Liquid cooling works by circulating fluid through a sealed loop. A cold plate conducts heat from the chip surface into the coolant, a pump carries that fluid to a heat exchanger, and the cooled fluid returns to repeat the cycle. Water holds roughly 3,500 times more thermal energy per litre than air, which is why this loop handles GPU rack densities air cooling cannot.
The Physics Behind Why Liquid Beats Air
Heat flows from a hotter surface to a cooler medium, and efficiency depends on the thermal properties of that medium. Air has low heat capacity and low thermal conductivity; fans help but consume power steeply as velocity increases. Water has a specific heat capacity of 4.18 joules per gram per degree Celsius versus 1.005 for air, and because water is far denser, the volume comparison is more striking still: water carries roughly 3,500 times more heat per litre. That is why air cooling methods reach a ceiling around 30 to 40 kilowatts per rack while liquid configurations can exceed 100 kilowatts.
The Closed Loop, Component by Component
A liquid cooling loop is a sealed circuit. Coolant travels in one direction continuously, absorbing heat at one point and rejecting it at another. Here is what each component does:
- Cold plate: A copper or aluminium block with internal channels, mounted directly on the processor. Coolant flows through the channels and absorbs heat by conduction from the chip surface. Mounting pressure and thermal interface material quality determine how well heat transfers from the die into the plate.
- Pump: Drives coolant through the loop at a controlled flow rate. Most data center deployments use variable-speed pumps that adjust to changing thermal load. Redundant pumps are standard because a pump failure stalls the loop and temperatures rise within minutes.
- Coolant lines and manifolds: Insulated hoses or rigid piping connect servers to the distribution system. Quick-disconnect fittings let individual servers be removed without draining the loop. A manifold feeds coolant from one supply line into multiple server-level branches simultaneously.
- Coolant Distribution Unit (CDU): The rack-level or row-level hub. The CDU houses the pump, monitors temperatures and flow rates, and contains a heat exchanger that transfers heat from the server-side coolant into a separate facility water circuit. It also manages loop pressure and houses a small reservoir to handle fluid volume changes.
- Heat exchanger: Inside the CDU, thin metal walls separate the hot server-side coolant from cooler facility water. Heat conducts across the wall without the two fluids mixing. The server-side fluid exits cooled and returns to the cold plates; the facility-side water exits warmer and carries the heat toward outdoor rejection equipment.
- Facility cooling loop: The building-level water circuit that moves heat from CDUs to chillers, cooling towers, or free-cooling heat exchangers outside. This is where the heat leaves the building entirely.
Single-Loop vs. Two-Loop (Primary/Secondary) Configurations
Most data center liquid cooling deployments use two separate fluid circuits. The server-side loop touches the hardware directly; its coolant is treated, deionized water with corrosion inhibitors and glycol, held to tight purity standards to protect the metals in cold plates and manifolds. The facility-side loop connects to the building’s chilled water plant and does not need to meet the same standards. The CDU heat exchanger separates the two without letting them mix, which means each loop can be serviced independently. This primary-secondary architecture is now the standard approach in hyperscale and colocation builds.
Direct-to-Chip vs. Immersion: Two Approaches to the Same Problem
Direct-to-chip (DTC) cooling, also called cold plate cooling, is the dominant approach in current data center deployments. Cold plates attach to the highest-power components, primarily processors and AI accelerators like the NVIDIA H100 and H200, which together account for the majority of heat generated in an AI training or inference server. Other components, memory, power electronics, networking chips, typically continue to be cooled by air within the same chassis. This hybrid approach means DTC servers can run in standard rack enclosures with conventional power distribution and server management infrastructure, which substantially reduces the cost and complexity of deployment.
Immersion cooling takes a different approach: the entire server, board, components, and all, is submerged in a dielectric fluid that cannot conduct electricity and will not damage electronics. Heat is absorbed by the fluid from every component simultaneously, removing the need for internal fans, heatsinks, or any air movement within the system. Single-phase immersion uses fluorocarbon or synthetic hydrocarbon fluids that remain liquid throughout the process; the heated fluid is pumped to an external heat exchanger and returned. Two-phase immersion systems use engineered fluids with low boiling points; the fluid vaporises on contact with hot components, rises to a condenser coil where it gives up its heat and returns as liquid, exploiting the much higher heat transfer available from a phase change.
Immersion supports higher power densities than DTC and removes the need for server fans entirely. The trade-off is operational: removing a server for maintenance means handling fluid, and the tank infrastructure is less standardised than rack-mount DTC. ASHRAE, which publishes the thermal guidelines data centers are designed around, has been updating its standards to formally address both DTC and immersion as they move from niche to mainstream.
What Coolants Are Actually Used
For direct-to-chip systems, the standard coolant is treated, deionized water with a corrosion inhibitor package. Tap water is avoided because dissolved minerals cause scaling and galvanic corrosion across the different metals in cold plates and manifolds. Most deployments add ethylene or propylene glycol, typically at 30 to 40 percent concentration, which lowers the freeze point to roughly minus 15 to minus 20 degrees Celsius. Biocides prevent bacterial growth in the loop over time.
Immersion cooling requires a different class of fluid entirely. The primary requirement is dielectric: the fluid cannot conduct electricity when it contacts live circuit boards. Fluorocarbon-based fluids and synthetic hydrocarbon mineral oils are the two main categories, each with different trade-offs around heat capacity, viscosity, environmental profile, and compatibility with PCB coatings and plastics that can swell or degrade with prolonged fluid contact.
Why AI GPU Racks Pushed Liquid Cooling Into the Mainstream
An NVIDIA H100 SXM has a thermal design power of 700 watts. A DGX H100 server holds eight H100 GPUs plus associated CPUs and networking, putting total server power draw above 10 kilowatts for a single 2U or 4U chassis. Pack a rack with several such servers and the rack-level draw reaches 40, 60, or 80 kilowatts, a density that makes conventional hot-aisle/cold-aisle air cooling impractical without very high airflow volumes and the fan power to drive them. A vendor like Vertiv, which manufactures both CDUs and DTC cooling infrastructure, notes that liquid cooling is now a design requirement rather than an option for racks above 30 to 40 kilowatts. The direction of travel is clear: as AI chip generations push thermal design power higher, liquid cooling becomes the only viable path for dense GPU deployments.
PC Liquid Cooling: The Consumer Version of the Same Principle
If you have seen an AIO (all-in-one) cooler in a desktop PC, you already understand the principle. A cold plate sits on the CPU, a small pump circulates coolant to a radiator at the case, and the radiator rejects heat into the room. The data center version does the same thing but rejects the heat outside the building entirely, via a CDU and facility water loop. The key difference is that data center liquid cooling removes heat from the building rather than adding to it, which is how well-designed facilities achieve Power Usage Effectiveness (PUE) values approaching 1.0 rather than the 1.3 to 1.6 typical of older air-cooled sites.
Frequently Asked Questions
How does liquid cooling work?
Liquid cooling circulates a fluid through a sealed loop. A cold plate conducts heat from a chip’s surface into the coolant. A pump moves the heated coolant to a heat exchanger in a Coolant Distribution Unit, which transfers the heat into a facility water circuit. The cooled fluid then returns to the cold plate to repeat the cycle, continuously removing heat from the processor.
What is direct-to-chip liquid cooling?
Direct-to-chip (DTC) liquid cooling attaches a metal cold plate directly to a processor or GPU package. Coolant flows through channels inside the cold plate, absorbing heat by conduction from the chip surface. The heated coolant travels to a CDU heat exchanger, gives up its heat to the facility loop, and returns cooled. Other components in the same server typically continue using air cooling in a DTC deployment.
What liquid is used in liquid cooling systems?
Direct-to-chip systems use treated, deionized water mixed with corrosion inhibitors and ethylene or propylene glycol to prevent freezing. Immersion cooling systems use dielectric fluids that will not conduct electricity, either fluorocarbon-based compounds or synthetic hydrocarbon oils. The specific fluid choice depends on the architecture, operating temperature range, hardware compatibility, and environmental regulations in the deployment region.
Is liquid cooling better than air cooling for servers?
Liquid cooling is significantly more effective at high power densities. Above roughly 30 to 40 kilowatts per rack, air cooling becomes impractical because the airflow volume and fan power required are no longer viable. For racks running AI GPU workloads at 60 to 100 kilowatts, liquid cooling is the only practical option. For lower-density deployments below 15 to 20 kilowatts per rack, air cooling remains cost-effective and mechanically simpler.
Do servers use liquid cooling?
Modern high-density servers, particularly those running AI accelerators, increasingly use liquid cooling. Most hyperscale operators building new AI infrastructure now deploy direct-to-chip cooling for GPU servers. General-purpose servers at lower power densities still predominantly use air cooling. The threshold for liquid cooling adoption in server deployments sits at roughly 30 to 40 kilowatts per rack, above which air systems cannot keep up without disproportionate fan power and airflow infrastructure.
