Direct-to-chip and immersion cooling are shifting the design centre of gravity in Australian AI data centres from air handling to hydraulics. Rack densities being briefed for new AI compute facilities routinely exceed what conventional air-cooled containment can dissipate — and the ELV scope around leak detection, monitoring and fire suppression is changing to match.
Why AI Compute Racks Have Outgrown Air Cooling
A conventional air-cooled Australian data centre hall is typically designed for rack densities in the 5-15kW range. AI training and inference racks — dense GPU clusters — routinely reach 40-100kW per rack and climbing, densities at which the volume of air needed to remove that heat becomes physically impractical within standard rack and aisle geometries. Liquid removes heat far more efficiently per unit volume than air, which is why liquid cooling has moved from a specialist HPC niche to a mainstream requirement for any Australian facility targeting genuine AI compute tenancy.
Direct-to-Chip vs Immersion: Different Design Consequences
- Direct-to-chip cooling circulates coolant through cold plates mounted on the CPU/GPU package, removing heat at the highest-density component while lower-heat parts of the server (RAM, storage, networking cards) still rely on conventional air cooling — a hybrid approach that keeps much of the existing rack and containment design familiar.
- Immersion cooling submerges the entire server in a dielectric fluid, removing heat from every component at once — more thorough, but a genuinely different rack form factor, maintenance procedure (servicing a server means lifting it out of fluid) and facility layout than a conventional data hall.
- Most current Australian AI facility briefs specify direct-to-chip as the near-term default, with immersion reserved for the highest-density specialist deployments where direct-to-chip's remaining air-cooling load still can't be managed conventionally.
Leak Detection Becomes Core ELV Scope, Not an Afterthought
Introducing liquid into a rack full of live electronics changes the risk profile in a way conventional room-perimeter water detection doesn't adequately address. A coolant leak at a cold plate connection or manifold joint can develop and cause damage faster than perimeter leak-detection cabling, laid only around the room's edges, would ever catch it. Leak detection sensors at cold plate connections, manifold joints, and beneath individual racks should be specified as standard ELV scope on any Australian liquid-cooled facility, integrated into the building management and DCIM platforms with automatic isolation valve triggers, not just an alarm to a human operator.
Design takeaway: Specify rack-level and connection-level leak detection with automatic isolation as core ELV scope for any liquid-cooled Australian facility — perimeter water detection designed for conventional air-cooled halls reacts far too slowly for a coolant leak inside a live, high-density rack.
Fire Suppression and Containment Rethinking
Liquid cooling loops also change fire-suppression design assumptions. Clean-agent gaseous suppression systems designed for a conventional air-cooled hall need re-evaluation against a rack containing pressurised coolant lines and, in some deployments, dielectric fluid — the suppression design team needs early coordination with the mechanical cooling design, not a standard gaseous-suppression spec applied unchanged from a previous project's air-cooled hall.
Frequently Asked Questions
What's the difference between direct-to-chip and immersion liquid cooling?
Direct-to-chip cooling circulates coolant through cold plates mounted directly on the CPU/GPU, removing heat at the component while the rest of the server still uses air cooling for lower-heat components. Immersion cooling submerges the entire server in a dielectric fluid, removing heat from every component simultaneously — a more thorough but more disruptive change to rack and maintenance design.
Does liquid cooling remove the need for a raised floor or air handling entirely?
Not entirely — even AI racks with liquid-cooled compute usually still need some air cooling for networking equipment, power distribution and other lower-heat-density components, so most Australian AI facilities design a hybrid system rather than eliminating air handling completely.
How does leak detection change with liquid cooling in the rack?
Leak detection sensors need to be specified at cold plate connections, manifold joints and beneath the rack itself as standard ELV scope — not an afterthought — since a coolant leak in a rack full of live electronics is a fast-developing failure mode that conventional water-detection cabling laid only at the room perimeter won't catch in time.