What Is a CDU in a Data Center? How to Size It for AI Racks

A coolant distribution unit (CDU) is the controlled interface between a data center's facility water system and the technology cooling system that serves liquid-cooled IT. It transfers heat, circulates coolant, controls temperature and pressure, filters the technology loop, and reports operating or fault conditions to facility controls.

A CDU should not be selected from a rack-kilowatt total alone. The unit has to meet the required heat-transfer duty at the specified supply temperatures, flow and available pump head, with the correct coolant, materials, redundancy and control behavior. Those limits must hold in the governing operating and maintenance cases.

This guide is for facility engineers, data center designers, operators and technical buyers converting an AI rack schedule into a CDU specification and commissioning plan.

TL;DR: CDU sizing for AI racks

• Build a rack-by-rack liquid heat schedule; rack input power is not automatically the CDU heat load.
• Calculate flow from heat duty, coolant properties and the permitted temperature rise, then verify the vendor's rating point.
• Size pumps from the complete hydraulic circuit, including remote racks, hoses, manifolds, valves and dirty filters.
• Check thermal capacity, pump head, temperature approach, coolant compatibility, controls and physical access as separate limits.
• Test complete-unit and common-system failures, then commission the connected loop rather than only the CDU cabinet.

Where the CDU sits in the liquid-cooling system

Liquid-cooled IT is commonly divided into two loops:

• Facility water system (FWS): the site side that rejects heat through chillers, dry coolers, cooling towers or another heat-rejection system.
• Technology cooling system (TCS): the controlled loop that serves cold plates, rack manifolds and other IT cooling components.

A liquid-to-liquid CDU normally separates the loops with a heat exchanger. Pumps circulate the TCS coolant while controls maintain supply temperature, differential pressure or another commanded condition. Filters, expansion volume, air management, instrumentation and leak-response interfaces support stable operation.

That separation limits direct exposure of IT components to facility-water pressure and chemistry, but it does not eliminate coordination. The CDU can transfer only the heat that the facility side can accept under the actual supply temperature and flow. The liquid cooling data center guide explains the broader architecture, while Does Liquid Cooling Use Water? separates the closed loops from outdoor heat rejection. This guide focuses on CDU sizing and acceptance.

Rack-level, in-row or facility-level CDU?

CDU placement changes the pipework, fault domain and service model. A rack-level CDU serves one rack or a small rack group, shortens the controlled technology-cooling circuit and makes rack-by-rack isolation straightforward, but it also creates more units to power, monitor and maintain. An in-row CDU shares capacity across several racks. A larger unit outside the data hall can reduce equipment count, although it requires longer secondary headers and can place more racks behind one cooling fault domain.

The right answer follows the selected rack system, maintenance strategy and phasing plan. Supermicro, for example, lists in-rack, in-row and liquid-to-air sidecar CDU options for its GB300 NVL72 system. Nidec lists a 4U, 250 kW in-rack CDU intended to cool one GB300 NVL72 rack. Those are vendor-specific examples, not interchangeable ratings.

A CDU has five independent limits

A headline thermal rating is useful only when its rating conditions match the project. Check these five limits together:

A unit can pass one limit and fail another. A heat exchanger may have enough surface area while the pumps cannot overcome the rack circuit. A pump may deliver the flow while the facility water is too warm to achieve the required TCS supply temperature. Selection is an operating-point decision, not a catalogue maximum.

Step 1: build the rack heat-load schedule

Start with the IT vendor's design data for each rack type and configuration. Record:

• maximum or design input power, not only average utilization;
• the fraction of heat rejected to liquid and the remaining room-side air load;
• required TCS supply-temperature range;
• allowable pressure, differential pressure and flow range;
• coolant chemistry, cleanliness and materials restrictions;
• rack manifold and connection details; and
• expected configuration changes over the design period.

Do not turn one named rack's thermal design power into a universal AI-rack value. Published systems already differ. NVIDIA's DGX GB200 hardware documentation, HPE's GB300 NVL72 integrated-rack QuickSpecs and other OEM implementations have their own power, temperature, flow and coolant requirements. Use the exact configuration purchased and its current technical release.

Step 2: convert liquid heat into coolant flow

For a single-phase liquid loop, the basic heat balance is:

Heat transferred (kW) = mass flow (kg/s) × specific heat (kJ/kg·K) × coolant temperature rise (K)

Rearranged:

Mass flow = heat transferred / (specific heat × temperature rise)

The following example uses water properties rounded for preliminary planning. It is deliberately generic, not a specification for a named rack or CDU.

The 90% capture and 10 K rise are assumptions. Replace them with the OEM's rack data. Glycol mixtures and other coolants have different specific heat, density and viscosity, so both flow and pump power change. Increasing the permitted temperature rise reduces required flow, but only if the rack, cold plates and IT vendor allow that operating point.

The residual 96 kW is also a real design load. It has to be removed from the room through air cooling or another specified path. “Liquid cooled” does not mean “zero air heat.”

Step 3: check the temperature approach

A heat exchanger needs a temperature difference to transfer heat. A preliminary counterflow relationship is:

Required FWS supply temperature ≈ required TCS supply temperature − CDU approach temperature

If the rack requires 32°C TCS supply and the CDU needs a 4 K approach at the selected duty, facility water must reach approximately 28°C. If the approach is 8 K at that same duty, the facility supply requirement becomes approximately 24°C.

That four-degree difference can change dry-cooler selection, water use, chiller hours and site energy. It is also why two thermal ratings printed on different CDU data sheets are not comparable until the test conditions are aligned.

For example, the published Vertiv Liebert XDU450 data sheet lists different nominal capacities at 4°C and 8°C approach conditions. The correct conclusion is not that one number is “the capacity.” It is that heat-exchanger duty varies with the operating temperatures and flow.

Use a performance map that covers design-day FWS temperature, minimum and maximum flow, part load, fouling allowance and the required TCS setpoint. If the plant resets water temperature seasonally, check the full envelope rather than one favorable point.

The facility-side result depends on climate and heat-rejection architecture. The data center cooling in extreme heat guide covers those design-day conditions. For the separate question of withdrawal and consumption, see how AI data centers use water.

Step 4: solve the hydraulic circuit

Thermal flow becomes a pump duty only after the system resistance is known. Build the pressure-drop schedule from:

• CDU internal exchanger, valves and filters;
• supply and return pipe length, diameter, fittings and elevation effects where relevant;
• row and rack manifolds;
• flexible hoses and quick disconnects;
• cold plates or immersion heat exchangers;
• balancing and control valves; and
• the dirty-filter condition at the replacement threshold.

Plot the system curve against the vendor pump curve for the operating configurations. Check the most remote rack, minimum-flow stability and what happens as additional racks open. Variable-speed control can regulate differential pressure, but it cannot create head beyond the pump curve.

Connection hardware must also match the service. The Open Compute Project's advanced-cooling connection guidance gives recommendations for fluid connections, including pressure and coolant compatibility. It is guidance, not a substitute for the chosen IT and CDU vendors' requirements or the project's pressure design.

Step 5: select coolant and materials as one system

Do not specify “water/glycol” as if every concentration and inhibitor package were interchangeable. Coolant choice affects heat capacity, viscosity, corrosion control, elastomer compatibility, electrical conductivity and maintenance.

Water, aqueous propylene glycol and aqueous ethylene glycol can all appear in liquid-cooling systems, but the fluid has to be approved for the particular loop. Glycol may be required where exposed pipework or heat-rejection equipment can see sub-zero conditions. It lowers the freezing point, while also changing heat-transfer properties and pump duty. Propylene glycol is less toxic than ethylene glycol; neither should be substituted into an IT-side loop without the rack and CDU vendors' approval.

OEM requirements can differ. Lenovo's Neptune water-quality guidance describes its supported chemistry and facility-water conditions. Dell's liquid-coolant guidance describes requirements for its own technology and facility loops. A mixed-vendor hall needs a project chemistry envelope accepted by every affected supplier.

The basis of design should state:

• fluid type and concentration;
• water source and treatment process;
• pH, conductivity, hardness and inhibitor limits where required;
• wetted materials and prohibited material combinations;
• sampling points, test method and test interval;
• filtration rating and change criterion; and
• fill, flush, storage and disposal procedures.

Approve the fluid before filling the system. Correcting incompatible chemistry after it has circulated through cold plates and hoses is an operating incident, not a paperwork change.

Redundancy has to cover capacity and common modes

“Dual pumps” can mean two pumps that share one controller, one electrical feed, one heat exchanger, one filter and one CDU shell. That may provide pump redundancy while leaving several single points of failure.

Test the arrangement against named failures:

A simple unit-count example shows why wording matters. Six 2.5 MW CDUs arranged as 4 duty + 2 standby have 15 MW installed. After one complete unit is unavailable, five units remain, equal to 12.5 MW of nameplate capacity. Whether that is sufficient depends on the required design load and whether common FWS, power, header or control failures have also been addressed.

Plan for condensation and leak response

Warm-water liquid cooling can reduce condensation risk, but the control system still needs a measured dew-point margin. Supply coolant temperature should remain above the applicable room or enclosure dew point under the stated transient and sensor-failure cases. The margin and response should come from the IT and cooling-system design, not a generic fixed value.

Leak detection also needs a control narrative. Detection without an agreed action only creates an alarm. Define sensor locations, zone naming, notification, automatic valve or pump actions, rack isolation, drainage, clean-up and return-to-service criteria. The Open Compute Project's cold-plate leak detection and intervention paper is a useful industry reference for designing that response.

Specify the CDU at project conditions

A procurement schedule should require a selection sheet, not a single kW field. Record:

• design and future liquid heat loads by rack group;
• TCS supply and return temperatures at each design point;
• FWS supply and return temperatures at design-day and seasonal conditions;
• coolant type, concentration and property basis;
• required TCS and FWS flow;
• available pump head at required flow and dirty-filter condition;
• maximum allowable pressure and pressure-control method;
• approach-temperature guarantee or performance map;
• pump, controller, electrical-source and complete-unit redundancy;
• heat-exchanger, pipe, seal and hose materials;
• filtration, air removal, expansion volume and water-quality instrumentation;
• communications protocol, points list, alarms and trend data;
• sound, dimensions, weight, service clearance and drainage; and
• factory, site and integrated test requirements.

ASHRAE's published Addendum b to Standard 127-2020 addresses liquid-to-liquid heat-exchanger rating conditions and terminology. Use the applicable edition and project specification with vendor performance data; do not compare unlike rating points.

Commission the loop, not only the CDU cabinet

Factory testing can verify the assembled CDU's internal wiring, controls, instruments and simulated sequences. Site commissioning has to prove the connected hydraulic and control system.

1. Inspect materials, cleanliness, supports, flexible connections and service access.
2. Confirm flushing and water-quality records before connecting sensitive IT equipment.
3. Document the pressure-test basis, limits, calibrated instruments and acceptance criteria agreed by the responsible engineers and vendors.
4. Verify sensor calibration, valve stroke, pump rotation, filter differential-pressure measurement and alarm naming.
5. Balance flow through headers and the hydraulically remote rack path.
6. Demonstrate stable control at minimum, partial and design flow.
7. Simulate one pump, one controller, one sensor and one electrical-source failure where the design claims tolerance.
8. Test loss of FWS flow, high temperature, high and low pressure, dirty filter and leak alarms.
9. Trend TCS and FWS supply/return temperatures, flow, pressure and pump speed during a representative heat load.
10. Record as-left setpoints, fluid sample results, valve positions, firmware, alarm delays and recovery procedure.

Do not prescribe a generic pressure-test multiplier without the system's pressure class, component limits and governing standard. The approved procedure has to respect the lowest-rated connected component and the IT vendor's isolation requirements.

Where ModulEdge fits

ModulEdge treats cooling, electrical power and IT space as coordinated interfaces in its factory-built modular data center platforms. For liquid-cooled projects, the rack schedule, heat-to-liquid fraction, TCS requirements and site FWS conditions must be defined before the module and cooling-plant interfaces are finalized.

For current GB300 NVL72 design work, ModulEdge generally prefers a rack-level CDU where the chosen OEM system supports it. The shorter secondary circuit and rack-level isolation can simplify phasing and preserve space otherwise assigned to a shared CDU. That is a design preference, not a universal rule: an in-row or external CDU can be the better choice when the rack vendor, capacity plan, service model or fault-domain analysis points that way.

The exact CDU supply, controls, coolant, field piping and commissioning boundary remains project-specific. This article does not claim that every ModulEdge configuration includes a CDU or that one standard CDU selection fits every AI platform.

For an initial cooling review, send ModulEdge the OEM rack data sheets, quantity by phase, liquid heat load, TCS flow and temperature requirements, coolant specification, site design conditions and proposed FWS temperatures. Those inputs allow the cooling interfaces to be assessed without turning a preliminary rack power figure into a facility promise.

Frequently asked questions

How do you size a CDU for AI racks?

Sum the design liquid heat load for the connected racks, including the planned phase and stated margin. Use the IT vendor's required coolant temperatures and allowable pressure range. Calculate coolant flow, solve the pressure losses through the complete circuit, and verify the heat exchanger and pump curves at those project conditions.

Is CDU capacity the same as total rack power?

Not necessarily. Some rack power leaves through the liquid loop and the remainder enters the room air. The CDU serves only the heat assigned to its technology cooling system, plus any relevant pump heat or design allowance. Use the hardware vendor's heat-to-liquid fraction rather than assuming that every electrical kilowatt reaches the CDU.

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

Flow depends on liquid heat load, coolant heat capacity, density and the permitted supply-to-return temperature rise. For water near typical operating conditions, a larger temperature rise reduces required flow for the same heat load. The final value must still satisfy the OEM's flow and temperature limits and the hydraulic balance across all connected racks.

What is the difference between a liquid-to-liquid and liquid-to-air CDU?

A liquid-to-liquid CDU transfers heat from the technology loop into a facility water loop. A liquid-to-air CDU rejects heat into the room air, so the room cooling system must carry that load outdoors. The choice changes facility interfaces, achievable temperatures, pump duties and where the heat-rejection capacity must be provided.

Does N+1 CDU redundancy guarantee cooling availability?

No. N+1 unit count covers only the stated capacity failure if the common headers, facility-water source, electrical supply, controller, network and isolation arrangement also support the required state. Define complete-unit and common-mode failures, the permitted load after each event and the recovery sequence.

What data should be included in a CDU specification?

Include liquid heat by phase, TCS and FWS temperatures and flows, coolant chemistry, pump head, pressure limits, approach temperature, materials, filtration, air removal, expansion volume, redundancy, power feeds, controls, alarms, communications, dimensions, access, drainage and the factory, site and integrated-test requirements.

Should a CDU be installed in the rack, in the row or outside the data hall?

Choose the location from the rack vendor's supported architecture, capacity per fault domain, pipe length, phasing, service access and maintenance strategy. Rack-level units shorten the secondary circuit and isolate individual racks; in-row and facility-level units consolidate equipment but serve larger rack groups. Compare the complete operating and failure cases rather than footprint alone.