Liquid Cooling for AI Data Centers

Quick answer

Liquid cooling in a data center moves heat from chips into a liquid loop instead of air. There are two main methods. Direct-to-chip runs coolant through cold plates on the processors, with air handling the rest; it is the common choice for AI racks. Immersion sinks whole servers in a non-conductive fluid and remains a smaller niche.

A liquid cooling data center question usually starts with a rack. The servers you plan to buy draw far more power than the hall was designed for, and every watt of it turns into heat. Air can only carry so much of that heat before the fans, the noise and the airflow stop making sense.

This guide covers why AI racks moved to liquid, how the loop works from the chip to the outdoors, direct-to-chip and immersion cooling side by side, and the decisions a tenant has to make before signing for space. For air cooling and the other methods, see the overview of data center cooling systems.

Why do AI racks need liquid cooling?

Because they make too much heat in too small a space for air to remove efficiently. A Lawrence Berkeley National Laboratory researcher put it simply: a rack of conventional servers used about 3 to 5 kW, while racks of AI servers use as much as 100 kW (Berkeley Lab, November 2024). A December 2025 Pacific Northwest National Laboratory report for the Department of Energy says present-day GPU racks reach 120 kW (PNNL, December 2025).

To move that much heat with air, a server has to push more air through the same box. That costs power. ASHRAE's data center committee (TC 9.9) notes that fan power of 10% to 20% of server power is not uncommon in some systems, and that in a 50 kW rack this means at least 5 kW spent on fans alone (ASHRAE, 2021). Liquid carries heat far better than air, so the same heat moves with less energy and less space.

The same ASHRAE paper recommends that any future data center include the capability to add liquid cooling. For what these rack figures mean for a whole site, see how much power a data center uses.

How does liquid cooling work in a data center?

Heat moves through a chain of separate loops, each handing it to the next. The names vary between vendors, but the chain is much the same:

At the chip. Heat passes into a cold plate bolted to the processor (direct-to-chip) or straight into the fluid around the server (immersion).
The technology loop. A closed, tightly controlled coolant loop carries the heat from the servers to a coolant distribution unit (CDU).
The CDU. Pumps, filters, controls and a heat exchanger. It holds the coolant at the right temperature, pressure and flow, and passes the heat to the building's water without the two fluids mixing.
The facility water loop. The building's water system takes the heat from one or more CDUs and carries it out of the hall.
Heat rejection. Dry coolers, cooling towers or chillers outdoors release the heat to the air.

Halls with no water near the racks can use liquid-to-air units that cool the technology loop with room air. ASHRAE calls this a workable bridge but less efficient than rejecting heat to facility water (ASHRAE, 2021). It suits a few liquid-cooled racks, not a full AI hall.

Direct-to-chip vs immersion cooling: what is the difference?

Direct-to-chip cooling puts liquid only where the heat is densest; immersion puts liquid around everything.

Direct to chip cooling (also called cold plate cooling) mounts a metal plate with fine internal channels on each processor and sometimes on memory and power parts. Coolant, usually a water and glycol mix, flows through the plates. The rest of the server, such as drives, network cards and power supplies, is still cooled by fans and room air. The server looks and services much like an air-cooled one, with hoses and quick-disconnect fittings added.

Immersion cooling sinks the whole server in a fluid that does not conduct electricity. In single-phase immersion the fluid stays liquid and is pumped through a heat exchanger. In two-phase immersion it boils on hot parts and condenses on a cooled coil. Either way, nearly all the heat goes into the fluid and the room needs very little air cooling.

A middle path is the rear-door heat exchanger: a water-cooled radiator on the back of the rack that cools the exhaust air. It often catches the air-side heat left in direct-to-chip racks.

Direct-to-chip vs immersion cooling for AI racks, from a tenant's point of view
Question	Direct-to-chip (cold plate)	Immersion (single- or two-phase)
How it works	Coolant flows through plates on processors; fans and room air cool the rest	Whole server sits in a non-conductive fluid that absorbs nearly all its heat
Density fit	Today's dense AI racks; the common design for GPU systems	Very dense or unusual hardware; also used for crypto mining and high-performance computing
Fluid	Usually water and glycol in a closed loop; conducts electricity	Engineered dielectric fluid; does not conduct electricity
Facility water needs	Facility water loop to CDUs; the server sets the water temperature class	Facility water loop to the tank or rack heat exchanger
Retrofit into an air-cooled hall	Moderate: piping, CDUs and heat rejection, standard racks stay	Harder: tanks or sealed racks, floor loading, new service routines
Maintenance	Close to air-cooled servers, plus quick-disconnects, coolant checks and filters	Servers lifted out of fluid; tanks may need a lift; fluid checks and top-ups
What the tenant must specify	kW to liquid and to air per rack, water class, CDU ownership, leak detection	Fluid type, hardware and warranty compatibility, tank layout, service procedures

Which liquid cooling method do AI data centers actually use?

Most AI halls being built today use direct-to-chip cooling for the processors and air for the rest. Immersion is used, but in a smaller set of deployments.

The reasons are practical. Dense AI server designs increasingly ship with cold plates, and direct-to-chip racks keep standard layouts and service routines. Immersion needs tanks or sealed racks and a check that every part is compatible with the fluid. ASHRAE notes immersion has been popular in cryptocurrency mining, oil and gas and high-performance computing, and that its service and warranty questions need working through first (ASHRAE, 2021).

Public modeling reflects the same mix. Lawrence Berkeley National Laboratory's 2024 report on US data center energy treats AI servers as using a mix of liquid and air cooling, and models its liquid-cooled AI case at ASHRAE water class W45 (LBNL, December 2024).

So the honest answer for most AI tenants is liquid for the chips and air for the remainder. Plan the hall for both.

What does a tenant have to specify for liquid cooling?

Liquid cooling turns a few cooling questions into a long list of shared decisions between you, your hardware vendor and the facility. Put these in writing early.

The heat split. For each rack type, how many kW go to liquid and how many to air. The air share is smaller than in an air-cooled rack, but not zero. Take it from your server vendor's specification.
Facility water temperature. ASHRAE groups liquid-cooled equipment into water classes named after the highest supply temperature they accept: W17, W27, W32, W40, W45 and W+ (the number is degrees Celsius). Most cold plate systems run comfortably on W32 water, but ASHRAE expects required water temperatures to fall as chip power rises. Ask which class your hardware needs now and which the next generation will need.
CDUs. Who supplies, owns and maintains them, whether they sit in the rack or at the end of the row, and how much spare pump and CDU capacity covers a failure.
Heat rejection and water use. Dry coolers use little or no water; cooling towers and other evaporative systems use a lot. LBNL notes the warmer loops of liquid cooling allow more "free" cooling. More on data center water usage.
Leak detection and fluid care. Sensors under and inside racks, automatic shutoff valves, drip trays, and a plan for coolant chemistry and filtration. Cold plate coolant usually conducts electricity; immersion fluid does not.
Cooling during a power event. If the pumps stop, a liquid-cooled chip heats up very quickly. PNNL notes that putting cooling on uninterruptible power is increasingly common at AI sites (PNNL, December 2025). Ask what is on UPS.
Service and warranty. Who connects and disconnects servers, who handles fluid, and whether your hardware warranty covers the coolant or immersion fluid used.
Contract wording. ASHRAE notes that some colocation agreements are written to exclude water near the racks. Make sure yours says the opposite, and states the liquid and air capacity per rack.

If you are comparing providers, these sit alongside the power, fiber and lease questions in our GPU colocation buyer's guide.

What are the benefits of liquid cooling in a data center, and the trade-offs?

The main benefit is density: you can run racks that air cannot cool at all. The others follow from it.

Less fan power. Less energy goes into moving air through servers and the room.
Warmer water, more free cooling. The loop can run warmer, so sites lean on outdoor air or water instead of chillers for more of the year.
Less floor space per kW. Fewer, denser racks and shorter cable runs for the same cluster.
A path to lower water use. Warm loops pair well with dry coolers. This is a design choice: the same loop can be cooled with evaporative towers that use a lot of water.

The trade-offs are real. Pipes, pumps, CDUs and fluid management cost more to build and need skilled people to run. Leaks have to be designed for. Standards for fittings and fluids are still settling, and retrofitting an air-cooled hall can mean new piping, heat rejection and floor work. For dense AI racks, the real question is how much of each, liquid and air.

Where does Accretion fit for liquid-cooled AI capacity?

Accretion Compute is in development. No site is built or operating yet, and there is no capacity to lease today. The model: we build data center capacity at the source of energy, with power generated on site (behind the meter), and lease it to AI companies on multi-year terms. Cooling is designed for the tenant's density, air or liquid, and liquid cooling is supported by design. Halls are built to the tenant's specification, so the questions in this guide are the ones we work through with each tenant before a build.

Good fit if

Your racks need direct-to-chip liquid cooling, or a mix of liquid and air, and you want the hall designed around your heat split and water class.
You own your GPU servers, or plan to, and want space built to your specification with room to expand on the same site.
You want power from firm on-site generation rather than waiting for a grid connection.
You can plan ahead and want a multi-year lease sized to the life of your hardware.

Probably not a fit if

You need liquid-cooled capacity live today. Nothing is built yet; existing colocation fits better.
You want GPU-hours or rented GPUs. We don't own or rent out GPUs.
You need a published PUE, uptime figure or tier rating before you talk. We don't publish those.
You only need a few liquid-cooled servers. A retail colocation hall with liquid-to-air units may suit you better.

Talk to us

Frequently asked questions

Can an existing data center be retrofitted for liquid cooling?

Often, yes, but how far depends on the building. A few direct-to-chip racks can run in an air-cooled hall using liquid-to-air units. A full AI hall usually needs facility water piped to the rows, CDUs, more heat rejection outdoors, and enough floor strength and power for dense racks. Immersion is harder to retrofit than cold plates because it changes the rack layout and service routine as well as the plumbing.

Is immersion cooling better than direct-to-chip?

Neither is better in general. Immersion captures nearly all the server's heat and needs very little air, which suits very dense or specialized hardware. Direct-to-chip fits standard racks and service routines, and most dense AI server designs ship ready for cold plates, which is why it is the common choice for AI halls. Pick the method your hardware is designed and warrantied for, then design the facility around it.

What is a CDU in data center liquid cooling?

A coolant distribution unit (CDU) sits between your servers' coolant loop and the building's water. It has pumps, filters, sensors, controls and a heat exchanger, keeps the server coolant at the right temperature, pressure and flow, and hands the heat to the facility water without the two fluids mixing. CDUs sit in a rack or at the end of a row. Agree who owns and maintains them, and how much spare capacity there is.

What happens if liquid cooling leaks in a data center?

In a well-designed system, a leak is caught early and contained. Racks carry leak sensors, drip trays and valves that shut off flow to the affected section, and quick-disconnect fittings seal when unplugged. The risk differs by method: cold plate coolant usually conducts electricity, so a leak onto electronics can damage hardware; immersion fluids do not. Ask any provider how leaks are detected, isolated and reported.

Do liquid-cooled servers still need air conditioning?

With direct-to-chip cooling, yes. Cold plates take the heat from the processors, but drives, network cards and power supplies are still cooled by fans and room air, so the hall still needs air cooling for that share. Rear-door heat exchangers can remove it close to the rack. With immersion, very little room cooling is needed.

Is direct-to-chip cooling a closed loop?

The part that touches your servers is. Coolant circulates in a sealed loop between the cold plates and the CDU and is not consumed. Whether the site uses water depends on heat rejection outdoors: dry coolers use little or none, while cooling towers and other evaporative systems consume water. Ask which one the site uses.

Liquid Cooling for AI Data Centers: Direct-to-Chip vs Immersion