Liquid cooling for data centers

Why air cooling hits its limit

Most data centers are still cooled by air, and most racks are still modest. Uptime Institute's 2025 global survey found the average of operators' most common rack densities at almost 9 kW, up from 8.3 kW in 2024. More than 80% of operators reported no racks at or above 30 kW. Racks above 100 kW exist, the survey said, but are still rare.

AI hardware sits at the far end of that distribution. NVIDIA put its GB200 NVL72 rack at about 120 kW in April 2025, against about 20 kW for a traditional data center rack, and said that cooling those racks with air would require air chilled below freezing or airflow close to gale force. IREN designed its Childress, Texas, AI facility, announced in February 2025, for Blackwell racks at 200 kW each. Power is heading higher still. Google said in April 2025 that it expects machine learning to need more than 500 kW per IT rack before 2030, and NVIDIA plans 800-volt DC power for 1 MW racks starting in 2027, alongside its Kyber rack for Rubin Ultra GPUs.

The physics explains the shift. Google notes that water carries about 4,000 times more heat per unit volume than air for a given temperature change, and conducts heat about 30 times better. Chip power has climbed from about 100 W to accelerators that exceed 1,000 W, so the heat leaving each square centimeter of silicon is beyond what fans can practically move.

Operators feel the pressure too, but adoption is gradual. Uptime Institute's 2025 Cooling Systems Survey, with 1,033 respondents, found that rising rack density remains the top driver for direct liquid cooling, and Uptime described adoption in July 2025 as slow and steady, with most operators still cooling IT equipment with air. For the wider picture of how density drives power, sites and costs, see our guide to AI data centers, and for one operator's view of the problem, designing for the AI cooling crisis.

Four ways to cool with liquid

Liquid cooling is not one technology. Four designs dominate, and they differ in how close the liquid gets to the chip, what fluid they use and how much heat one rack can reject. The kW figures below are vendor ratings for specific products, dated where the vendor gave a date. Real capacity depends on water temperature, hardware and layout.

Rear-door heat exchangers

A rear-door heat exchanger replaces the back door of a standard rack with a water coil. Servers still blow hot air out the back, and the coil takes the heat into facility water before the air re-enters the room. Active doors add fans to help the servers push air through the coil. Motivair rates its ChilledDoor at up to 75 kW per rack with 55°F to 65°F supply water, per its 2023 brochure. Vertiv launched its CoolLoop RDHx globally in March 2025 at up to 80 kW per rack, and said it can run on its own or alongside direct-to-chip cooling.

The appeal is that servers do not change. Any air-cooled hardware works, and the room stays roughly neutral in temperature. The limit is that the heat still travels through air inside the server, so rear doors cap out well below the densities of current AI racks. They are now often paired with cold plates to catch the heat that cold plates miss.

Direct-to-chip cold plates

Direct-to-chip cooling, also called direct liquid cooling or cold-plate cooling, puts a metal plate with internal channels on each GPU and CPU. Coolant, usually treated water or a water and propylene glycol mix, flows through the plate and carries heat to a coolant distribution unit. It is the standard design for NVIDIA's rack-scale systems: the GB200 NVL72 rack at about 120 kW, and the Vera Rubin NVL72, whose compute tray NVIDIA described in October 2025 as 100% liquid cooled with 45°C liquid cooling. Google has used cold plates since its TPU v3 in 2018 and said in April 2025 that it runs liquid cooling at gigawatt scale across more than 2,000 TPU pods, with about 99.999% uptime.

Cold plates do not capture all of a server's heat. Memory, power supplies and network cards still shed heat into air, which is why many halls run cold plates and a rear door or some air handling side by side. Per-GPU power and operating cost for current liquid-cooled parts, such as the NVIDIA B300 SXM and the AMD MI355X, are listed on aPanel.

Single-phase immersion

In single-phase immersion, servers sit vertically in a tank of dielectric fluid, usually a synthetic oil, that never boils. Pumps move the warm fluid through a heat exchanger, which hands the heat to a water loop. Fans are removed. GRC's ICEraQ system, on its current product page, is rated at up to 368 kW per rack with 13°C chilled water, or 200 kW with 32°C water and no chiller.

Immersion captures nearly all of a server's heat, removes fan power and keeps dust out. It also changes operations. Technicians lift servers out of fluid, hardware warranties must cover immersion, and floors must carry heavy tanks: GRC lists floor loading of 903 kg per square meter. Single-phase is the dominant immersion design in Bitcoin mining, and Bitmain sells mining machines built specifically for single-phase dielectric baths.

Two-phase immersion

Two-phase immersion uses a fluid with a low boiling point. Heat from the chips makes it boil, the vapor rises to a condenser coil in the tank lid, turns back to liquid and drips back down. Microsoft ran a two-phase tank in production in Quincy, Washington, in April 2021, with a 3M fluid engineered to boil at 122°F, and reported 5% to 15% lower power use per server. LiquidStack, spun out of Bitfury's Allied Control, rated its DataTank at up to 252 kW per 48U rack in April 2021.

Boiling is a very efficient way to move heat, but the fluids are the catch. Almost all commercial two-phase fluids have been fluorinated chemicals, which is why the design has stalled while single-phase and cold plates have spread. Uptime noted in July 2024 that renewed interest in two-phase cooling is now mostly in direct-to-chip form, where a refrigerant boils inside a cold plate, rather than in open tanks.

Rear-door heat exchanger

Water coil on the rack door. Up to 75 kW (Motivair, 2023) or 80 kW (Vertiv, March 2025) per rack. Works with unchanged air-cooled servers.

Direct-to-chip cold plates

Water or glycol through plates on GPUs and CPUs. The standard for NVIDIA GB200 (about 120 kW per rack) and later systems; IREN designed for 200 kW racks in 2025.

Single-phase immersion

Servers in non-boiling synthetic dielectric fluid. Up to 368 kW per rack with chilled water on GRC's rating. Widely used in Bitcoin mining.

Two-phase immersion

Servers in a fluid that boils and condenses. Up to 252 kW per 48U rack on LiquidStack's 2021 rating. Fluid supply hit by the PFAS phase-out.

Fluids and the PFAS problem

The fluid is the least visible and most regulated part of a liquid cooling system. Cold-plate loops mostly run on treated water or propylene glycol. Google's Project Deschutes CDU specification, for example, is designed for deionized water or PG25, a 25% propylene glycol mix. Single-phase immersion uses synthetic dielectric fluids such as GRC's ElectroSafe. Two-phase immersion has relied on fluorinated fluids such as 3M's Novec and Fluorinert lines, which are PFAS, the class of persistent chemicals often called forever chemicals.

On December 20, 2022, 3M announced it would exit all PFAS manufacturing, including fluorinated fluids, by the end of 2025. Its manufactured PFAS products brought in about $1.3 billion in annual net sales at the time. DCD reported in February 2023 that two-phase vendors including Zutacore, LiquidStack and Wiwynn referenced Novec in their materials, that the EPA had moved in January 2023 to tighten handling and reporting rules for certain PFAS, and that the EU was weighing broader restrictions. Single-phase immersion providers were largely unaffected. Uptime wrote in July 2024 that 3M's decision marked the beginning of the end for Novec in liquid cooling and hit operators who had committed to two-phase immersion hard.

Replacements are arriving slowly. Chemours first announced Opteon 2P50, a hydrofluoroolefin dielectric fluid for two-phase immersion, in August 2023. In May 2025 it signed Navin Fluorine to manufacture it, with commercialization planned beginning in 2026, a year later than its original 2025 target. Chemours also began a trial of the fluid with NTT Data and Hibiya Engineering in Japan, announced in March 2025. Buyers weighing two-phase systems in late 2026 should confirm fluid supply, regulatory status and disposal terms in writing.

CDUs and facility water

Every liquid cooling system except the simplest rear door needs a coolant distribution unit. The CDU is a pump and heat exchanger package that sits between two loops. The technology cooling system, or secondary loop, carries clean, filtered coolant to the cold plates or tanks. The facility water system, or primary loop, carries the heat away to chillers, dry coolers or cooling towers. Keeping them separate protects the delicate cold-plate channels from the dirt and chemistry of building water, and lets each loop run at its own pressure and temperature.

CDUs have grown with the racks. Vertiv announced its CoolChip CDU 2300 in November 2024 at 2.3 MW in a liquid-to-liquid design, and by May 2026 sold a CDU family from 100 kW to 2.3 MW. Google's Project Deschutes specification, version 0.8 effective July 2025 and published through the Open Compute Project, rates one CDU at 2 MW of heat with a 3°C approach temperature, 500 gallons per minute on each side, and N+1 pumps. Google said its Deschutes CDUs have achieved about 99.999% availability across its fleet since 2020. By October 2025, Boyd, Vertiv, Stulz, Delta, nVent, Cooler Master, Envicool and Nidec were showing CDUs built to the design.

Liquid-to-liquid or liquid-to-air

A liquid-to-liquid CDU needs facility water piped to the white space. A liquid-to-air CDU instead rejects heat from the coolant into the room through its own coil and fans, so it can serve a few liquid-cooled racks in a building without facility water at the rack. Vertiv's 350 kW liquid-to-air CoolChip unit, announced in November 2024, is aimed at exactly that retrofit case. The trade-off is that the heat ends up back in the room air, so the building's air system must still remove it.

Warmer water

Liquid cooling lets facility water run much warmer than the cold water air-cooled halls need. NVIDIA's Vera Rubin racks use 45°C liquid cooling. Warm water can often be cooled by dry coolers on outside air for much or all of the year, cutting chiller energy and evaporative water use. Berkeley Lab's December 2024 report modeled AI liquid-cooled facilities at ASHRAE class W45 and noted that the higher operating temperature of liquid cooling supports better water and energy performance.

PUE and water use

Power usage effectiveness (PUE) is total facility energy divided by IT energy. Uptime's 2025 survey put the weighted global average at 1.54, essentially flat for the sixth year in a row. Newer facilities, built within the past five years, averaged 1.48, and sites of 20 MW or more averaged 1.44. NVIDIA, citing McKinsey, said in April 2025 that cooling can account for up to 40% of a data center's electricity.

Liquid cooling attacks that overhead in two ways. Removing server fans cuts IT power, and warmer water cuts chiller power. Microsoft reported 5% to 15% lower power per server in its 2021 two-phase tank. GRC claims a partial PUE below 1.03 for its immersion system, a vendor figure that covers only the cooling system rather than the whole building. NVIDIA claimed in April 2025 that its liquid-cooled GB200 NVL72 delivers 25 times the energy efficiency and 300 times the water efficiency of traditional air-cooled infrastructure, comparisons that combine newer chips with the cooling change. Uptime's survey suggests that wider use of liquid-cooled hardware may lift the share of efficient facilities, but it has not yet moved the global average.

Water usage effectiveness (WUE) is liters of water consumed on site per kWh of IT energy. Berkeley Lab found the US average site WUE stayed just over 0.36 L/kWh through 2023 and projected 0.45 to 0.48 L/kWh by 2028 as AI growth adds load. Microsoft reported a fleet average of 0.30 L/kWh in fiscal 2024. Since August 2024 all its new data center designs use closed-loop chip-level cooling with no evaporation, which Microsoft says avoids more than 125 million liters of water per year per data center, with a nominal rise in energy use. Pilot sites in Phoenix and Mt. Pleasant, Wisconsin, are expected online in late 2027.

The lesson is that liquid cooling does not by itself save water. A closed loop rejecting heat through dry coolers uses almost none. The same loop rejecting heat through evaporative cooling towers still evaporates water, only less per unit of computing than an air-cooled hall.

Costs: retrofit or new build

Liquid cooling raises the capital cost of a building. Datacentres.com, reporting JLL's 2026 forecast in May 2026, put the global average construction cost at about $11.3 million per megawatt, with premium markets above $15 million. It said liquid cooling adds $1 million to $2 million per megawatt over an air-cooled design, and AI-ready requirements add 15% to 25% to shell-and-core cost. The cooling system itself is only part of the bill: heavier floors, piping, leak detection and higher-capacity power distribution come with it.

Retrofitting is the harder path. Uptime reported in July 2025 that operators most often judge a liquid cooling system by how readily it integrates with their existing infrastructure. A typical retrofit adds rear doors or liquid-to-air CDUs to a few rows, which serves moderate densities without new facility water pipework. Bringing chilled or warm water to every row, reinforcing floors for immersion tanks and adding the power for 100 kW-plus racks usually costs closer to a new build.

New builds can design around liquid from the start: facility water to every row, warm-water heat rejection, and power distribution sized for the next rack generation. Speed matters as much as cost. IREN's 75 MW Childress facility, designed for 200 kW Blackwell racks with direct-to-chip cooling, was presented in February 2025 as a response to scarce liquid-cooled capacity. Modular, factory-built liquid-cooled halls are one way operators shorten that timeline; see MOD42's modular AI data centers.

Standards: ASHRAE and OCP

Two bodies set most of the reference points. ASHRAE Technical Committee 9.9 publishes the thermal guidelines that define allowable temperatures for IT equipment. Its fifth edition, published in March 2021, renamed the liquid cooling facility water classes by their maximum supply temperature: W17, W27, W32, W40, W45 and W+ for anything above 45°C. Equipment compliant with a class must run unthrottled at every temperature in that class, so a building designed for W32 should support any W32-compliant hardware. Higher classes allow more hours of chiller-free cooling.

The Open Compute Project now publishes much of the open liquid cooling work. Google's Project Deschutes CDU specification is one example. NVIDIA contributed portions of the GB200 NVL72 electro-mechanical design and said in October 2025 it would contribute the Vera Rubin rack and compute tray innovations. Google, Meta and Microsoft are also working at OCP on Mt. Diablo, a standard for +/-400 VDC power to racks of up to 1 MW, announced in April 2025. Shared specifications like these matter to anyone buying CDUs, manifolds and quick-disconnect fittings from more than one vendor.

What Bitcoin miners already know

Bitcoin miners ran liquid cooling at industrial scale years before AI made it mainstream. Their machines run flat out around the clock, so cooling cost drives profit directly, and that pushed miners to hydro and immersion early. LiquidStack's history shows the link: it began as Allied Control, which built a 500 kW two-phase immersion data center in Hong Kong in 2014 and was bought by Bitfury in 2015. By April 2021, Bitfury and Allied Control had deployed more than 160 MW of two-phase immersion, according to Network World, when Wiwynn invested $10 million to spin LiquidStack out for cloud data centers.

Today mining hardware is sold in liquid-cooled versions. Bitmain launched the Antminer S21 XP Immersion and S21 XP Hydro on June 17, 2024: the immersion model at 300 TH/s and 4,050 W for single-phase dielectric baths, and the hydro model at 473 TH/s and 5,676 W on a closed water loop. Immersion also lets miners overclock. Braiins showed in May 2022 that an Antminer S19j Pro rated at 104 TH/s on air could reach about 136 TH/s at 4,000 W in immersion with its firmware. It estimated immersion infrastructure at about $280 to $600 per kW, against $150 to $400 per kW for air.

Much of that experience carries over to AI: specifying dielectric fluids, managing tanks and manifolds, running closed water loops and dry coolers, monitoring leaks and fluid chemistry, and training crews to service hardware that is wet. IREN, a Bitcoin miner, designed its Childress AI facility for direct-to-chip cooling with power redundancy. The differences are real, though. AI tenants expect redundant power and cooling, strict service-level agreements and cold plates rather than tanks. For examples from the mining side, read about the return on hydro and immersion mining, closed-loop hydro reliability, Arctic Systems' immersion work and Bitmain's liquid cooling for AI data centers.

This crossover between miners who know liquid cooling and AI operators who need it is the subject of Intelligence Disrupt 2027, March 22-24, 2027, at the Irving Convention Center at Las Colinas in Dallas-Fort Worth, co-located with Mining Disrupt, the Bitcoin mining expo. See passes or exhibit.

Questions

What is liquid cooling in a data center?

Liquid cooling moves heat out of servers with a fluid instead of room air, either through cold plates on the chips, a water coil on the rack door, or by submerging servers in a dielectric fluid. Google notes that water carries about 4,000 times more heat per unit volume than air for the same temperature change.

At what rack density do you need liquid cooling?

There is no single cutoff, but rising rack density is the top reason operators adopt direct liquid cooling, according to Uptime Institute's 2025 cooling survey. NVIDIA's GB200 NVL72 rack runs at about 120 kW, which NVIDIA says air cannot practically cool.

What is the difference between single-phase and two-phase immersion cooling?

In single-phase immersion the fluid stays liquid and is pumped through a heat exchanger. In two-phase immersion the fluid boils on the hot components and condenses on a coil above the tank; Microsoft's 2021 tank in Quincy, Washington, used a 3M fluid that boils at 122°F.

Why is two-phase immersion cooling affected by PFAS rules?

Most two-phase fluids, including 3M's Novec line, are fluorinated chemicals. In December 2022 3M said it would exit PFAS manufacturing by the end of 2025, and Chemours expects to commercialize its HFO-based replacement, Opteon 2P50, beginning in 2026.

What does a coolant distribution unit (CDU) do?

A CDU separates the clean coolant loop that runs to the servers from the facility water loop, using a heat exchanger and pumps. Google's Project Deschutes specification, published through the Open Compute Project in 2025, rates one CDU at 2 MW with a 3°C approach temperature.

How much does liquid cooling add to data center construction costs?

Datacentres.com, reporting JLL's 2026 forecast, puts liquid cooling at about $1 million to $2 million per megawatt over an air-cooled design, against an average build cost of about $11.3 million per megawatt.

Does liquid cooling reduce data center water use?

It can. Microsoft's closed-loop, chip-level cooling design, used for all new Microsoft data center designs since August 2024, avoids more than 125 million liters of water per year per data center compared with evaporative cooling.

What is the average data center PUE?

Uptime Institute's 2025 survey put the weighted average annual PUE at 1.54, little changed for six years. Facilities of 20 MW and above averaged 1.44.

More guides

• AI data centers: power, cooling, sites and the build-out
• GPU cloud: providers, prices and the power behind them
• Data center power: how much it takes and how to get it

Sources

1. Uptime Institute, Global Data Center Survey 2025 (full report) · 2025-07
2. Uptime Institute, Global Data Center Survey 2025 summary · 2025-07-29
3. Uptime Institute, 2025 Cooling Systems Survey results · 2025-06-26
4. Uptime Institute, DLC adoption remains slow and steady (Data Report 181) · 2025-07-30
5. Uptime Institute, Interest in two-phase cooling warms · 2024-07-05
6. NVIDIA, Blackwell platform boosts water efficiency with liquid cooling · 2025-04-22
7. NVIDIA, 800 VDC architecture for next-generation AI factories · 2025-05-20
8. NVIDIA, Partners drive gigawatt AI factories for Vera Rubin · 2025-10-13
9. Google Cloud, Enabling 1 MW IT racks and liquid cooling at OCP EMEA Summit · 2025-04-29
10. Open Compute Project, Project Deschutes data center facilities specification · 2025-09-05
11. DCD, Companies show off Google-inspired Project Deschutes CDUs · 2025-10-17
12. Upsite Technologies, ASHRAE fifth edition thermal guidelines: liquid cooling chapter · 2024-07-31
13. Vertiv, Global launch of CoolLoop RDHx rear door heat exchanger · 2025-03-18
14. Vertiv, Two new high-capacity coolant distribution units · 2024-11-19
15. Vertiv, Liquid cooling portfolio expansion in EMEA · 2026-05-26
16. Motivair, ChilledDoor rear door heat exchanger brochure · 2023
17. GRC, ICEraQ Flex single-phase immersion system · 2024
18. Network World, LiquidStack launches as a stand-alone company · 2021-04-05
19. Microsoft, To cool datacenter servers, Microsoft turns to boiling liquid · 2021-04-06
20. Microsoft, Next-generation datacenters consume zero water for cooling · 2024-12-09
21. 3M, 3M to exit PFAS manufacturing by the end of 2025 · 2022-12-20
22. DCD, Two-phase cooling will be hit by EPA rules and 3M's PFAS exit · 2023-02-02
23. DCD, Chemours selects Navin Fluorine to make its two-phase immersion fluid · 2025-05-12
24. Berkeley Lab, 2024 United States Data Center Energy Usage Report · 2024-12
25. Datacentres.com on JLL construction costs · 2026-05-12
26. IREN, Q2 FY25 results (Childress Horizon 1) · 2025-02-12
27. Braiins, The economics of immersion cooling for Bitcoin miners · 2022-05-09
28. Hashrate Index, Antminer S21 XP Immersion hardware overview · 2024-07-01