There is a thermodynamic identity that makes data center design simpler than most people assume. Every watt of electricity that enters a data center exits as heat. Not approximately. Exactly. The chips don't store the energy, the servers don't convert it into anything useful in the physical sense, and the cooling system's only job is to move that heat from the inside of the building to the outside.

A 100-megawatt data center is, from a physics standpoint, a 100-megawatt heater. The building's electricity bill and its heat output are the same number, just expressed in different units. The conversation about data center efficiency is, at root, a conversation about how cleverly the building moves that heat to somewhere it isn't a problem. As the chips inside the building have gotten more power-dense, the cleverness required has gone up by orders of magnitude, and the entire industry is in the middle of a cooling transition that is reshaping how data centers get built.

The Air Cooling Era

For the first 30 years of commercial data center operation, cooling worked the same way. Cold air came up through a raised floor. Hot air came out the back of the servers. The hot air got pulled across cooling coils that ran chilled water through them. The chilled water carried the heat outside to a cooling tower, where evaporation rejected it to the atmosphere. The whole loop was air on the inside, water on the outside, and the engineering was well-understood.

This worked because rack densities were low. A standard rack drew 5 to 10 kilowatts. Air at moderate flow rates could carry that much heat away without exotic equipment. The downside was that air is a terrible heat transfer medium. It has low heat capacity, low density, and requires a lot of fan power to move. Older data centers spent 30 to 50 percent of their total electricity just on the cooling system, which is the inverse of what you want when the goal is to spend electricity on computing.

The Power Usage Effectiveness metric (PUE), which measures total facility power divided by IT equipment power, captured the inefficiency directly. A PUE of 2.0 meant the cooling system used as much energy as the servers. A PUE of 1.5 was considered good. The hyperscalers chased lower numbers through the 2010s, with Google's fleet averaging about 1.09 against an industry average still near 1.56, and the rest of the industry followed slowly behind.

That whole architecture is now obsolete for AI workloads. Air cooling cannot remove heat fast enough from a 100-kilowatt rack. Not at any fan speed. Not with any aisle containment strategy. The thermal density is just too high, and the air molecules can't carry it. The industry has known this since 2018 and has been preparing for the transition since then.

What Liquid Cooling Actually Does

Liquid cooling, in its mature form, brings water or a coolant directly to the chip. There are two architectures.

Direct-to-chip liquid cooling routes a cold plate, a thin metal manifold with channels for fluid, into thermal contact with the CPU and GPU dies. Coolant flows through the cold plate, picks up heat, and returns through pipes to a heat exchanger. The fluid never contacts the electronics directly. The plumbing runs to every server in every rack. The work of moving heat is done by the fluid, not the air. This is the architecture Nvidia ships with the GB200 and B200, and it is becoming standard across hyperscale builds.

Immersion cooling submerges the entire server in a dielectric fluid (a liquid that doesn't conduct electricity). The fluid carries heat away by direct contact with every component. There are two kinds: single-phase, where the fluid stays liquid the whole time and circulates through a heat exchanger; and two-phase, where the fluid boils on the hot components and condenses on cooler surfaces, using the latent heat of vaporization to move energy at extremely high rates. Immersion is more efficient than direct-to-chip but harder to operate, harder to service, and harder to retrofit into existing buildings. Adoption has been slower as a result, but the technology is mature and several specialty applications are scaling it.

Both architectures change the fundamental thermodynamics. Water carries roughly four times the heat of air per unit mass, and because it is far denser, on the order of 3,000 to 4,000 times more per unit volume. The flow rates required to remove a fixed amount of heat collapse by orders of magnitude. The fan power that air cooling required largely goes away. The PUE numbers that air-cooled facilities optimized for stop being the right metric. New liquid-cooled builds are landing in the 1.05 to 1.15 PUE range, with most of the cooling overhead being pump work rather than fan work.

The heat itself, however, has not gone anywhere. It still has to come out of the building.

The Outdoor Side

Whatever the indoor cooling architecture, the outdoor side is where the heat actually leaves. The conventional choices are evaporative cooling towers (which use water and reject heat to the atmosphere through evaporation), dry coolers (which use only air and reject heat through forced convection), and adiabatic hybrids (which use a small amount of water to pre-cool the air before a dry stage, getting most of the efficiency of full evaporation with most of the water savings of dry-only).

Phoenix is in an interesting position on this. The dry climate makes evaporative cooling unusually effective, which is why older facilities here used it heavily. The water concerns make dry cooling more attractive for new builds, even though the energy penalty is real. Adiabatic hybrid designs have become the dominant choice for new construction in the metro because they balance water and energy in a defensible way.

The interesting frontier is what happens to the heat after it's rejected. In the conventional design, the heat dissipates into the air or the water and is gone. The atmosphere absorbs it. There is no recovery, no second use, no commercial value. The heat is a waste product.

This is where the conversation gets interesting.

The District Heating Question

Several European cities have spent decades building district heating networks: hot water distribution systems that move heat from a central source through insulated pipes to thousands of buildings for space heating and hot water. The district heating model is well-established in Stockholm, Copenhagen, Helsinki, and a long list of Scandinavian and German cities. It is unusual in North America and essentially nonexistent in the American Southwest.

Several European data centers are now exporting waste heat into district heating networks. Stockholm captures data center excess heat into the city's heat grid. Near Dublin, Amazon recycles waste heat from an AWS data center through the Tallaght District Heating Scheme, initially heating about 47,000 square meters of public buildings, commercial space, and 135 affordable apartments. In Denmark, Meta's Odense campus feeds surplus heat into the local Fjernvarme Fyn network, supplying heat for thousands of homes.

The economics work because the data center is generating heat as a continuous byproduct, the city has heat demand as a continuous baseline, and the infrastructure to move heat from one to the other already exists. The data center sells (or in some cases gives away) the heat. The utility avoids burning natural gas to produce it. The carbon math improves substantially.

This does not transfer cleanly to Phoenix. The Sonoran Desert does not have winter heating demand at residential scale. There are no district heating networks. The heat that a Phoenix data center produces has nowhere to go that creates value, at least not in the conventional building heating sense.

The Hot Phoenix Idea

There are emerging applications for waste heat that fit a desert climate better than district heating does.

Greenhouse agriculture, particularly for high-value crops in winter, can use 80-degree water year-round to maintain growing temperatures and to reduce energy costs. Several European data centers have integrated with adjacent greenhouse operations. Phoenix has an active greenhouse industry growing leafy greens, tomatoes, and specialty crops in controlled environments.

Aquaculture, particularly tilapia and other warm-water species, requires sustained water temperatures that match what data center waste heat produces. There is no significant aquaculture industry in central Arizona today, but the inputs (warm water, available land, established freight infrastructure) all line up.

Industrial processes that consume low-grade heat (food drying, brewing, certain chemical processes, water desalination feedwater preheating) could in principle be sited near data center campuses to use waste heat as a free input. The siting decisions to do that have not been made at scale, but the economic logic is intact.

The most interesting application may be water itself. Several pilot projects globally have explored using data center waste heat to drive water desalination or brackish-water purification through humidification-dehumidification (HDH) cycles. The technology is not mature, but the inputs (hot water, brackish source water) match what Arizona has access to. A future hyperscale campus could in principle treat its waste heat as the energy input for an adjacent water reclamation facility, closing some of the water loop the industry currently exports.

None of this is built today. All of it is plausible.

The Honest Frame on Heat

The waste heat question matters less in Arizona than in Northern Europe, but it is not zero. The heat that a Phoenix campus produces today goes to the atmosphere as a thermodynamic dead end. The heat that a Phoenix campus could produce in 2030, integrated with greenhouses or aquaculture or water reclamation, is a different conversation entirely.

The cooling architecture is changing because the chips changed. The heat output isn't changing in the aggregate (every watt in still equals every watt out), but the temperature of the rejected heat is going up as liquid cooling moves heat at higher concentrations. Higher-temperature waste heat is more useful, more recoverable, and more economic to integrate with secondary uses.

The next generation of data center campuses, the ones being designed today for delivery in 2027 and 2028, are being engineered with waste heat recovery as a possibility rather than an afterthought. The economic models that justify the integration are still developing. The point is that the question is being asked, in a way that 5 years ago it wasn't, and the answer for the desert Southwest is going to be different than the answer for Stockholm. That is fine. Both answers can be useful. Neither one has been written yet.

Sources