How Data Centers Stay Cool: Air, Liquid, Immersion and Closed-Loop Water

A pipe carrying 750 gallons of water a minute sounds like a water bill racing out of control. In a closed cooling loop, however, it can be the same water going around again. The number measures circulation, not the amount a data center takes from the town supply.

That distinction is the key to understanding data center cooling. “Air-cooled,” “liquid-cooled” and “closed-loop” describe different parts of a system, not three mutually exclusive kinds of building. An air-cooled server room can depend on an evaporative cooling tower outside. A liquid-cooled AI rack can send its heat to a dry cooler that evaporates no water at all.

Here is how the main approaches work, which types of data centers tend to use them, and what to ask when a company promises water-efficient computing. This is the engineering companion to our look at AI’s electricity, water and local-grid footprint, rather than another tally of the industry’s demand.

Follow the heat, not just the water

Nearly all the electricity used by servers ultimately becomes heat. Cooling moves that heat along a chain: chip → air or liquid → building cooling system → outdoors or a useful heat customer. Every link matters.

At the chip, the choice is a heat sink with moving air, a liquid-fed cold plate, or immersion in an electrically insulating fluid. At the building boundary, the choice might be a dry heat exchanger, a refrigeration plant, an evaporative tower, a body of cold water, or a heat-recovery network. These choices can be combined.

A closed loop tells you where the coolant goes. It does not, by itself, tell you where the heat goes.

Which cooling approaches fit which data centers?

Facility labels mostly describe ownership, customers and location. They do not prescribe cooling. The useful comparison is the workload density, existing infrastructure, local climate and service requirements behind each label.

Data center type	Common cooling fit	Main constraint
Hyperscale / AI	Contained air cooling for general servers; direct-to-chip liquid loops for dense AI; purpose-built outdoor heat rejection	Rack density, water supply and peak summer power
Colocation	Shared chilled-water or air plant, with rear-door exchangers and dedicated liquid-ready zones	Tenant hardware and agreed cooling limits
Enterprise or on-premises	CRAC units or CRAHs with containment; selective liquid retrofits	Keeping an existing room operating during upgrades
Edge and modular	Packaged air conditioning, sealed air-to-air exchangers, or factory-integrated liquid systems	Small footprint, harsh conditions and limited on-site service
Research and HPC	Warm-water direct-to-chip, sometimes immersion and heat reuse	High sustained loads and specialist support

Hyperscale is not synonymous with liquid cooling

A large cloud campus can contain storage, networking, conventional CPU servers and GPU clusters with very different thermal needs. Building from scratch makes it easier to install pipework and a coordinated heat-rejection plant, but retrofitting an older hall is a different project. A mixed campus may keep substantial air cooling even after adding liquid-cooled AI.

Colocation has to serve more than one hardware owner

A colocation operator provides the building, power, connectivity and cooling while customers bring their IT equipment. Shared infrastructure therefore needs clear boundaries: who supplies the coolant, maintains the distribution unit, detects leaks and guarantees the inlet temperature? A building advertised as “liquid-ready” still needs to demonstrate that a particular rack’s flow, pressure and heat load are supported.

Edge, enterprise and HPC face different priorities

A hospital server room cannot simply stop for a plumbing overhaul. A roadside edge enclosure may need to keep dust out and operate without permanent staff. A research supercomputer can justify a purpose-built warm-water plant. Modular describes how a facility is packaged and expanded, not a cooling method: the module can contain conventional air equipment, a liquid circuit, or both.

Air cooling: CRAC, CRAH and keeping the aisles separate

In traditional air cooling, server fans move room air over heat sinks. The warm exhaust returns to equipment that cools it again. Mitsubishi Electric’s 2025 guide distinguishes two familiar systems: a CRAC, or computer room air conditioner, uses direct-expansion refrigerant cooling; a CRAH, or computer room air handler, blows air across coils supplied by a chilled-water plant.

CRACs are a practical fit for smaller or legacy rooms because they can be deployed without a central chilled-water network. CRAHs and large fan-wall systems make more sense where a shared plant can serve a large hall. Neither name alone establishes the building’s water consumption: trace the heat to the outdoor equipment.

Hot/cold aisle layouts put server intakes toward a cold aisle and exhausts toward a hot aisle. Containment adds doors, roofs or barriers to stop supply and exhaust air mixing. Blanking unused rack spaces and sealing unintended bypass paths help prevent a room from being overcooled just to rescue a few hot spots. Both DOE’s cooling guidance and Mitsubishi emphasize airflow management before more elaborate upgrades.

Free cooling is not literally free

An air-side economizer uses favorable outdoor air conditions to reduce compressor cooling; filters and humidity controls protect the equipment. Indirect air systems transfer heat across an exchanger without bringing the outdoor air itself into the server hall.

A water-side economizer lets sufficiently cold outdoor conditions cool a liquid circuit through a heat exchanger, reducing or bypassing the mechanical chiller. Fans and pumps still consume electricity. And, as ENERGY STAR explains, a water-side economizer using an evaporative tower can still use substantial water. “Free cooling” means less refrigeration, not automatically zero water.

Evaporative cooling, adiabatic assist and chilled-water plants

Evaporation removes heat because changing liquid water into vapor takes energy. Direct evaporative systems cool an air stream with water; indirect systems keep the wetted process separate from server supply air. In an adiabatic-assist arrangement, wetted pads or a mist can precool outdoor air before it reaches heat-rejection equipment. Performance depends on humidity as well as temperature: dry air offers more evaporative cooling potential than already humid air.

A conventional water-cooled chiller plant has distinct circuits. The chilled-water loop carries heat from indoor coils or other exchangers to the chiller. Refrigeration moves it to a condenser-water loop, which takes it to an outdoor cooling tower. Some tower water evaporates; additional water is drained as blowdown to control dissolved minerals. Makeup water replaces these losses. The indoor chilled-water circuit can be closed while the tower still needs a continuing water supply.

This is often an electricity-versus-water choice, not a simple efficient-versus-wasteful choice. Towers can reduce compressor work, but consume water. Reclaimed wastewater can reduce demand for drinking-quality water without eliminating evaporation. Hybrid plants run dry when weather permits and use evaporative assistance when needed; their annual average and hottest-day demand can look very different.

Closed-loop water cooling: what the 750-gallon-a-minute clip shows

The Innovation Council post on X shares a short CNBC video of chilled-water pipes, AMD-branded hardware and a coolant distribution unit supplying Helios racks. In the full CNBC report’s transcript, that segment describes about 750 gallons per minute of chilled water circulating in a completely closed loop. It is a flow rate through the system, not 750 gallons of fresh water consumed every minute.

“750 gallons of water flowing every MINUTE, and it’s being reused.”

Innovation Council, sharing CNBC’s cooling video — watch the clip on X.

It is also important not to attach the wrong company to that number. The wider CNBC report separately visits an Equinix colocation facility in San Jose running NVIDIA Grace Blackwell racks. The Helios/CDU segment in the shared clip is an AMD example; the adjacent Equinix visit is not evidence that the same flow number applies to that site.

Inside a liquid-cooled rack, coolant picks up heat and returns to a coolant distribution unit (CDU). A liquid-to-liquid CDU transfers heat across an exchanger to a separate facility circuit while controlling conditions on the sensitive IT side. The two fluids need not mix. Real systems still require commissioning fills, maintenance and leak management, even when normal operation does not consume coolant through evaporation.

Closed loop plus a dry cooler can mean near-zero cooling-water consumption

A dry cooler is like an outdoor radiator: fans move ambient air across a finned coil carrying the warm liquid. No deliberate water evaporation is required. It works best when the outdoor air is sufficiently cooler than the required coolant temperature; hotter conditions can require a larger exchanger, more fan work, refrigeration or a different operating mode.

That is why CNBC’s November 2026 report separates the internal loop from the external cooling process. A closed rack loop connected to an evaporative tower is not the same thing as a whole facility designed for zero evaporative cooling water.

Microsoft’s zero-water designs go beyond recirculating the rack coolant

Microsoft’s design announcement describes chip-level liquid cooling with a closed circuit between servers and chillers, avoiding evaporation. It estimates that the design will avoid more than 125 million liters of water annually per data center, based on its FY2024 fleet-average withdrawal WUE. That is Microsoft’s design estimate, not a universal saving for every building or an independently measured result at every future site.

Microsoft also explicitly says replacing evaporation with mechanical cooling increases PUE, while warmer chip-cooling water and economizing chillers help limit the energy penalty. Its original rollout plan identified Phoenix and Mt. Pleasant pilots in 2026 and new sites beginning to come online in late 2027. Its June 2026 update reiterates the closed-loop, direct-to-chip design; that does not mean its entire existing fleet has already converted. “Zero water for cooling” also does not cover restrooms, construction, equipment manufacturing or the water used to generate purchased electricity.

Direct-to-chip and rear-door cooling: two different liquid upgrades

Cold plates take the hottest load out of the air stream

Direct-to-chip cooling places a metal cold plate against a processor or GPU. Liquid flows through channels inside the plate, absorbing heat without intentionally touching the electronics. Manifolds and hoses connect the plates to a CDU and facility cooling. Coolant chemistry, corrosion control, filtration, condensation avoidance and leak detection are engineering requirements, not optional extras.

The attraction for AI is heat concentrated in a very small area. NVIDIA’s DGX GB200 rack guide lists approximately 120 kW rack consumption and describes cold plates on CPUs and GPUs, with air cooling for other components. Its GB300 NVL72 reference architecture lists a liquid-cooled rack requiring up to 142 kW. These are published reference figures, not a promise that every vendor configuration draws the same power.

At those densities, simply forcing more room air through the hottest components becomes an increasingly difficult way to remove heat. Moving liquid to the chips shortens the heat-transfer path and allows denser computing. It does not remove the need to cool residual air loads or reject the captured heat outdoors.

Rear-door heat exchangers cool the exhaust, not the chip directly

A rear-door heat exchanger replaces a rack’s rear door with a liquid-fed coil. The servers still use air internally, but the warm exhaust gives its heat to water before returning to the room. Passive doors rely on server fans; active doors add fans. This can be a useful bridge for colocation and enterprise retrofits because the server hardware can remain largely conventional.

A rear door cannot make an air-only server equivalent to a purpose-designed liquid-cooled AI tray. Door weight, service clearance, airflow resistance, plumbing and the existing plant’s capacity all matter. It can also complement cold plates by capturing heat from components outside the chip loop.

Immersion: when the entire server goes into liquid

Immersion replaces the air around compatible electronics with a dielectric fluid that does not conduct electricity under its specified operating conditions. It is not a tank of ordinary water. As MANN+HUMMEL’s technical overview explains, the two main approaches differ in what the fluid does with the heat.

Single-phase immersion

The fluid warms up but remains liquid, then transfers heat through an exchanger to another cooling circuit. Removing much of the server fan requirement can be attractive for specialized HPC, dense compute and some modular installations. Fluid compatibility with cables, seals and components, handling procedures and manufacturer support must be checked before hardware is submerged.

Two-phase immersion

The working fluid boils at hot surfaces. Vapor rises to a condenser, turns back into liquid and returns to the bath. The phase change transports heat, but the condenser still needs an external heat sink. Fluid containment, loss control, servicing and the environmental and regulatory profile of the chosen coolant become part of the operating decision. Neither immersion method guarantees zero facility water use; that still depends on the external cooling plant.

Seawater, lakes, the ground and useful heat

Natural cold sources are a location advantage, not a universal recipe

At Lefdal Mine in Norway, the operator says seawater taken from the fjord at 8°C cools a separate closed freshwater circuit through heat exchangers. That freshwater can serve air handlers, direct liquid cooling or immersion. Saltwater is not being pumped through the processors. Lefdal advertises PUE of 1.08–1.15 depending on configuration and scale; that is an operator-reported site-specific range, not a rating for all seawater cooling.

Lake-water cooling follows the same broad principle when a suitable cold-water resource and permits exist. Intake ecology, fouling, corrosion, pumping energy and the temperature of returned water matter. A large withdrawal returned to a lake or sea is different from evaporation, but it is not an impact-free transaction.

Ground-coupled cooling transfers heat into soil or rock through boreholes, or uses a permitted groundwater system. DOE’s Epic Systems case study describes borefields and lake heat sinks supporting a 3.5 MW data center. This is shallow geothermal heat exchange, not the same thing as a geothermal power station supplying electricity. Long-term ground heat balance, geology and available land limit where it works well.

Heat reuse needs someone who actually wants the heat

A warm liquid loop can feed a district heating network, greenhouse or nearby building, often with a heat pump raising its temperature. The Hamina heat-recovery project was designed to send Google’s warm cooling water to a heat-pump plant serving the local heating network. Its published design is a useful example of connecting a data center to a heat customer, not evidence that every data center has one nearby.

The engineering challenge is matching temperature, distance and demand. Homes need less heat in summer while servers keep working. Pipes and heat pumps cost money and energy, and a fallback heat-rejection path is still needed. Recovering useful heat can replace other heating energy without making the data center’s own electricity demand disappear.

PUE and WUE: read the pair, then read the fine print

Power usage effectiveness (PUE) is total facility energy divided by IT equipment energy. Lower means less overhead for cooling, power conversion and other support systems; it does not say how useful or efficient the computation itself is. Google reports a 2025 fleet-wide PUE of 1.09, with continuous annual measurement across its stabilized large-scale facilities. The same page cites 1.54 as the average in Uptime Institute’s 2025 respondent survey. A single company’s large-scale fleet and a survey sample are not identical populations.

Water usage effectiveness (WUE) relates annual site water use to IT energy, usually in liters per kilowatt-hour. Microsoft’s June 2026 update reports 0.27 L/kWh for 2025. Before comparing companies, check whether the numerator counts withdrawals or consumption, cooling only or other site uses, and owned sites or leased capacity. A fleet average does not reveal a particular town’s peak-day demand.

The tradeoff is real but not absolute. Evaporation can reduce electrical overhead while increasing water consumption. Dry heat rejection can cut direct water demand but increase electricity use in hot weather. Warmer liquid loops, good controls and favorable climates can improve both. Water consumed at power stations also means a site with near-zero direct cooling WUE can still have an indirect water footprint.

What we know—and what a video cannot establish

We know that circulating flow is not the same as fresh-water demand, that cold plates make very dense AI systems practical, and that the final heat-rejection stage is central to water consumption. What the short X clip does not disclose is a whole-site annual WUE, peak-day water draw, the external plant’s complete design or the electricity needed to run it. Those require operating data, not an impressive pipe diameter.

Does liquid cooling always save water?

No. It can reduce fan work and support warmer coolant, but a liquid-cooled rack can still depend on an evaporative tower. Ask about the complete system, not just the rack.

What should a community ask before approving a data center?

Ask for annual and peak-day withdrawals and consumption, the water source, drought operating plans, the external cooling method, design versus measured PUE and WUE, and what backup mode runs on the hottest days. The most useful question is simple: after the water has cooled the chips, what cools the water?

Further reading