Liquid cooling: the definition
Liquid cooling uses a circulating liquid to transport heat from equipment toward a heat exchanger or rejection system. In a data center, chip-level heat capture, coolant circulation and final heat rejection are separate engineering stages.
The key points
- Cooling transports heat; it does not make the heat disappear.
- Direct-to-chip, rear-door and immersion approaches put liquids at different boundaries.
- A CDU can separate technology and facility circuits while transferring heat between them.
- Water use and energy use depend on the entire heat-rejection system and operating conditions.
Begin with the heat path
Computing equipment consumes electricity and produces heat that must be removed to remain within supported operating conditions. Liquid cooling changes how that heat is collected and transported. The useful picture is a chain: heat leaves components, enters a coolant path, crosses an exchanger where needed and is ultimately rejected or reused. Focusing only on the cold plate omits much of the facility system. [1]
Imagine a hypothetical rack producing 100 kW of heat. Moving that heat into water does not reduce it to zero. Another system must accept approximately that heat flow, together with relevant added loads, and keep the coolant within its operating conditions. A cold plate without a usable downstream path is not a complete cooling solution.
Cold plates, rear doors and immersion
Direct-to-chip systems use interfaces such as cold plates to capture heat close to selected components. Rear-door heat exchangers can transfer heat from server exhaust air into a liquid circuit. Immersion places compatible equipment in an appropriate electrically nonconductive fluid. These approaches differ in equipment compatibility, service procedures and the fraction of heat captured through the liquid path. [1]
A liquid-cooled server is not necessarily entirely liquid-cooled. Some components can continue rejecting heat into room air, creating a hybrid requirement. The cooling description should identify which loads enter which path. Assuming that liquid-cooled GPUs eliminate every fan or room-cooling obligation can leave important supporting loads unaccounted for. [1]
For an illustrative allocation, suppose 80 kW from a 100 kW rack enters cold plates while 20 kW remains in air. The facility must support both streams. Advertising 100 kW of liquid capacity elsewhere does not establish that the residual air load is handled correctly. Match the equipment's actual heat split to the room and liquid-system design.
The role of a coolant distribution unit
A coolant distribution unit can circulate and regulate coolant and transfer heat between technology-side and facility-side circuits. A liquid-to-liquid design uses a heat exchanger to separate the fluids while passing heat. Open Compute Project's Deschutes-based CDU descriptions illustrate this separation together with pumps, monitoring and filtration. Other designs have different boundaries, so the acronym does not define one universal configuration. [2]
Separating circuits allows them to have different fluid-quality, pressure and operating requirements. It does not remove the need for compatible materials or correct maintenance. The manufacturer specifies acceptable coolant, temperatures, pressure drops and interfaces for the equipment. A generic description of water cooling is not sufficient to select a fluid or commission a server installation. [1][2]
Think of a hypothetical exchanger as a boundary that passes thermal energy but not the intended bulk flow of the two fluids. A warm technology return gives up heat to a suitable facility circuit; the technology supply returns cooler. The achievable temperatures depend on exchanger performance and flow conditions. The facility cannot provide an arbitrary supply temperature independently of its own heat-rejection conditions.
A simple heat-balance calculation
For a single-phase illustrative fluid path, heat transport can be approximated by mass flow multiplied by specific heat capacity and temperature rise. This is an energy-balance relationship, not a complete design method. Pressure drop, pump efficiency, fluid properties, local component limits and transient behavior still need to be considered by qualified designers.
Assume hypothetical water with an approximate specific heat of 4.18 kJ per kilogram-kelvin and a 10 K temperature rise while carrying 100 kW. The required idealized mass flow is 100 divided by 4.18 divided by 10, about 2.39 kg/s. With density approximated as 1 kg/L, that is about 143 L/min. These assumptions are for teaching; a glycol mixture, different temperature or real hydraulic system changes the result.
At the same hypothetical heat load, halving the allowed temperature rise doubles the idealized flow requirement. This does not mean increasing pump speed indefinitely is an acceptable solution. Equipment pressure limits and the resistance of the actual path constrain operation. The equation helps explain why temperature, flow and heat load are linked, not how to override a manufacturer's limits.
Where the heat goes next
A facility can reject heat through different combinations of dry coolers, evaporative equipment, chillers and heat-reuse systems. Outside conditions and permitted coolant temperatures affect which modes are available. A national-laboratory data-center account illustrates how warm-water cooling, heat recovery and hybrid heat rejection interact; it is a specific operating example, not a guarantee for every geography. [3]
A closed technology loop describes circulation within one boundary. The facility loop can still use an evaporative cooling tower, and the broader site can consume water elsewhere. Conversely, a suitable dry-rejection design can reduce on-site evaporative water use while having other energy or climate constraints. Always identify which boundary a zero-water claim addresses. [4][3]
Suppose a hypothetical server loop never discharges coolant in normal operation, but its facility heat is rejected through evaporating water. Calling the server loop closed is reasonable; calling the whole facility water-free is not established. A claim about one component must not silently become a claim about the full system.
Warm water and useful heat
Heat can be reused when there is a suitable recipient and temperature relationship. The laboratory example describes using recovered computing heat for building needs. The benefit depends on the surrounding system, seasonal demand and the energy that would otherwise be used. Capturing heat is not the same as having a valuable destination for it all year. [3]
Imagine a hypothetical campus with a steady data-center heat output but little heating demand in summer. A large annual heat-generation figure does not establish an equal amount of useful recovered heat. The analysis must match supply and demand over time. PUE also does not automatically capture that benefit; heat-reuse and water metrics have separate definitions.
Cooling is an operating responsibility
Liquid systems require attention to fluid compatibility, cleanliness, leak detection, connections, pressure and maintenance. Condensation is another concern when surfaces operate below the relevant dew point. ASHRAE's discussion treats interfaces and operating limits as part of the design. These are reasons for documented procedures and qualified commissioning, not instructions to improvise modifications to energized equipment. [1]
A useful operational question is what happens when a pump, power feed or control component fails. Redundant parts do not prove the whole path remains available, and cooling continuity may differ from server-power continuity. Acceptance testing should examine the supported failure cases and recovery procedures, using the actual installed configuration rather than a generic marketing diagram.
Reading a liquid-cooling claim
Ask which components are cooled, the supported rack load, fluid and temperature requirements, residual air load, facility interface and final rejection method. Distinguish design capability from commissioned availability. Measure water and energy at declared boundaries, and retain climate and load conditions alongside the result.
The central lesson is a continuous heat path. Liquid cooling can make high-density systems practical and can support efficient heat reuse, but neither outcome follows from a single label. Chip interfaces, loops, exchangers, heat rejection and operations must work together. Understanding the full path turns a fashionable term into a concrete infrastructure requirement.
Check your understanding
Try answering before opening the explanation. Your answers are not collected or scored.
1. Does a closed server loop establish zero facility water use?
No. Other loops or heat-rejection equipment can still consume water. The claim's measurement boundary must be specified.
2. Why can a liquid-cooled rack still need air cooling?
Some components may reject heat into air rather than the liquid path. The residual load depends on the equipment design.
3. What happens to idealized flow if the permitted temperature rise halves at the same heat load?
The heat-balance equation requires twice the mass flow, before real hydraulic, fluid-property and equipment constraints are considered.
Sources & editorial note
Reference documentation is listed below with its recorded check date. Technical statements are attributed; passages framed as our view or recommendation are editorial interpretation. Examples are hypothetical unless explicitly identified otherwise. No independent Kovara hardware testing is claimed.
- ASHRAE · Data-center cooling systems ↗ (opens in a new tab)Engineering handbook · Checked 28 September 2026
- Open Compute Project · Deschutes CDU architecture example ↗ (opens in a new tab)Open hardware product record · Checked 28 September 2026
- National laboratory · A decade of greener computing ↗ (opens in a new tab)Operator's technical account · Checked 28 September 2026
- The Green Grid · Water Usage Effectiveness ↗ (opens in a new tab)Metric-body explanation · Checked 28 September 2026
Prepared with AI assistance. Publication authorized by Tommaso Luci; this does not claim independent technical peer review. Kovara Research is the publication label, not a claim of an independent laboratory or a named analyst team.