Immersion Cooling vs Air Cooling for High-Density Racks

Why cooling is now a buying decision

For most of the history of colocation, cooling was something the facility handled and the customer rarely thought about. You rented a rack, plugged in standard servers, and cold air arrived at the front of the cabinet. That still works for a large share of enterprise equipment.

High-density computing is changing the conversation. GPU servers for AI training and inference, and high-performance computing clusters, can pack far more power into a single rack than traditional workloads, and almost all of that power becomes heat. At some point, moving enough air through a cabinet stops being practical, and liquid cooling becomes part of the plan.

Immersion cooling is the most radical of the liquid options: the servers themselves sit in fluid. This guide goes deeper than our overview of data centre cooling technologies and compares immersion with air cooling for dense racks.

How air cooling works, and where it runs out

Air cooling moves conditioned air to the front of the rack, through the servers, and out the back as hot exhaust. Good facilities separate hot and cold air with aisle containment, blanking panels and careful cable management, and may add in-row coolers or rear-door heat exchangers to handle denser cabinets.

Its strengths are well known:

• Compatibility. Virtually all standard servers, storage and network equipment are designed for it.

• Familiar operations. Technicians, spares and maintenance processes are standard across the industry.

• Flexibility. Equipment can be swapped, moved and upgraded without special handling.

The limit is physics. Air carries heat far less effectively than liquid, so cooling a denser rack means moving much more air, using more fan power and producing more noise. Hotspots become harder to control, and above a certain density per rack, air alone becomes impractical. Where that point sits depends on the hall design and the hardware, so ask each facility what density it supports with air in the specific hall you would use. Our power and redundancy guide explains how kW per rack is measured and priced.

How immersion cooling works

In immersion cooling, servers are placed in a tank filled with a dielectric fluid, a liquid that does not conduct electricity. Heat moves from the components directly into the fluid, which is then cooled through a heat exchanger connected to the facility's water or cooling loop.

There are two main approaches:

• Single-phase immersion. The fluid stays liquid. It is pumped or circulated through the tank and a heat exchanger. Fluids are typically engineered oils or synthetic hydrocarbons. This is the more common approach in commercial deployments.

• Two-phase immersion. The fluid boils at a low temperature on the hot components, and the vapour condenses on a cooled coil and drips back down. It can remove heat very effectively, but it relies on specialised engineered fluids, and some of the fluids historically used belong to the PFAS family of chemicals, which face growing regulatory and supply scrutiny around the world. Check the fluid's environmental, safety and supply position before choosing this route.

Immersion is different from direct-to-chip liquid cooling, where coolant runs through cold plates attached to processors and the rest of the server is still air cooled. Many dense AI deployments use direct-to-chip, so compare both options rather than assuming immersion is the only alternative to air. Our AI infrastructure guide covers the wider requirements of GPU clusters.

Air vs immersion: the trade-offs

Density and heat removal. Immersion can cool much higher densities per rack-equivalent than air, because the liquid is in direct contact with every component.

Energy use. Removing server fans and large air-handling loads can reduce the energy spent on cooling. Ask for measured results from comparable deployments rather than relying on brochure claims, and read our guide to checking green data centre claims.

Hardware compatibility. Not all equipment is suitable for immersion. Servers often need modification, such as removing fans and replacing some thermal materials, and spinning hard drives are generally not suitable. Confirm with your hardware vendor whether immersion affects the warranty, and whether they offer immersion-ready models.

Space and floor loading. Tanks are horizontal rather than vertical, so they use floor area differently from racks, and a full tank of fluid is heavy. The facility needs to confirm floor loading and layout.

Operations. Swapping a component means lifting a server out of the fluid, letting it drain and handling it carefully. That needs trained staff, the right tools and clear procedures. Ask whether the facility's remote hands team is trained for immersion work.

Fluid management. The fluid has a cost, needs monitoring and eventually replacement, and must be handled and disposed of according to its safety data sheet and local regulations.

Resilience. Pumps, heat exchangers and the facility water loop all need redundancy, just as air systems do. Ask how the design achieves N+1 or better redundancy and what happens to tank temperatures if a pump or chiller fails.

Where immersion starts to make sense

Immersion is not a default upgrade. It tends to suit:

• Dense GPU or HPC clusters where air cooling cannot meet the heat load.

• Workloads with sustained high utilisation, where cooling efficiency matters over years.

• Deployments where the hardware is chosen with immersion in mind from the start.

• Sites where space, noise or the energy cost of air handling is a constraint, including some edge and regional deployments and modular builds.

Air cooling, often with containment or rear-door heat exchangers, usually remains the simpler choice for mixed enterprise equipment, moderate densities and environments where hardware changes often. Many facilities will end up running both, with immersion or direct-to-chip in dedicated areas.

Facility readiness in Australia

Liquid cooling capability varies widely between Australian facilities, and even between halls in the same building. Some operators have announced liquid-cooled capacity for AI customers, while many existing halls were designed around air. "Liquid-cooling ready" can mean anything from a fully commissioned water loop to space that could be fitted out later.

Before you shortlist, ask each provider:

• Do you support immersion, direct-to-chip, or both, in a hall available now?

• Who supplies and owns the tanks, fluid and coolant distribution units?

• What are the floor loading limits, water supply temperatures and redundancy design?

• Who is responsible for leaks, fluid top-ups, disposal and server handling?

• How are tank temperatures and fluid condition monitored, and how are alarms escalated?

• How is power metered and billed for liquid-cooled deployments?

Compare answers across facilities in Sydney, Melbourne, Brisbane and Perth, and see our GPU hosting guide if you are still deciding between owning hardware and renting it.

DChunt is independent and does not sell data centre space or cooling equipment. Use our comparison tool and provider directory to shortlist facilities, or the free assessment to check which sites can support your density.