The arrival of dense AI and high-performance-computing racks has pushed heat loads past what moving air can practically handle, and the industry's answer is to cool with liquid instead. Liquid cooling brings a fluid into direct or near-direct contact with the hottest components, using water's or a dielectric fluid's far greater capacity to carry heat. This guide defines the two main families, direct-to-chip cold plates and immersion, explains why hundred-kilowatt racks force the move off air, introduces the coolant-loop concepts, and covers the flow, pressure, and leak points that a liquid loop demands.
Liquid Cooling in one line: Liquid cooling in a data center uses a liquid rather than air to carry heat away from IT equipment, because liquid can absorb and move far more heat per unit volume than air. The two main approaches are direct-to-chip cooling, where a cold plate sits on the processor and coolant flows through it, and immersion cooling, where whole servers are submerged in a non-conductive dielectric fluid. Both are driven by the very high rack densities of AI and HPC hardware, which exceed what air cooling can practically remove.
Direct-to-chip cooling, sometimes called cold-plate cooling, brings the liquid right to the hottest components. A cold plate, a metal block with internal channels, is clamped onto the processor or accelerator in place of a conventional air heatsink, and coolant is pumped through the plate to carry heat straight off the chip. The servers keep most of their conventional form, but the biggest heat sources are now cooled by liquid, with only the remaining lower-power components still cooled by air. The coolant that leaves the cold plates is warm, and its heat is passed through a coolant distribution unit into the facility's cooling system.
Immersion cooling takes a more radical approach: entire servers, boards and all, are submerged in a tank of dielectric fluid, a liquid that does not conduct electricity so it can safely bathe live electronics. In single-phase immersion, the fluid stays liquid throughout, is warmed by the components, and is pumped through a heat exchanger to shed its heat before returning to the tank. In two-phase immersion, the fluid is chosen to boil at a low temperature right at the hot components, and the vapour rises, condenses on a cooler surface above, and drips back, using the latent heat of boiling to move very large amounts of heat with little pumping. Two-phase is extremely effective but depends on specialised fluids and careful containment of the vapour.
The two families suit different situations. Direct-to-chip is often the more practical retrofit, because it can be applied to otherwise conventional racks and coexists with air cooling for the rest of the server. Immersion delivers the highest densities and eliminates server fans entirely, but it changes how equipment is racked, serviced, and handled, and it commits a facility to fluid-filled tanks. Many operators adopt direct-to-chip first as the bridge and consider immersion for the very densest, most specialised deployments.
Air is a fundamentally weak carrier of heat. To remove a lot of heat with air you must move a large volume of it fast, and there is a hard practical ceiling on how much airflow you can force through a rack before the fan power, noise, and air velocity become self-defeating. For years that ceiling was comfortably above typical rack loads, so air cooling, helped by containment, in-row units, and rear-door exchangers, kept pace. But dense AI accelerators concentrate enormous power into small packages, and racks packed with them can draw well past a hundred kilowatts, which is simply more heat than air can strip off the chips before they overheat.
Liquid changes the arithmetic. A given volume of water can absorb roughly thousands of times more heat than the same volume of air for the same temperature rise, so a modest flow of coolant through a cold plate removes what would take a gale of air. Bringing the coolant into direct contact with the chip also cuts out the thermal bottleneck of transferring heat into air and then chasing that air around a room. The result is that liquid cooling not only makes extreme densities possible, it does so while using less energy to move heat, which improves overall efficiency at the same time as it solves the raw capacity problem.
This is why the move to liquid is being driven by the workload rather than by preference. Operators would generally rather keep the simplicity of air if they could, but the physics of hundred-kilowatt racks leaves no room: the heat has to be taken off the silicon by something denser than air. As these workloads spread, facilities are being designed or retrofitted with the water loops, distribution units, and monitoring that liquid cooling requires, and air cooling is increasingly reserved for the lower-density equipment that still lives alongside the accelerators.
A liquid loop is a piece of process plumbing, and it must be instrumented like one. The first essential measurements are flow and temperature: the coolant flow rate to the cold plates or through the immersion loop, and the supply and return temperatures, which together reveal how much heat the loop is carrying and whether it is keeping the components in their safe range. A falling flow or a rising supply temperature is an early sign that the loop is not delivering the cooling the hardware needs, and because the components are so power-dense, they can overheat quickly when cooling falters.
Pressure is the second pillar. A liquid loop runs at a controlled pressure, and both too little and too much are problems: low pressure or a lost pump can starve the cold plates, while excessive pressure stresses fittings and raises the risk of a leak. Monitoring supply and return pressure, and the differential across the equipment, confirms the pumps are delivering and the loop is intact. Pump status and speed sit alongside pressure, because a loop typically runs redundant pumps and the failover of one to another is an event operators must see.
The point that matters most for safety is leak detection, because liquid and electronics are a dangerous combination even when the fluid is dielectric. Leak-detection sensing along manifolds, hoses, cold-plate connections, and tank boundaries must alarm instantly and, in many designs, trigger isolation. All of these signals, flow, temperature, pressure, pump status, and leak alarms, are exactly the kind of tags a cloud SCADA platform such as Merobix is built to gather, trend, and alarm on. The same platform that watches pumps, pressures, and flows in water and oil and gas facilities can carry a data center's liquid-cooling loop, so operators see the health of the loop that keeps their most expensive hardware alive in the same monitored, remotely visible way they see the rest of their plant.
Direct-to-chip cooling puts a cold plate on the hottest components and pumps coolant through it, while the rest of the server may still be air-cooled. Immersion cooling submerges entire servers in a non-conductive dielectric fluid so the whole board is bathed in liquid. Direct-to-chip is often the easier retrofit, whereas immersion reaches the highest densities and removes server fans entirely but changes how equipment is racked and serviced.
Air carries very little heat per unit volume, so removing a large heat load requires moving a huge volume of air, and there is a practical limit to airflow before fan power, noise, and air velocity become self-defeating. Dense AI accelerators pack enormous power into small chips, and racks of them can exceed a hundred kilowatts, more heat than air can strip off before the components overheat. Liquid, which carries far more heat, is needed to keep up.
In single-phase immersion the dielectric fluid stays liquid the whole time, is warmed by the components, and is pumped through a heat exchanger to shed its heat. In two-phase immersion the fluid boils at a low temperature right at the hot components, and the vapour rises, condenses on a cooler surface, and drips back, using the heat of boiling to move very large amounts of heat efficiently. Two-phase is more effective but depends on specialised fluids and careful vapour containment.
Merobix reads your field devices into a cloud SCADA - the real thing behind these terms, live in days from any browser.