Direct-to-chip liquid cooling: what it changes in the plant room

For most of the last two decades, data centre cooling has meant moving air: CRAH units, raised floors or fan walls, hot and cold aisle containment, and chilled water at temperatures chosen to keep the air cold enough. Accelerated computing has changed the starting point. Processors and GPUs in current AI and HPC platforms produce heat densities that air struggles to remove at rack level, and several hardware vendors now design their high-end platforms around direct liquid cooling.

For estates teams, M&E consultants and contractors, the result is less about the server and more about the plant behind it. This article sets out what direct-to-chip cooling adds to a facility, and where the risks sit.

The basic architecture

Direct-to-chip (DLC) cooling puts cold plates on the hottest components. A coolant, usually treated water or a water and propylene glycol mix, flows through the cold plates and carries heat away from the server. Industry guidance, including ASHRAE's work on liquid cooling, describes two separate circuits:

• Technology cooling system (TCS): the secondary loop serving the racks and cold plates. Small passages, tight water quality limits, and connections that IT staff may make and break.
• Facility water system (FWS): the primary loop from the building's heat rejection plant: chillers, dry coolers, cooling towers or a combination.

A coolant distribution unit (CDU) sits between them. It contains a heat exchanger, pumps, filtration, expansion and controls, and keeps the two loops hydraulically separate. CDUs can be in-rack, in-row, or facility-scale units in a plant room serving many racks.

Warmer water, different plant

Cold plates can work with much warmer coolant than air-based systems. ASHRAE defines liquid-cooling facility water classes by maximum supply temperature, and many direct-to-chip designs target the warmer classes. Warmer facility water has three practical effects:

1. More free cooling. Dry coolers can reject heat without compressors for more of the year, and in some climates for all of it.
2. Smaller or fewer chillers on the liquid-cooled load, although most halls still carry a residual air load that needs conventional cooling.
3. More useful waste heat. Higher return temperatures make heat recovery easier. We cover that in a separate article on heat reuse.

The catch is that most facilities will run mixed loads for years: liquid-cooled racks alongside air-cooled ones, each wanting different water temperatures. That often means two facility water temperatures, or a primary system with local temperature control, which adds pumps, heat exchangers and valves to the plant room.

Water quality becomes a design discipline

Cold plate passages are narrow. Particulates, corrosion products, biological growth and incompatible materials can foul them and cause throttling or failure. In practice, the TCS needs:

• Fine filtration, often with duty and standby filters so elements can be changed without stopping flow.
• Materials compatible with the coolant and the server manufacturer's wetted-materials list, and no mixing of incompatible metals.
• Thorough flushing and cleaning before connection to IT equipment, with records.
• Ongoing monitoring and dosing, agreed with the operator and the hardware supplier.

These are plant room tasks. They are also easier to control in a workshop, where pipework can be cleaned, capped and pressure tested before it reaches a live site.

Redundancy and failure modes

An air-cooled hall has some thermal inertia: lose cooling and temperatures rise over minutes. A liquid-cooled processor relies on continuous flow, so loss of pumping is a far more immediate event. That changes the redundancy conversation:

• Pumps: N+1 as a minimum on CDUs and facility pumps, with automatic changeover and pumps on separately protected power supplies.
• Heat exchangers: duplicated, or sized so one can be isolated for cleaning without derating the system.
• Distribution: ring mains or dual A and B feeds to rows, with isolation valves positioned so any section can be maintained.
• Power: pumps and controls on UPS-backed supplies, so flow continues through a mains failure until generators pick up.
• Leak detection: under CDUs, along pipe routes and at rack manifolds, linked to the BMS or EPMS.

Whether a site needs N+1, 2N or a concurrently maintainable arrangement is a decision for the operator and design team. Once made, it needs to be visible in the plant room layout: in valve positions, bypasses and access space, not only on the schematic.

Why prefabricated skids fit liquid cooling

Liquid cooling arrives in phases, often hall by hall or row by row as tenants deploy hardware. That suits a repeatable, factory-built approach:

• Repeatability: the same CDU primary skid, pump skid or heat exchanger module can be built for each phase, with the same drawings, spares and operating procedures.
• Cleanliness: pipework is fabricated, cleaned, pressure tested and sealed under workshop conditions.
• Testing: factory acceptance tests confirm pump duties, valve operation, controls and BMS points before delivery.
• Programme: skids are built while the hall is being fitted out, then installed in short windows with fewer trades on a live site.

Questions to settle early

1. What share of the IT load will be liquid-cooled, and over what timescale?
2. Which facility water temperatures will the hardware accept, and how will residual air load be handled?
3. Where will CDUs sit, and who owns the TCS: the operator, the tenant or the hardware vendor?
4. What redundancy level applies to pumps, heat exchangers and distribution?
5. Who sets the water quality specification, and who maintains it after handover?
6. How will new rows connect without shutting down live ones?

Answering these before design freeze avoids the most expensive problem in liquid cooling: plant rooms laid out for air and retrofitted for water.

Published 12 November 2026 by Prefab Plant Systems, a trading name of Combicert Ltd. General guidance only; every project needs its own design.