Heat Rejection: Chillers, Dry Coolers, Towers, Adiabatic & Economizers

Everything upstream of this chapter — the cold plate, the CDU, the secondary loop, the facility water loop — moves heat from the die to a pipe of warm water on the roof or in the yard. Heat rejection is the last meter: the act of dumping that heat into the atmosphere so the loop can come back cold enough to do it again. It is the least glamorous subsystem in the building and the one that most directly sets your annualized PUE, your annual water bill, and whether the local utility and the local watershed will tolerate you. The physics is a single sentence — you can reject heat to dry air, to evaporating water, or to a refrigeration cycle that pumps it uphill — but the consequences of which one you lean on, and for how many hours a year, compound into the largest controllable line in operating cost.

The master variable is the one you already set two chapters ago: the temperature of the water you push to the rack. Warm-water DLC (W32–W45 in ASHRAE's classes) is what makes the cheap rejection paths reachable at all; a low-temperature chilled-water loop forecloses them. So this chapter reads as a cascade from loop temperature → rejection technology → climate-bin economics → the water/energy operating point → the redundancy and ride-through the plant must carry, with a downstream cost attached to each fork. → loop temperature and the W-classes are set in Chapter 5.7; the metrics we score against are canonical in Chapter 15.1.

There are four families of heat-rejection plant, and the honest way to order them is by the resource each one spends to move a watt of heat into the sky. From most-water-cheapest-energy to least-water-most-energy:

Open cooling towers (evaporative). The thermodynamic champion: they reject heat by evaporating a fraction of the circulating water, so their floor is the ambient wet-bulb temperature, which can sit 8–15 °C below dry-bulb on a hot afternoon. That low approach is why a tower-cooled plant can hold a low PUE in climates where a dry cooler would have surrendered to a compressor. The price is liters: an evaporative plant is the dominant reason a data center's WUE lands at industry-average ~1.8–1.9 L/kWh rather than near zero, plus blowdown, drift, biocide, and an ASHRAE-188 program. Open towers also expose the facility loop to airborne fouling, which is why they almost always sit behind a plate heat exchanger.

Adiabatic / evaporatively-assisted dry coolers. The hybrid middle and the fastest-growing class in 2026. A dry cooler (a finned air-to-water coil with fans) runs bone-dry most of the year; on the handful of hot, high-dry-bulb hours it cannot meet alone, a pre-cooling stage wets pads or sprays a mist into the inlet air to pull the entering air down toward its wet-bulb, recovering capacity without committing to year-round evaporation. You spend water only on the worst bins — often a small single-digit percentage of annual hours — and trade it directly for not buying or running a chiller. This is the canonical water-vs-energy compromise instrument.

Dry coolers (sensible only). Zero process water by construction — they reject heat to dry air across a coil, like a giant car radiator. Their floor is the ambient dry-bulb, so their fortunes are entirely a function of how warm your loop is and how cool your air gets. A W40+ loop in a temperate or cold climate can run dry-cooler-only for most or all of the year, hitting a near-zero WUE and a low PUE simultaneously — the holy grail. The same dry cooler on a 18 °C chilled-water loop in Phoenix is a fan that cannot keep up, and a chiller has to carry the day.

Mechanical chillers (vapor-compression). The brute-force backstop: a refrigeration cycle that can deliver loop water colder than the ambient wet-bulb, at the cost of a compressor whose electrical draw is the single largest non-IT load in a chiller-based plant. Chillers exist to cover the bins the free-cooling and adiabatic paths cannot, and to serve any loop too cold for them. Every hour a compressor runs is a direct, measurable hit to PUE. In a 2026 warm-water DLC facility, the design goal is to make the chiller a rarely-used insurance policy — or to delete it entirely.

"Free cooling" is a misleading name — the fans and pumps still draw power — but the term of art means rejecting heat without running a compressor. There are two economizer families, and the choice between them is mostly settled by the fact that you are cooling liquid, not air.

Air-side economization ducts filtered outside air directly to the IT, or uses an air-to-air wheel. It is the workhorse of legacy air-cooled hyperscale halls and can deliver near-1.1 PUE in cool climates — but it scales with airflow, not with a liquid loop, and it imports humidity, particulates, and (in some regions) corrosive gases that the IT must tolerate. For a dense liquid-cooled AI hall, air-side economization cools at most the residual ~10–15% air fraction of the rack, not the ~85% the cold plates carry. Water-side economization is the relevant lever for liquid plants: when the ambient wet-bulb (tower) or dry-bulb (dry cooler) is low enough, a plate heat exchanger lets the rejection loop cool the facility loop directly and the chiller idles. Water-side economizing is what turns a chiller plant into a part-time chiller plant.

The engineering that matters is not the nameplate — it is the bin-hour count. Take your site's TMY (typical meteorological year) wet-bulb and dry-bulb distribution, lay your loop's approach temperature over it, and read off how many of the 8,760 hours each mode covers. A warm-water loop pushes the crossover so far that, in temperate and cold climates, the compressor-hours fall to a rounding error; water-side economizing on cooling towers below the wet-bulb threshold can cut annual cooling energy by roughly 40–65% in temperate climates versus mechanical chilling alone (Alfa Laval; ScienceDirect free-cooling studies, 2023–2025). The same overlay in a hot-humid site shows the compressor carrying real load — and tells you, quantitatively, what you are buying when you site there.

A modern AI heat-rejection plant is rarely one machine. It is a hybrid that runs three modes against changeover setpoints, and the controls logic that sequences them is where annualized PUE and WUE are actually won or lost — far more than the equipment selection. The canonical sequence, coldest ambient to hottest:

• Free-cooling (dry) mode. When ambient dry-bulb is comfortably below the loop return temperature, modulate dry-cooler fans only. Zero process water, lowest energy. A warm loop keeps the plant here for the bulk of the year in most of the populated temperate world.
• Adiabatic-assist mode. As dry-bulb rises and the dry coil starts to lose its approach, stage on pre-cooling — wet the pads or pulse the spray — to drag entering-air temperature toward the wet-bulb and recover capacity. You begin spending water, but only on these bins, and you have still not started a compressor.
• Mechanical (chiller / compressor) mode. Only when even the adiabatic-assisted approach cannot hold the loop setpoint do you commit the compressor. In a well-sequenced warm-water plant in a temperate climate, this can be a few hundred hours a year or fewer; in some Nordic and high-altitude siting, effectively never.

The decision embedded in the sequence is the changeover wet-bulb (and dry-bulb) setpoints and the hysteresis band around them. Set them too aggressive — chasing the last fraction of free cooling — and the plant hunts: fans and water and compressors cycle on the boundary, churning energy and water and stressing equipment. Set them too conservative and you start the compressor on hours the adiabatic stage could have covered, paying PUE you did not owe. This is a control-tuning problem, and it is the thermal sibling of the setpoint-stability problem the next-but-one chapter treats for the loop side. → control-loop tuning, anti-hunting, and the no-inertia transient in Chapter 5.12.

Strip away the equipment and heat rejection is a single trade: liters of water against kilowatt-hours of energy. Evaporate more, and your wet-bulb floor drops, your compressor runs less, and your PUE improves — at the cost of WUE. Refuse to evaporate, and your WUE goes to zero while your PUE rises and (in hot climates) a compressor draws the difference. There is no free lunch on this frontier; there is only the operating point that minimizes your blended cost of water and power under your permit constraints.

The numbers anchoring the frontier are stark. Industry-average WUE sits around 1.8–1.9 L/kWh, dominated by evaporative plants; best-in-class evaporative designs reach 0.3–0.7 L/kWh; and closed-loop / dry designs approach zero process water (Vertiv/NREL synthesis, 2025). Microsoft's reported FY2025 fleet WUE of ~0.30 L/kWh, with next-generation closed-loop designs at ~0 and saving over 125 million liters per facility per year, is the hyperscaler signal that the industry is moving down the water axis — accepting a PUE penalty in exchange for water that is, in many basins, the harder permit to get. The fork is genuine and it is regional: in a water-rich, power-expensive Nordic site, you may evaporate freely; in an arid, power-cheap Southwest US or Gulf site, the social license and the discharge permit push you toward dry rejection even though the climate punishes it.

The moment you commit to any evaporative stage — open tower or adiabatic assist — you have taken on a water-treatment and public-health obligation that is not optional and is increasingly a permit condition. A cooling tower is, by aerosol physics, the single highest-risk water system in an industrial facility for Legionella pneumophila: it warms water to the bacterium's growth range and then sprays it into the air. ANSI/ASHRAE Standard 188 (current edition 2021) is the consensus standard that governs the response — it requires every facility with a cooling tower to maintain a written Water Management Program: a documented analysis of the water system, defined control limits at critical control points (disinfectant residual, temperature), monitoring, corrective actions, and verification records. In the US this is reinforced by CDC guidance; in much of Europe and Asia, registration and sampling regimes for evaporative cooling are statutory. Skipping it is not an O&M shortcut — it is a liability and, in several jurisdictions, an operating-permit violation.

The operational mechanics that the program governs:

• Biocide regime. Oxidizing (chlorine/bromine) and non-oxidizing biocides on a dosing schedule, plus dispersants to break biofilm — the substrate Legionella actually hides in. The treatment chemistry interacts with corrosion and scale control, so it is a coupled water-chemistry problem, not three independent ones.
• Cycles of concentration and blowdown. As water evaporates, dissolved solids concentrate; "cycles of concentration" is how many times you let them concentrate before bleeding off (blowdown) and replacing with fresh makeup. More cycles save makeup water but raise scaling and corrosion risk and concentrate the blowdown stream you must discharge. This single ratio sets both your makeup-water draw and your discharge volume.
• Drift. Droplets carried out of the tower by the airflow — minimized by drift eliminators both to cut water loss and to limit aerosol dispersal of any contamination.

The discharge side is where arid siting collides with environmental permitting. Tower blowdown is a concentrated brine; a facility cannot always send it to a municipal sewer (volume, salinity, biocide residuals), and in a closed-basin arid region there may be no surface water to discharge to at all. Zero-liquid-discharge (ZLD) is the answer that unlocks those sites: an on-site treatment train — typically reverse osmosis followed by thermal evaporation/crystallization — that recovers most of the water for reuse and reduces the waste to a solid salt cake for landfill, discharging essentially no liquid. ZLD is energy- and capex-intensive and is rarely chosen for its own sake; it is chosen because it is the price of admission to evaporative cooling in a basin that will not issue a discharge permit otherwise. Framed as a decision: in an arid, discharge-restricted site, your real options are dry rejection (no blowdown to discharge) or evaporative rejection plus ZLD (blowdown recovered on site) — and the ZLD capex/energy is the toll for keeping the wet-bulb advantage. → discharge and withdrawal permitting as a siting critical-path item in Chapter 3.9; the water screen in siting in Chapter 3.1.

Heat-rejection plant must be sized not just for the worst climate bin but for the worst climate bin with a unit down for maintenance. Concurrent maintainability — the Tier-III-and-up property that any single component can be taken out of service for planned work without dropping below design load — drives N+1 (or N+2) on towers, dry coolers, pumps, and chillers, with isolation valves and headered piping so a cell can be drained and serviced while the rest carry the heat. The redundancy count is set by your maintenance philosophy and your tier commitment; the harder, AI-specific problem is what happens in the seconds after an unplanned loss.

This is where dense DLC breaks the old playbook. A legacy chilled-water plant carried enormous thermal inertia: thousands of liters of cold water in tanks and pipes, plus the thermal mass of a slab full of air. When a chiller tripped or the utility blinked, that stored cold rode the load through the seconds it took generators to pick up and pumps to restart. A direct-to-chip facility has almost none of that buffer. The water volume in the cold plates and CDUs is tiny relative to the heat flux; a GB200-class rack can carry ~115 kW of liquid load with seconds of ride-through, not minutes. Lose flow and the junction temperatures spike in seconds, and the GPUs throttle — or, past the limit, trip — long before a slow plant recovers. There is no chilled-water inertia to hide behind.