawatt facility placed across twelve data center hubs in the U.S., evaporative cooling reduced the electricity required to reject heat by a median of 30.5% and a maximum of 41% when operating with continuous spray (DOI: 10.2139/ssrn.7238869, preprint). This is the result of calculating 29 out of 36 combinations of regions and design conditions where both methods are feasible. The design condition is the outside air temperature set as a benchmark based on the rare, hot days that visit that region.

These twelve locations are a sample chosen in the first place for being favorable to evaporative cooling, and humid climates are omitted from this paper entirely. The authors write that humid climates are left to a forthcoming study currently in preparation. The savings range above comes from climates suited to evaporative cooling.

Up to this point, this looks like a battle of technologies, but the conclusion of the same study points to a map. Under the harshest design conditions in Phoenix, Las Vegas, and Tucson, dry cooling failed to reach the target cooling water outlet temperature of 107 degrees Fahrenheit (41.7 degrees Celsius) at all, making it unusable as the primary method. Which option is cheaper is determined more by the price ratio of electricity to water than by climate alone, and El Paso, Tucson, Albuquerque, and Reno mark the boundary where the advantages of evaporative cooling break down.

The water story gets only a single paragraph here. What the same study points out is the trade-off ratio where every kilowatt-hour of electricity saved requires 8 to 56 gallons (30 to 210 liters) of water on-site. A kilowatt-hour is a unit that measures how much electricity was used. This on-site water is roughly an order of magnitude greater than the water saved on the power generation side by using less of the same electricity. Therefore, evaporative cooling does not reduce operational water; it shifts the water burden from the power grid side to the site itself. How cooling and water are traded off is tracked by "AI Drinks Water — But Saving It Spends Electricity".

At the opposite extreme lies a different answer. Free cooling is an operation that rejects heat using only outside cold air or cold water without running chillers. It is close to turning off the air conditioner and opening a window.

Pyhäjoki, Finland is a place where 96% of the year is below 20 degrees. A data hall design targeting this location places dry coolers and an ice thermal storage tank on top of this climate (DOI: 10.36227/techrxiv.175355214.44070668/v1, preprint). A data hall is a room where server racks are gathered, and an ice thermal storage tank is a container that freezes ice and draws out coldness by melting it later. When it is cold, the dry coolers handle the entire load while simultaneously freezing ice, and during periods exceeding 20 degrees, that ice is melted to keep the server inlet temperature within specifications. The result is a configuration that does not run chillers year-round. This is a design proposed by the authors (the abstract does not present measured values), with free cooling handling most of the year and the ice frozen in advance by that free cooling filling in the remaining hot period.

The exact same dry cooling does not work under the hottest conditions in Arizona, yet in a design envisioned for Finland, it erases chillers entirely. Where and how to reject heat cannot be completely resolved inside the blueprint; it is bound roughly halfway to which sky the building stands under.

The Question That Finally Remains

The cooling system of a supercomputer named Frontier has 25 units that distribute cooling water. How many circuits to bundle these units into is a decision made once during construction and never changed, and there is a preprint that calculated all 611 possible ways of doing so (arXiv:2606.15408, preprint). What made the difference between the optimal configuration found by the calculation and the configuration actually built was the piping, rather than the energy used during operation. When even the backup systems actually in place for contingencies were priced in, the optimum shifted elsewhere. The authors' summary is that deciding how many circuits to divide into is a matter of backup policy, neither cost nor carbon. And because remodeling already-built facilities to redivide them is unrecoverable, they conclude that this framework is a tool for newly built sites.

Heat is currently the most belatedly discussed and longest-lasting item in data centers. Chips change every few years and software changes every week, but the piping, the cooling towers, and the latitude where the building stands remain the same. Whether a plot of land allows natural cooling or not is determined by its climate before any blueprint, and which method is cheaper is heavily weighed by the electricity-to-water tariff ratio.

And these decisions are mostly invisible from the outside. How many megawatts a data center uses appears in news articles, but whether that heat dissipates from cooling towers into the sky or enters a neighboring district's piping is rarely written down. The three conditions seen in the previous sections make that difference. They are what temperature the heat comes out at, how close the recipient is, and whether the chillers can be turned off on that land. All three are mostly decided before the building stands.

Only one thing is certain. Nearly all the electricity used by computing equipment becomes heat, and that heat must go somewhere. Everything else is a choice, and every choice comes with a price tag.

So when you hear stories about data centers, one more question has been added to ask. Along with how much electricity they use, it is where they are sending the heat that electricity turned into. Whether they are scattering it into the sky, feeding it into a neighboring district's heating pipes, or no one has decided yet is the answer to that question. These three look identical on the electricity meter, yet they are completely different choices.

Source Papers

Articles published on arXiv, SSRN, and Research Square are marked as "preprint," and whether they were subsequently published in an academic journal was not separately verified. Simulation and modeling results are noted in the main text whenever applicable.

• "Commissioning of a Liquid-Cooled High-Density Data Center Rack" (2024, conference/experiment) — DOI: 10.1115/ipack2024-140672 — In a 1,000-watt simulated heater rack, cooling water is 32 degrees and the surface is 64.5 degrees, with about 97.1% of server heat captured by facility chilled water.
• "Materials Design for Data Center Thermal Management" (2026, preprint) — DOI: 10.2139/ssrn.7322018 — Data center electricity is almost entirely discharged as heat. If the immersion fluid outlet is 65 to 75 degrees, it combines with district heating without a heat pump.
• "Analysis of Alternative Data Center Cooling Approaches" (2007, conference) — DOI: 10.1115/ipack2007-33176 — At the time, racks were 1 to 3 kilowatts, with projections to exceed 30 kilowatts and 500 watts per square foot within 2 to 4 years.
• "Common Cold Plate for Pumped Single- and Two-Phase Direct Liquid Cooling" (2024, conference/experiment) — DOI: 10.1115/ipack2024-141367 — Accelerators approach 1 kilowatt, AI racks approach 100 kilowatts, and test heat flux is 300 W/cm². A broader component ecosystem is needed for two-phase cooling diffusion.
• "Pumped Two-Phase R2L Direct-to-Chip Cooling in High-Density Racks" (2026, preprint/experiment) — DOI: 10.21203/rs.3.rs-9293095/v1 — Chip heat generation baseline exceeds 1 to 2 kilowatts, and results from a 160-kilowatt rated test rack (simulated heater) lead to the conclusion that processing over 160 kilowatts per rack is possible.
• "Single Phase Immersion Cooling for Hyper Scale Data Centers" (2023, conference) — DOI: 10.1115/ht2023-107598 — Premise that air cooling requires both a low supply temperature and a high air volume. Immersion touches all components without distribution piping, but due to a lack of standards, mineral oil specifications were modified and used.
• "Overview of Single-phase Immersion Liquid Cooling Technology" (2025, review) — DOI: 10.54097/dhmrxm38 — An overview of single-phase immersion cooling based on the premise that air cooling has hit its physical limits.
• "Impact of Cold Aisle Containment on Thermal Performance" (2013, conference/laboratory case study) — DOI: 10.1115/ipack2013-73201 — A 22% reduction in cooling energy in a case study blocking the cold aisle.
• "Hybrid Cooling Data Center With Rear Door Heat Exchangers" (2013, conference/numerical analysis) — DOI: 10.1115/ipack2013-73101 — Contemporary estimate that thermal management systems consume roughly 30 to 40% of total energy.
• "Enough Hot Air: The Role of Immersion Cooling" (2022, preprint) — arXiv:2205.04257 — Immersion cuts energy by about 50% and space by about two-thirds, but increases maintenance and IT failures, and retrofitting existing air-cooled facilities is not recommended.
• "Power Density in the Context of Two-Phase Immersion Cooling" (2022, conference/overview) — DOI: 10.1115/ipack2022-96370 — When exceeding 7 kilowatts per rack, IT area shrinks to half of the computer room. At 3 kilowatts per square meter, components are 5 millimeters below the building level, and boiling immersion exceeds floor density by 6 times.
• "Bringing 2-Phase Immersion Cooling to Hyperscale Cloud" (2022, presentation abstract) — DOI: 10.1364/ofc.2022.tu2a.4 — A report on applying boiling immersion to large-scale clouds.
• "Single-phase immersion cooling for GPU servers" (2026, preprint/numerical analysis) — DOI: 10.2139/ssrn.7151773 — Less than 50% of the cooling liquid passes through components, leaving ample room for flow path design, and GPU heatsink thermal resistance is about 3 times that of CPUs.
• "Investigation of Single-Phase Immersion Cooling for Modern Data Centers" (2024, experiment/300-500 watts) — DOI: 10.37934/arfmts.111.2.141153 — Deionized water 349.29 / propylene glycol 194.69 / white mineral oil 74.44 W/m²·K.
• "Multi-Strategy CHF Enhancement in Immersion Cooling" (2026, preprint/experiment) — DOI: 10.2139/ssrn.7380728 — Definition of the limit where nucleate boiling transitions to film boiling, and design guidelines combining pump circulation, cooling liquid, and pressure to raise test heating surface limits by 215% (surface roughness tested separately).
• "2PhaseEx – Two-Phase Thermosiphon Loop for Chip Cooling" (2026, preprint/initial measurements) — DOI: 10.14293/p2199-8442.1.sop-.prlksr.v1 — Stable up to 1.2 kilowatts, local dryout at 1.3 kilowatts. Based on a baseline finless evaporator.
• "Evaluation of a Two-Phase Liquid-Immersion Cooling System" (2021, experiment + simulation) — DOI: 10.3390/en14051395 — Measured partial metric 1.053→1.037, minimum server spacing obtained via simulation 15/20/30 millimeters.
• "Reliability Analysis of Liquid Coolants for Cold Plates" (2023, conference/experiment) — DOI: 10.1115/ipack2023-110576 — Measures differences among manufacturers using corrosion specifications to fill the gap where cooling water reliability testing is not standardized.
• "High-density data center via two-phase liquid immersion cooling" (2025, CFD simulation) — DOI: 10.1007/s10973-025-14006-0 — In a simulation of up to 1,168 kilowatts, power usage effectiveness drops from 2 for air cooling to 1.01 for immersion.
• "Data center cooling in a hot and humid climate: the limits of PUE" (2026, preprint/measured data) — DOI: 10.2139/ssrn.7169838 — 8,736 Dhaka instances. Despite a 23.3% reduction in peak cooling power, the metric improves by 0.83%, creating a 1.65% ceiling driven by a 28.6% cooling share.
• "Is Power Usage Effectiveness (PUE) Still Fit for Purpose?" (2026, review) — DOI: 10.38126/jspg280101 — Reviews the suitability of this metric in AI infrastructure and outlines supplementary metrics for water, carbon, and equipment utilization.
• "Heat pump energy storage for data center waste heat recovery" (2026, preprint/model) — DOI: 10.2139/ssrn.7396487 — 35 to 45 degree waste heat is blocked by applications and economics; proposes a configuration that heats and stores it at 100 to 200 degrees (model coefficient of performance 2.36).
• "Hybrid Data-Centre Cooling in District Heating Applications" (2025, dynamic simulation) — DOI: 10.3390/app16010323 — Compared to the baseline configuration, combining 5th-generation cuts primary energy and carbon dioxide by 88% with a payback under 1 year, while 3rd-generation takes 14 years.
• "Heat recovery and heating system for liquid-cooled data center" (2026, simulation) — DOI: 10.18686/cest630 — 20 to 30% reduction in cooling power, lowest partial metric 1.13. Factoring in demand temperature constraints and demand fluctuations allows up to 23% reuse.
• "Waste-to-Energy-Coupled AI Data Centers: Corridor Screening" (2026, preprint/screening) — arXiv:2607.26324 — Corridor calculation combining heat attenuation, absorption conversion, and auxiliary power. Authors emphasize it is for filtering candidate sites.
• "Data Center Waste Heat in District Heating in Latvia" (2024, dynamic simulation) — DOI: 10.3390/en17020445 — 51.37 gigawatt-hours of 65-degree waste heat in 2022, with projections to utilize 201 gigawatt-hours out of 257 gigawatt-hours in 2050.
• "Data Centre Waste Heat Potential in District Heating in Latvia" (2023, conference/survey) — DOI: 10.7250/conect.2023.007 — Results showing that uneven data center distribution in Latvia limits national-level utilization, alongside waste heat potential estimates and adoption recommendations based on system dynamics simulation (heat pump integration cheapness cited as background from other regional cases).
• "When Does Adiabatic Cooling Outperform Dry Cooling?" (2026, preprint) — DOI: 10.2139/ssrn.7238869 — In twelve locations favorable for spraying (excluding humid climates), median heat rejection electricity drops by 30.5%. Costs 8 to 56 gallons of on-site water per conserved kWh, shifting the burden from the grid to the site. Under design conditions of 0.4% in three locations including Phoenix, dry cooling cannot reach the target outlet temperature of 41.7 degrees, making it unusable as a primary method.
• "Nordics Prototype Data Hall Free Cooling Design" (2025, preprint/proposed design) — DOI: 10.36227/techrxiv.175355214.44070668/v1 — Proposes a chiller-free configuration using dry coolers and ice thermal storage in a location where 96% of the year is below 20 degrees.
• "Data Center Life Cycle Co-Design Optimization" (2026, preprint) — arXiv:2606.15408 — Evaluation of 611 configurations dividing 25 frontier cooling water distribution units. Deciding the number of circuits is a preliminary policy; once built, that facility cannot be repartitioned to recover resources.