March 15, 2026
Liquid Cooling for Data Centers: System Design and Selection Guide
Compare direct-to-chip, rear-door, and immersion cooling. Learn rack interfaces, CDUs, facility loops, water, commissioning, and procurement requirements.
Updated September 19, 2026 to reflect current liquid-cooling terminology, rack and facility interfaces, and a clearer design and procurement process.
Liquid cooling for a data center is a heat-transfer system that connects the IT equipment to the facility and then to the outdoor environment. The label can describe rear-door heat exchangers, direct-to-chip cold plates, immersion systems, or hybrids. These designs do not have the same server interfaces, liquid loops, residual-air loads, service procedures, or heat-rejection requirements.
The right choice starts with the selected IT equipment and workload. Rack density is useful, but no single kW-per-rack threshold decides the architecture. A buyer needs the rack schedule, the heat split between liquid and air, the allowable temperatures and pressures, the site climate, the operating model, and the required behavior during maintenance and failures.
This guide follows the heat from the chips to the outdoor plant, compares the main approaches, and sets out the information a supplier should return before a liquid-cooling design is accepted.
What does “liquid-cooled data center” mean?
Water or another fluid can appear at several points in the cooling path. That does not make every system direct liquid cooling.
| Architecture | Where liquid removes heat | What still needs attention |
|---|---|---|
| Air cooling | Room or row cooling equipment transfers heat from air to a refrigerant or water loop | Airflow, containment, fan energy, room temperature, humidity, and outdoor heat rejection |
| Rear-door heat exchanger | A water-cooled door captures hot server exhaust at the rack | The servers remain air cooled; door airflow, condensation control, hoses, and any remaining room load must be designed |
| Direct-to-chip cold plates | A technology-cooling loop removes heat from selected processors or other components | Uncooled components still reject heat to air; CDUs, manifolds, connections, coolant quality, and leak management become part of the system |
| Immersion cooling | IT equipment is immersed in a dielectric fluid, in a single-phase or two-phase system | Equipment compatibility, fluid handling, tank loading, service procedures, warranties, and outdoor heat rejection require a purpose-built operating model |
A hybrid design may combine several of these methods. For example, cold plates can remove most processor heat while memory, power supplies, storage, switches, and other components remain air cooled. The facility must be sized for both paths.
Start with the IT equipment, not a density rule
Claims such as “air cooling stops at 30 kW per rack” or “direct liquid cooling is mandatory above 50 kW” are too broad for design. Server airflow, inlet-temperature limits, permitted temperature rise, rack geometry, fan capability, cold-plate coverage, and the room layout all matter. The equipment manufacturer's installation and environmental specifications are the controlling inputs.
Build a rack schedule that records:
- server, network, storage, and management equipment by rack;
- nominal and peak IT power, rather than accelerator power alone;
- required air flow and the heat rejected to air;
- required coolant, supply temperature, return temperature or temperature rise, flow, pressure, and water-quality limits;
- rack dimensions, weight, manifolds, hose routes, connection type, and service clearances;
- normal, maintenance, start-up, shutdown, and fault behavior.
ASHRAE's liquid-cooling classes use the number in W17, W27, W32, W40, and W45 as the maximum facility-water supply temperature for that class. They do not describe chip temperature, and a class name is not a substitute for the complete OEM envelope. ASHRAE's liquid-cooling overview also makes clear that facilities and IT manufacturers need a common interface definition.
Follow the complete heat path
A workable design accounts for every transfer step:
- Equipment to coolant or air. Cold plates, immersion fluid, or server fans collect heat from the IT equipment.
- Rack distribution. Manifolds, hoses, quick disconnects, or rear doors move heat out of the rack while preserving access and containing leaks.
- Technology cooling system. A CDU can control temperature, pressure, flow, filtration, and water quality on the IT side.
- Facility water system. A heat exchanger transfers heat to pumps, headers, and the facility loop without assuming that the two fluids or pressure regimes are compatible.
- Outdoor heat rejection. Dry coolers, adiabatic systems, cooling towers, chillers, or hybrids reject the heat under the site's design conditions.
The Open Compute Project's cold-plate liquid-cooling requirements identify material compatibility, coolant properties, cleanliness, leak detection, flow, pressure, temperature, and CDU capacity as connected design variables. A CDU is not simply a pump skid; it establishes and controls the boundary between the technology and facility loops.
The basic heat relationship is:
Heat rate = mass flow × specific heat × temperature rise
A larger temperature rise can reduce the required flow for the same heat load, but the permitted temperatures and approaches remain limited by the IT equipment and heat exchangers. Warmer water can improve opportunities for compressor-free heat rejection and heat reuse, but only if the selected equipment supports it and the site conditions make it useful.
Rear-door heat exchangers
Rear-door heat exchangers can extend the useful range of air-cooled servers by intercepting hot rack exhaust. Active doors use fans to overcome pressure drop and control airflow; passive doors rely more heavily on the server fans and the room design. Both need a defined chilled-water circuit, condensate strategy where applicable, hose routing, door-swing and service access, controls, leak detection, and remaining room-cooling capacity.
This approach can be attractive when the operator wants to retain standard air-cooled IT, but it should not be described as direct-to-chip cooling. The server still moves heat through air before it reaches the water circuit.
Direct-to-chip liquid cooling
Cold plates place a liquid path close to the highest-heat components. This can reduce server fan demand and transfer a large share of rack heat to liquid, but it rarely removes the air system entirely. The buyer should request the exact liquid-capture percentage or, better, the kW assigned to the liquid and residual-air paths for each rack configuration.
Specify the technology loop and facility loop separately. The design should state:
- fluid type and approved material set;
- supply and return temperatures, allowable ramp rates, dew-point control, and heat-exchanger approach;
- minimum, nominal, and maximum flow and differential pressure;
- filtration, chemistry, corrosion control, sampling, filling, flushing, and maintenance;
- CDU and pump topology in normal, maintenance, and failure states;
- leak detection, isolation zones, drip containment, and alarm response;
- the residual-air load and how it is cooled during every operating state.
Quick disconnects and manifolds must be evaluated for pressure drop, drip performance, cycle life, serviceability, and compatibility with the OEM warranty. The most efficient hydraulic design is not useful if routine server replacement becomes unsafe or excessively slow.
Our direct-to-chip cooling requirements guide follows the cold plates, coolant connections, residual-air duty and service procedures that must be agreed between server, rack and facility suppliers.
Immersion cooling
Immersion removes the conventional rack-air interface, but it changes the equipment and operating model. Servers or boards must be approved for the fluid and tank arrangement. Cables, connectors, storage devices, labels, seals, and component materials need compatibility review. The design must also resolve fluid inventory, fire strategy, floor loading, lifting, draining, filtration, contamination control, service tools, spare-part handling, and end-of-life fluid management.
Compare immersion with cold plates on the actual workload, hardware availability, service process, supplier ecosystem, and lifecycle plan. Neither is automatically the better liquid-cooling system.
For equipment compatibility, service procedures and facility qualification, read our immersion cooling guide. For selecting between the two cycles, use the single-phase versus two-phase comparison.
Does liquid cooling reduce energy and water use?
It can reduce fan energy and allow higher coolant temperatures, which may increase free-cooling hours or improve heat-reuse potential. It does not guarantee a particular PUE, annual energy saving, or water result. Pumps, CDUs, residual-air systems, and outdoor heat rejection still consume energy. An adiabatic or evaporative system may consume water even when the IT loop is closed.
Compare options using the same IT workload, climate file, load profile, availability state, and measurement boundary. Report annual facility energy alongside PUE, direct water consumption alongside WUE, and the operating conditions behind both. The PUE guide and AI water-use guide explain those boundaries in more detail.
Reliability and serviceability
Liquid cooling replaces some air-side constraints with hydraulic ones. Reliability depends on the whole system rather than on the absence of water near electronics. Review common-mode failures, power to pumps and controls, CDU and heat-rejection redundancy, valve positions, sensor failure, trapped air, fouling, loss of facility water, condensation risk, and the response to a leak.
Maintenance plans should cover coolant sampling, filter changes, pump and valve service, connection inspection, firmware and controls, spare CDUs or components, and the isolation method for one rack or branch. State whether IT can continue during each maintenance activity and at what load.
Commissioning and acceptance
Factory and site testing should prove the interfaces, controls, and failure responses, not just the nameplate capacity. A project-specific test plan should include:
- cleaning, flushing, pressure testing, water-quality records, and material certificates;
- sensor calibration and verification of flow, pressure, temperature, and heat balance;
- leak detection, isolation, alarm routing, and safe shutdown sequences;
- normal, partial-load, full-load, maintenance, and selected failure states;
- CDU, pump, residual-air, and outdoor heat-rejection control sequences;
- thermal stabilization and a representative IT or load-bank test;
- handover of setpoints, trend data, water-treatment records, spares, and maintenance procedures.
Agree the pass criteria and test boundary before fabrication. A supplier test that stops at the module connection does not demonstrate the site heat-rejection system, and a site test without the intended rack heat split may not demonstrate production operation.
How ModulEdge applies the cooling choice
ModulEdge offers factory-built modular data-center platforms from 20 kW to 2.3+ MW per deployment unit, scalable to multi-megawatt capacity. Across the portfolio, project configurations support 5–150 kW per rack. The selected IT equipment, rack schedule, site, and project availability objective determine the final cooling architecture.
The 500 kW MDC-500 provides ten 42U rack positions at up to 50 kW each. Its defined architecture keeps the IT equipment air cooled and uses active rear-door heat exchangers, a closed chilled-water loop, and modular air-cooled chillers with free-cooling capability when conditions allow. It is positioned for dense air-cooled H200-class and RTX PRO 6000 Blackwell Server Edition systems, subject to the selected server and project engineering.
The MDC-2000 is a prefabricated modular AI building designed for rack-scale liquid-cooled infrastructure. Its current public GB300 layout uses 16 compute racks plus 16 network and service racks and combines warm-water direct liquid cooling with residual-air cooling. It can also be configured for AMD Helios-class infrastructure; rack count, load, layout, and interfaces follow the selected platform. The issued product reference and the current public layout have different rack allocations, so the detailed project load schedule should accompany a proposal.
Define the cooling system from the rack outward
Send ModulEdge the equipment list, rack loads, liquid and air interfaces, site climate, and availability objective. We will map the heat path, facility interfaces, external plant, and test boundary for the relevant modular platform.
Liquid cooling procurement checklist
- Exact rack schedule, operating cases, and liquid-versus-air heat split
- OEM environmental, hydraulic, chemistry, connection, and warranty requirements
- Technology loop, CDU, facility loop, and outdoor-plant boundaries
- Temperatures, flow, pressure, approach, filtration, and material compatibility
- Residual-air cooling in normal, maintenance, and failure states
- Leak prevention, detection, containment, isolation, and response
- Annual energy and water model with climate and load assumptions
- Heat-reuse temperature, demand, interface, and metering where relevant
- FAT, SAT, integrated test procedures, pass criteria, and handover records
- Operations, maintenance, spares, training, and responsibility matrix
A useful proposal makes every interface visible. It should show where the supplier's responsibility begins and ends, what the system will do under stated conditions, and how the buyer will verify it.
