Data Center Liquid Cooling: CDUs, Cold Plates, Rear Doors and TCS Piping

Data center liquid cooling carries heat away from processors and accelerators in a pumped liquid circuit rather than in room air, and Paul Industries installs, pipes, flushes, pressure-tests, fills, starts up and maintains the piping and connections that circuit depends on. We do not manufacture cold plates, coolant distribution units, rear doors or servers. We build the facility-side branches, the technology cooling system piping, the CDU connections, the row manifolds and the rear door connections, then prove them clean and tight before the first server is plugged in. The plant, loops and water use are covered on our data centers page and our data center cooling systems page.

What is data center liquid cooling?

It is any cooling arrangement in which a liquid, rather than air, collects heat at or very near the electronics and carries it out of the white space to a heat exchanger. The three forms are cold plates on the hottest chips, rack-mounted coils that catch exhaust air, and tanks of dielectric fluid holding whole servers; all three end at a water connection to the heat rejection plant.

The driver is heat density. An ASHRAE TC 9.9 white paper notes that a 40 to 50 kilowatt air-cooled rack can need up to about 5,000 cubic feet per minute of air, while the best perforated floor tile delivers around 1,900, and that fans can draw 10 to 20 percent of server power in dense air-cooled equipment. Analysts put the practical ceiling for air near 50 kilowatts per rack, while accelerator racks are planned at 100 kilowatts and above, which is why GPU liquid cooling has moved from research labs into commercial halls. A 2024 Lawrence Berkeley National Laboratory report for the Department of Energy put data centers at about 4.4 percent of United States electricity in 2023 and 6.7 to 12 percent by 2028. An AI data center liquid cooling system is a precondition for that hardware, not an efficiency option.

A data center liquid cooling diagram, loop by loop

Read from the outside in, the diagram has five segments: heat rejection equipment, the facility water system, the coolant distribution unit, the technology cooling system and the IT equipment itself. ASHRAE calls the rack-side loop the technology cooling system, or TCS; many specifications call it the technical cooling system or the secondary fluid network. The two sides of the CDU follow different water quality rules, pressure ratings and often owners.

A data center liquid cooling system built this way keeps the delicate part of the circuit small and isolated from building water. Liquid cooling systems for data centers can skip the CDU and run facility water straight to cold plates, but ASHRAE notes that those designs need cold plates with larger passages, filtration around 100 microns at the rack and facility water that never drops within 2 degrees Celsius of the room dew point, so most installations keep the CDU.

What is direct to chip cooling?

Direct-to-chip cooling clamps a liquid-filled cold plate onto each processor or accelerator package so that the chip’s heat conducts through a thin interface material straight into the coolant. The direct to chip cooling cold plate is a copper or similar block whose internal fins are often only a few tenths of a millimeter apart. Cold plate cooling is widely used in new AI halls because it fits standard rack and server form factors and leaves the rest of the server serviceable in the usual way.

How does direct to chip cooling work, step by step?

The CDU pumps conditioned coolant through a row header to a manifold in each rack. Hoses with dripless quick disconnects feed each server, where the fluid passes through its cold plates and returns to the CDU exchanger. The CDU holds supply temperature at least 2 degrees Celsius above the room dew point, as ASHRAE recommends, so nothing condenses on server boards. Water carries one kilowatt with about 1.4 liters per minute at a 10 degree Celsius rise, or about 0.38 gallons per minute per kilowatt at an 18 degree Fahrenheit rise, and glycol mixtures need somewhat more because their heat capacity is lower. Cold plate liquid cooling captures the processor and accelerator heat, which is most of the load; memory, drives, network cards and power supplies still shed heat to air unless they carry their own plates.

Single-phase vs two phase direct to chip cooling

Single-phase systems, the more common type, keep the coolant liquid throughout. In two-phase systems a low-boiling dielectric fluid or refrigerant boils inside the cold plate and carries heat as vapor to a condenser, usually inside a refrigerant-to-liquid CDU that hands the heat to facility water. Boiling holds the chip near the fluid’s saturation temperature, and because the heat travels as latent heat the loop needs much less pumping than water for the same load. The rack-side circuit is sealed and charged by its supplier; the building side is still a facility water connection.

Direct to chip cooling vs immersion cooling

The difference is what touches the electronics: a cold plate puts liquid on a few components inside an otherwise normal server, while immersion submerges the entire server in a dielectric fluid. For the mechanical contractor the bigger difference is where the piping ends and what the floor has to carry.

Data center liquid cooling solutions are not mutually exclusive. Many halls run cold plates on accelerator racks, rear doors on storage and network racks, and conventional air for everything else, all fed from one facility water system at different temperatures.

How does a coolant distribution unit work?

A coolant distribution unit takes facility water on its primary side, passes it through a plate heat exchanger to cool the rack-side fluid on its secondary side, and pumps that secondary fluid to the IT equipment at a controlled temperature, pressure and flow. Bypass loops, proportional valves and variable-speed pumps hold supply temperature, and the controls raise the setpoint when the room dew point climbs. The CDU also lets the TCS run at lower pressure, with a fluid chosen for the cold plates rather than the plant.

The liquid cooling components inside a typical unit are a plate exchanger with plate spacing of about 2 to 8 millimeters, redundant pumps, secondary-side filters, an expansion tank or reservoir, a makeup connection, instruments, a leak or level alarm and a controller tied to the building management system. All of that arrives with the unit. The installer provides what surrounds it: primary supply and return with isolation valves, strainers, unions or flanges for exchanger removal, secondary connections to the row headers, drains, high-point vents and the instrument wiring interface.

Liquid-to-liquid, liquid-to-air and in-rack CDUs

A liquid-to-liquid coolant distribution unit is the efficient choice wherever facility water can reach the row. A liquid to air coolant distribution unit is a bridging tool: it lets cold plate servers run in an air-cooled hall, but every kilowatt it removes is handed back to the room, so the existing CRAH capacity still sets the limit. An in-rack coolant distribution unit confines a failure to one rack, at the cost of more units and facility water branches.

How does a rear door heat exchanger work?

A rear door heat exchanger replaces the back door of a server cabinet with a finned water coil, so the hot exhaust from the servers passes through the coil and gives up its heat to the water before it re-enters the room. Lawrence Berkeley National Laboratory describes rear doors as a bridge between cold air and warm-liquid cooling, because they eliminate bypass and recirculation air around the rack.

Active vs passive rear door heat exchanger

A passive door has no fans: the server fans push exhaust through the coil and work slightly harder against its resistance. An active door adds its own fans and usually a larger coil, which gives a closer approach between entering water and leaving air and supports higher rack loads. Berkeley Lab’s own installations found that active doors unload the server fans, so measured PUE goes up because door fans count as facility power, yet total energy goes down because door fans are more efficient. Rear door performance falls when the load is concentrated in one part of the rack, so select doors against the real equipment layout and water temperature.

A chilled water rear door heat exchanger can be fed straight from the facility chilled water system, from a CDU secondary loop, or from a dedicated closed loop cooled by tower water, provided the supply stays above the room dew point. Each door needs supply and return hoses with isolation valves, a means of setting flow, from a manual throttling valve to a calibrated balancing valve to an automatic control valve, and filtration to the door maker’s specification. Berkeley Lab added filtration for exactly that reason when it expanded rear door cooling.

Data center liquid cooling temperature: the ASHRAE W classes

ASHRAE TC 9.9 classifies liquid-cooled IT equipment by the highest facility water supply temperature it accepts; the 2021 fifth edition of its thermal guidelines renamed the classes so the number is that limit in degrees Celsius, with a shared lower limit of 2 degrees. The class decides what heat rejection plant the building needs.

The rack-side supply temperature is the facility water temperature plus the CDU’s approach, so a hall designed for W32 water delivers TCS fluid a few degrees warmer than that. ASHRAE’s own guidance carries a warning for owners: most cold plate products today run comfortably on W32 water, but higher-power processors and memory may need cooler water, pushing some sites down toward W27 and trimming economizer hours. Chiller capacity in reserve, and piping sized for the higher flow a cooler supply implies, keeps that option open.

Data center liquid cooling piping: materials, manifolds and valves

Technology cooling system piping is commonly welded 304L or 316L stainless steel, which tolerates treated water and glycol and can be cleaned and passivated after fabrication. Copper appears on many wetted-material lists and is joined by brazing, which ASHRAE recommends over soldering; soldered joints are porous, and neither method belongs on steel or stainless. EPDM hose and certain engineered plastics also appear where the CDU and cold plate makers list them. Galvanized pipe is excluded, aluminum is avoided unless it is isolated from the other metals, and carbon steel stays on the facility side, where a steel-specific inhibitor protects it. Every pipe, fitting, gasket, valve seat, hose and sealant must appear on the end-to-end wetted-materials list. Our stainless steel piping page covers grades and joining, and we use orbital welding for smooth, repeatable internal beads.

Row headers and the rack manifold

A data center liquid cooling manifold is the vertical distribution bar inside a rack, with supply and return headers and a quick disconnect for every server. Row headers feed the rack manifolds from the CDU, usually overhead in new halls and underfloor in many retrofits. Good header design balances flow with reverse return or flow-limiting devices at each rack, and avoids two problems ASHRAE singles out: low-velocity zones and dead legs, which encourage biofilm, and branches installed for future racks, which sit stagnant for months and grow scale and corrosion product that lands in the first server connected. Future taps belong close to the header with an isolation valve and a flushing connection, not as long capped stubs.

Quick disconnects need to be dripless and rated for the loop’s pressure and temperature. The Open Compute Project’s Universal Quick Disconnect specification defines the coupling interface openly so compliant parts from different makers mate.

Valves, air, filtration and expansion

Data center liquid cooling pumps sit in two places: redundant pumps inside each CDU move the TCS fluid, and variable-speed plant pumps move facility water to the CDUs. N+1 or 2N redundancy on both sides needs piping to match: isolation on every pump, check valves where pumps run in parallel, and a path to service one pump while the other carries the load.

Data center liquid cooling leak detection

Leak detection works in layers: sensing cable under headers, CDUs and drip trays, spot sensors at the low point of each rack, CDU level and pressure alarms, and a makeup meter that reveals slow losses. All of it reports to the building management system, alongside a dew-point margin alarm. The best leak prevention is mechanical: correctly torqued and supported joints, hoses within their bend radius, and no load hanging on a quick disconnect.

Coolant choice, water quality and filtration

The data center liquid cooling coolant in the rack-side loop is either treated water, typically deionized water with a corrosion inhibitor and a biocide, or a premixed inhibited propylene glycol fluid, and the server maker’s approved fluid list decides which. The Open Compute Project’s PG25 guideline describes a fluid of 24.5 to 27.5 percent propylene glycol by volume, which it considers bio-static within that range and which freezes at about minus 10 to minus 13 degrees Celsius. ASHRAE’s limits for the two loops show why the CDU boundary matters.

ASHRAE warns that the common mistake is applying the wrong column: TCS limits are unachievable in a building loop, and facility-grade water on the rack side puts servers at risk. ASHRAE puts cold plate microchannels between about 50 and 500 microns and suggests a filter rated at one-half to one-tenth of the smallest passage, so a plate with 50-micron gaps calls for filtration between 25 and 5 microns. The Open Compute guideline adds that CDUs commonly carry 25- or 50-micron inline filters, that a side-stream filter taking roughly 5 to 15 percent of flow is strongly recommended once a loop holds more than about 250 gallons, and that over-fine filtration can strip antifoam additives from a glycol fluid. The same guideline does not recommend adding biocide to a PG25 fluid kept within its glycol range; the ASHRAE table above is a water quality guideline written for water-based coolant. Facility water quality and makeup treatment are covered on our industrial water treatment page, and filter housings and elements on our process filtration page.

Pressure testing, flushing and filling before IT connection

Site-installed TCS piping is pressure tested, flushed, filled and proven clean before any server is connected, and the servers only get a leak check at the final coupling. ASHRAE TC 9.9 recommends that site interconnecting piping follow ASME B31.3, which tests at 1.5 times design pressure, and advises against field pressure testing IT equipment. Commercial building loops built to ASME B31.9 are held at 1.5 times design pressure for at least 10 minutes.

The Open Compute guideline recommends using the same fluid through testing, cleaning and filling wherever practical, to avoid introducing a second fluid into the loop. These steps are the mechanical core of the wider commissioning program described on our data center commissioning page, and the test method itself is on our hydrostatic testing page.

Liquid cooling retrofit in an air-cooled hall

Retrofitting brings liquid to selected racks in a room designed for air, and the starting point depends on whether facility water can reach the row. Where chilled water already runs to perimeter units, a new branch with a hot tap or a planned shutdown can feed in-row liquid-to-liquid CDUs or rear doors. Where it cannot, liquid-to-air CDUs and passive rear doors let the first dense racks arrive while the piping is planned.

Feasibility turns on five checks: chilled water temperature relative to the dew point, spare pump and plant capacity, floor loading for CDUs and filled racks, a pipe route clear of cable paths with drip protection, and a tie-in that does not risk live load. We prefabricate headers and CDU connection spools off site so the work in the operating hall is short and clean; see our guide to skid and spool fabrication.

Data center liquid cooling maintenance and coolant monitoring

Maintenance in a liquid cooling data center means keeping the rack-side fluid in specification and catching small changes early. ASHRAE notes that conductivity, pH and turbidity can be measured online, while bacteria need periodic samples. The Open Compute guideline sets a minimum of one laboratory analysis per TCS loop per year, sampling more often after commissioning.

Top up only with the approved premixed fluid, and treat every server added to a live loop as a possible contamination source; ASHRAE notes that a poorly purged node can foul an established system. Our preventive maintenance plans cover the mechanical side, and our emergency repair service responds to leaks, failed pumps and fouled exchangers.

What drives data center liquid cooling cost

For the mechanical scope of a liquid cooling system for data center white space, cost follows the number of rack connections and the length of TCS distribution far more than the CDUs, which the owner or IT vendor often buys directly. Each rack adds a manifold connection, valves, hoses and a leak check. Overhead routing adds support steel and drip protection; underfloor routing needs access and clearance. Work in a live hall adds hot taps, off-hours shifts and protection of running equipment. Redundancy, N+1 or 2N CDUs with dual headers, multiplies valving. Shop prefabrication, the cleanliness standard, the initial fluid charge and oversized headers for future racks complete the list.

Facility-side upgrades, new chillers, dry coolers or plate exchangers, sit on top of this and depend on the W class chosen. Our chilled water systems and heat exchangers pages cover that plant work.

How to compare data center liquid cooling companies

Data center liquid cooling companies fall into three groups: equipment makers, IT integrators that deliver liquid-cooled racks, and the mechanical contractors who build and maintain the data center liquid cooling infrastructure linking them to the building. The question for the third group is whether one firm can route, fabricate, install, test, flush, fill and maintain the piping. Ask for a written cleanliness procedure, the wetted-materials list they will build to, sample test and flush records, and their method for tie-ins to a live plant. Engineering support for layout and hydraulics is available through our process engineering and design team, and the general contractor’s view of the trade is on our data center mechanical contractors page.

What Paul Industries handles on a liquid cooling job

On a liquid cooling project Paul Industries routes, fabricates and installs the facility water branches to CDUs and rear doors, the TCS row headers and connections to rack manifolds, the valves, strainers, filters, vents, drains and expansion components around them, and the supports and drip protection. We set CDUs supplied by others, pressure test, flush, fill and vent the loops, support startup and balancing, and maintain and repair the system afterward. We do not manufacture CDUs, cold plates, rear doors, immersion tanks or servers, and we do not supply coolant or treatment chemicals as a vendor program; electrical work is by others. Request a quote for a new liquid-cooled hall or a retrofit.

Standards referenced: ASHRAE TC 9.9 Thermal Guidelines for Data Processing Environments · ASHRAE Liquid Cooling Guidelines for Datacom Equipment Centers · OCP PG25 Guidelines · OCP Universal Quick Disconnect specification · ASME B31.3 · ASME B31.9 · ISO 16889

Frequently asked questions

Are AI data centers liquid cooled?

The racks built around current accelerators are, because their heat output exceeds what air can practically remove. Most AI halls still run some air cooling for networking, storage and the parts of each server without cold plates, so they are hybrid rather than all-liquid.

What is liquid cooling in data center design, as opposed to a liquid-cooled PC?

A PC loop is a sealed kit with one pump and radiator. In a data center the liquid circuit is building infrastructure: CDUs, piping across the hall, manifolds in every rack and a water connection to the plant, with redundancy, leak detection and fluid chemistry managed for years.

Is direct to chip cooling closed loop?

Yes. The rack-side circuit is a closed, pressurised loop between the CDU and the cold plates, and it never mixes with facility or cooling tower water. The heat crosses the CDU's plate exchanger, which is the only thing the two fluids share.

Does direct to chip cooling use water?

Usually, as deionized water with inhibitors and biocide or as a water and propylene glycol mixture. Two-phase systems are the exception: they use a low-boiling dielectric fluid inside the cold plates, though the building side still uses water.

What is cold plate cooling?

It is the use of a metal plate with internal coolant channels clamped to a hot component, so heat moves by conduction into the liquid. In data centers the plates sit on CPUs and GPUs and are connected through hoses to the rack manifold.

Cold plate vs immersion cooling: which asks more of the building?

Immersion asks more of the structure and layout, because tanks are heavy and horizontal, while cold plates ask more of the piping, because every rack needs a manifold, couplings and fine filtration. Both need a clean facility water connection.

Coolant distribution unit vs CRAC: what is the difference?

A CRAC is a room air conditioner with its own compressor that cools air. A CDU cools no air at all in its liquid-to-liquid form: it exchanges heat between facility water and the rack-side liquid and pumps that liquid to the servers.

Can rear door heat exchangers be installed on existing racks?

Often yes. Doors are made to fit common cabinet sizes, and a retrofit needs a water supply above the dew point, hose routes to each cabinet and clearance for the deeper door. Check the cabinet frame and hinge side before ordering.

How much coolant flow does a liquid-cooled rack need?

It scales with heat and temperature rise. With water at a 10 degree Celsius rise, a 100 kilowatt rack needs roughly 140 liters per minute, about 38 gallons per minute; a smaller rise or a glycol mix needs more. The server maker's flow requirement governs.

What filter rating should a technology cooling system use?

ASHRAE suggests filtering to between one-half and one-tenth of the smallest passage in the loop, and cold plate passages can be as small as about 50 microns. The cold plate, quick disconnect and CDU makers should agree the rating before installation.

Who owns water quality in the rack-side loop?

Someone has to be named. ASHRAE notes that ownership of the TCS often falls between facilities staff and the IT group, and recommends a single responsible owner for adding equipment, servicing and sampling.

Can a PG25 loop be topped up with plain water?

It should not be. Water dilutes both glycol and inhibitor, and below about 24.5 percent glycol the fluid loses the bio-static behavior PG25 relies on. Use the same premixed fluid, from a batch with a retained sample.

Do quick disconnects from different manufacturers fit together?

Only if they are built to a shared interface. The Open Compute Project's Universal Quick Disconnect specification defines one so compliant couplings mate across makers, but the server maker's required coupling still decides what goes on the manifold.

Does data center liquid cooling still need chillers?

It depends on the ASHRAE W class. Equipment rated for W45 can often run on dry coolers or towers alone, while W17 and W27 sites normally keep chillers. Chiller capacity in reserve protects against future hardware that wants cooler water.

What pressure is a TCS loop tested to?

Site piping is commonly hydrotested at 1.5 times design pressure following ASME B31.3, before the IT equipment is attached. The servers are not field pressure tested; their connection is leak checked instead.

Why are unused future branches a problem in liquid cooling piping?

Water standing in a capped branch for months grows scale, corrosion product and biofilm, which flushes into the first server connected. Keep future taps short, valved at the header and flushable.

How is the ASHRAE W class different from the A class?

A classes describe the air temperature and humidity an air-cooled server can take at its inlet. W classes describe the facility water supply temperature a liquid-cooled server can accept. A liquid-cooled hall usually has to meet one of each.

What is a secondary fluid network?

It is another name for the rack-side loop, the TCS: the piping, manifolds and hoses between the CDU's secondary side and the cold plates.

Related: Data Centers · Data Center Cooling Systems · Chilled Water Systems · Process Chiller Installation & Service · Sanitary Heat Exchanger Installation & Service · Plate Heat Exchanger · Request a quote