In brief
Direct-to-chip sends coolant through processor cold plates. Immersion puts compatible IT hardware into a dielectric fluid. Rear-door heat exchangers cool rack exhaust air. The suitable solution depends on server specifications, the residual air load, facility water and the operating model.
Liquid cooling is not new to data centers, but rising compute density has taken it into the mainstream in the space of a few years. A conventional rack draws a few kilowatts. A rack built for AI workloads can draw tens of kilowatts or more than a hundred. Density at that level is not cooled with air.
In practice liquid cooling divides into three solutions: cooling taken directly to the chip, submerging the whole machine in a dielectric fluid, and combinations of the two. This article compares them and sets out what each means for plant design, operation and maintenance. The wider picture is in our complete guide to data center cooling.
What a buyer should provide
When comparing options or retrofitting a hall, share rack power, IT vendor coolant requirements, available air and water cooling capacity and the phasing plan. Finkova can help define the plant-side installation, metering and commissioning interfaces.
Where air cooling runs out
Air is a poor medium for moving heat. Per unit of mass, water carries roughly four times the heat that air does, and per unit of volume the difference is more than three thousandfold. Moving the same heat with air therefore takes an enormous volume flow.
Fan power rises roughly with the cube of flow, so doubling the air volume multiplies the fan electricity several times over. At some point the limit arrives: the added cooling capacity costs more in electricity than it is worth, and noise and floor space grow with it. In practice a well executed air cooled hall with aisle containment handles rack loads whose airflow and temperature can be controlled, a little more in exceptional cases. AI racks exceed this several times over.
Liquid cooling captures the heat where it is generated, rather than spreading it into the hall air first and collecting it back afterwards. That improves efficiency and raises the temperature of the recovered heat, which is the precondition for making use of it.
Direct-to-chip
In direct-to-chip cooling a cold plate is fitted on top of the processor, with narrow flow channels inside it. Coolant passes through them, picks up the heat and travels via the rack manifolds to a CDU, which transfers the heat into the facility loop.
The single-phase version is by far the more common. The coolant stays liquid throughout, and is typically a water-glycol mix. In the two-phase version the coolant boils inside the cold plate and condenses back to liquid, which moves more heat but makes the system considerably more complex.
One practical feature of direct-to-chip cooling gets too little attention in projects: the residual load. Cold plates cover the processors and accelerators, not the rest of the machine. Memory, power supplies, network cards and drives still produce heat, and that heat leaves with the air. The proportion depends on the hardware, but broadly 70 to 80 per cent of the rack’s heat goes into the liquid. The rest has to be cooled with air, which means a liquid-cooled hall still needs air cooling. If that has not been sized, the hall heats up even though the liquid loop is working perfectly.
The strength of the solution is that it works with conventional server hardware and can be retrofitted. The weakness is that pipework, pressure and joints come inside the rack, and every quick disconnect is a potential leak point.
Immersion cooling
In immersion cooling the whole server is submerged in a tank filled with a non-conductive, or dielectric, fluid. No fans are needed at all, and in practice all the heat the machine produces goes into the fluid.
In single-phase immersion the fluid stays liquid and is pumped from the tank through a heat exchanger. The fluid is typically a synthetic or mineral based oil. In two-phase immersion the fluid boils on the surface of the components and condenses on a coil in the tank lid, which gives very high heat transfer capacity.
Immersion solves the density problem thoroughly, but it costs elsewhere. Tanks are heavy and change both floor loading and the whole hall layout. Servers have to be modified to suit the tank, which limits hardware choice and affects warranty terms. Maintenance means lifting a wet server out of a tank, so servicing procedures and drip management have to be designed separately. Retrofitting into an existing air-cooled hall is rarely sensible.
Two-phase solutions and PFAS regulation
Two-phase immersion fluids and their supply chains vary. Some fluorinated fluids fall within PFAS definitions. The applicable regulatory status and any exceptions need a substance-specific check. 3M announced in 2022 that it would exit PFAS manufacturing by the end of 2025.
Before selecting a two-phase solution, obtain written confirmation of the fluid formulation, long-term availability, service procedure, waste handling and the current EU regulatory position from the supplier. Treat that assessment as part of the life-cycle business case.
Rear-door heat exchangers as an intermediate step
A fourth option often goes unmentioned, even though in many projects it is the most sensible first move. A rear-door heat exchanger is fitted to the back of the rack, and the warm air leaving the rack is cooled by liquid in the door before it returns to the hall.
The solution brings liquid into the hall but not into the rack. The server internals are untouched, but confirm warranty implications with the IT vendor. No quick disconnect sits inside a server. Capacity typically reaches well above what air cooling manages, though not to AI rack levels. In many sites a rear-door heat exchanger is a good way to raise the density of an existing hall while building out the facility-side liquid loop ahead of the real transition.
Comparing the solutions
Single-phase direct-to-chip: supports dense compute. Plan residual air cooling and connection integrity.
Two-phase direct-to-chip: vendor-specific option. Assess fluid availability, servicing and life-cycle support.
Single-phase immersion: requires compatible server hardware and a different floor layout and maintenance process.
Two-phase immersion: assess the particular fluid’s life-cycle availability and regulatory status.
Rear-door heat exchanger: useful for some retrofits when facility water and airflow capacity are available.
Hybrid is the practical answer
Pure solutions are rare in reality. A typical project today cools the compute racks with liquid, the network gear, storage and ancillary loads with air, and the residual load of the liquid-cooled racks with air as well. The site is therefore built around two parallel cooling systems that have to work together.
This affects the design in three ways. Air cooling cannot be designed out even when the main load moves to liquid. The liquid loops and CDUs have to be sized for a growing load, because density rises over the life of the site rather than staying at its installation level. And both systems have to survive the same fault scenario, which means redundancy cannot be designed separately for each.
What the choice means for the plant
The choice of liquid cooling method has a direct bearing on:
- floor loading, hall layout and service clearance
- hardware choice and manufacturer warranty terms
- the temperature of the recovered heat, and therefore its value
- serviceability in operation, meaning what can be maintained without stopping the hall
- leak risk and how leak detection is implemented
- fluid availability and regulation over the whole life of the installation
- the ability to raise density later
Of these, the last two most often decide whether the choice is regretted. A regulatory change can leave a working installation maintainable but impossible to extend, and the lack of headroom bites exactly when more capacity is needed.
Common misjudgements
The residual load is forgotten. A liquid-cooled hall still needs air cooling. This is the single most common sizing error.
The temperature level is set too low. A low supply temperature feels safe, but it cuts free cooling hours and lowers the value of the recovered heat. The level is agreed with the hardware manufacturer, not chosen as a precaution.
Servicing procedures are left undefined. Who disconnects the quick connector, how the loop is drained and what happens to spilled fluid. These are agreed before commissioning, not at the first service call.
Cleanliness is treated as an installation detail. Weld scale or swarf ends up directly in a cold plate flow channel and does not come back out. The subject is covered in the article liquid cooling water chemistry, filtration and material compatibility.
Headroom is sized for the load at installation. Density rises over the life of the site. CDU and loop sizing is covered in the article on CDUs in the data center.
Finkova’s scope and next step
Finkova installs and commissions data center cooling in the Nordics, with instrumentation and lifecycle support. The work scope, OEM requirements and acceptance criteria are agreed for each project.
Contact Finkova with the project inputs above to discuss the relevant installation, testing and handover interfaces.
Frequently asked questions
What is direct-to-chip cooling?
In direct-to-chip cooling a cold plate is fitted on the processor, and coolant flows through its narrow channels. The coolant picks up heat directly from the chip and carries it to a CDU. The single-phase version, using a water-glycol mix, is now by far the most common liquid cooling method.
What is immersion cooling?
In immersion cooling the whole server is submerged in a tank of electrically non-conductive fluid. No fans are needed, and in practice all the heat goes into the fluid. It does require modified servers, imposes heavy floor loads and changes servicing practice, so it suits new halls best.
What is the difference between single-phase and two-phase liquid cooling?
In single-phase cooling the fluid stays liquid throughout. In two-phase cooling it boils at the heat source and condenses back to liquid, which moves more heat. A two-phase system is more complex. Check the exact fluid’s classification, availability and applicable regulation.
Why does PFAS affect immersion cooling?
Some two-phase immersion fluids are fluorinated. 3M announced its exit from PFAS manufacturing by the end of 2025. Other products must be assessed by substance, supplier availability and the status of EU restrictions. Do not assume that every replacement fluid has the same regulatory position.
Does a liquid-cooled hall still need air cooling?
Yes. Cold plates cover processors and accelerators, but memory, power supplies, network cards and drives still produce heat that leaves with the air. The fraction of rack heat captured by liquid depends on the IT hardware and cold-plate coverage. Forgetting this residual load is the most common sizing error in liquid cooling projects.
What is a rear-door heat exchanger?
A rear-door heat exchanger is fitted to the back of the rack, and the warm air leaving the rack is cooled by liquid in the door before it returns to the hall. The server internals are untouched. Confirm any warranty implications with the IT vendor. It is a good way to raise density in an existing hall.
How much can air cooling handle?
The air-cooling limit depends on the server hardware, fan curves, airflow path, containment and the temperature conditions specified by the IT vendor. High-density AI racks commonly require a liquid-cooled approach.
Can liquid cooling be installed in an existing hall?
Yes. Direct-to-chip cooling works with conventional server hardware and can be retrofitted, as can rear-door heat exchangers. Immersion cooling, by contrast, changes floor loading and hall layout so fundamentally that retrofitting it into an air-cooled hall is rarely sensible.
What is the biggest risk in liquid cooling?
Leaks and contamination. Liquid cooling brings pipework, pressure and joints into the rack, and every quick disconnect is a potential leak point. The other risk is cleanliness: particles can obstruct cold-plate channels and may require component service or replacement. Both are addressed in installation and commissioning.
Why does liquid cooling improve waste heat recovery?
Liquid cooling concentrates heat into one loop at a higher temperature than air cooling, where the heat is spread through the whole hall. A higher temperature reduces the lift required of the heat pump, which improves recovery efficiency and makes selling heat into a district heating network more viable.
