In short
- Liquid becomes the practical choice above about 40–50 kW per rack; GB200-class racks at 120–132 kW are designed for it.
- Direct-to-chip cold plates capture typically 70–90 % of the heat; the rest still needs air.
- A CDU separates the clean server loop (TCS) from the facility water (FWS) with a heat exchanger.
- Immersion captures almost all the heat and removes most fans, but changes the hardware and the service model.
- Warm liquid loops (W32 and up) allow chiller-less operation in much of Europe and make heat reuse easier.
When does a data center need liquid cooling?
Air carries little heat per volume. At a 12 K rise, a 132 kW rack would need about 33,000 m³/h of air, more than any rack, fan or aisle can move sensibly. Water carries roughly 3,500 times more heat per volume than air, so a few litres per second do the same job.
In practice the switch happens between 40 and 50 kW per rack. Below that, contained air and rear doors still work; above it, liquid moves from option to requirement. Uptime Institute's survey, as reported by mgrid, found 19 % of operators had deployed liquid cooling by early 2026 and 36 % plan to within 12–24 months.
Direct-to-chip cold plates
Metal plates with fine internal channels sit on the GPUs, CPUs and sometimes memory and voltage regulators. Coolant flows through them via manifolds at the back of the rack and quick disconnects on each server.
- Single-phase: a water-glycol mix stays liquid and warms by 8–15 K. This is the common design for AI racks today.
- Two-phase: a refrigerant boils inside the plate and absorbs heat as latent heat. Higher capacity per plate, more complex fluid handling.
Cold plates typically capture 70–90 % of the rack heat. Power supplies, drives and network cards stay air-cooled, so the room still needs fan walls or CRAHs for the remainder. Size both sides in the heat balance.
The coolant distribution unit (CDU)
The CDU is the interface between the IT and the building. It holds a liquid-to-liquid heat exchanger, pumps, filters and controls:
- Technology cooling system (TCS): the secondary loop through the racks, with tightly controlled chemistry and fine filtration.
- Facility water system (FWS): the primary loop from the dry coolers or chillers.
- Approach: a few kelvin between facility supply and TCS supply, often around 3 K.
CDUs come in-rack, in-row or as floor-standing units serving several rows. Liquid-to-air CDUs (sidecars) reject heat into the room and suit pilots in halls without facility water, but they only move the problem back to the air system.
Immersion cooling
Servers sit vertically in tanks of dielectric fluid that does not conduct electricity.
- Single-phase: a hydrocarbon or synthetic oil is pumped through a heat exchanger. Simple, robust, heavy tanks.
- Two-phase: an engineered fluid boils on the hot components and condenses on a coil in the tank lid. Very high capacity, but most such fluids are fluorinated and face growing regulatory pressure on PFAS.
Immersion captures nearly all the heat and removes most fans, so it is close to silent. The trade-offs are servers adapted for the fluid, cranes or lifts for service, floor loading and a smaller choice of vendors.
Direct-to-chip vs immersion, side by side
| Criterion | Direct-to-chip | Immersion |
|---|---|---|
| Heat to liquid | ~70–90 % | ~95–100 % |
| Residual air cooling | Needed | Minimal |
| Rack form factor | Standard racks with manifolds | Tanks; servers adapted for fluid |
| Service | Like air-cooled, plus quick disconnects | Lift servers out of the fluid |
| Floor loading | Similar to dense air racks | High; tanks full of fluid |
| Vendor support for AI hardware | Mainstream | Narrower |
| Retrofit into an existing hall | Feasible with CDUs | Harder |
For most new AI halls direct-to-chip is the default because the GPU makers design for it. Immersion fits edge sites, crypto and specialised HPC, and operators willing to own the service model.
Water quality, leaks and service
- Water quality: keep TCS chemistry within the maker's limits: corrosion inhibitors, biocide, conductivity and fine filtration. Materials in the loop must be compatible.
- Leaks: use dripless quick disconnects, leak detection cables under manifolds and, where offered, negative-pressure loops that pull air in instead of pushing water out.
- Commissioning: flush and clean the loop before IT is connected and prove the CDUs at load. See allinCX for the L1–L5 sequence.
- Water temperature: warmer facility water saves energy; check the IT maker's allowed supply range and the ASHRAE W classes.
Questions about liquid cooling
Is liquid cooling more efficient than air?
Usually yes. Fans in the servers and the room slow down or disappear, and warmer water allows more free cooling. The gain depends on how warm the facility water can run.
Can I add liquid cooling to an existing data center?
Yes, mostly with direct-to-chip and CDUs. You need facility water near the racks, or liquid-to-air CDUs as a stopgap, plus floor capacity and space for manifolds.
Does liquid cooling remove the need for chillers?
Often. With W32 or warmer water and a 6 K dry-cooler approach, dry coolers carry the full load up to about 26 °C outdoor. Hot sites keep trim chillers.
What happens if a cold plate leaks?
Leak detection isolates the affected rack or loop. Dripless connectors, pressure monitoring and negative-pressure designs keep the risk low; the TCS fluid is usually treated water-glycol.
Sources
- ASHRAE TC 9.9 liquid cooling guidelines: TCS and FWS terminology, W classes.
- Uptime Institute survey on liquid cooling adoption, as reported by mgrid.
- Capture shares and densities reflect common industry practice; check the IT maker's specifications.