Industry Solutions / AI and High-Density Computing

UPS and Precision Cooling for AI, GPU and High-Density Infrastructure

Most Australian data halls were designed around an assumption that no longer holds. A raised floor, a row of perimeter CRAC units and a comfortable 4 kW to 8 kW per rack was a sound design for a decade of enterprise virtualisation. A single NVIDIA DGX-class node breaks it. Put eight of them in a rack and the heat rejection problem stops being an HVAC tuning exercise and becomes a question of whether air can physically carry the load at all.

The same is true on the electrical side, and it gets less attention. GPU clusters do not draw a smooth, predictable load. They step. A training job starting, checkpointing or finishing moves tens of kilowatts in seconds, and that behaviour interacts badly with UPS systems and generators sized on the old assumption of a flat IT load with a modest diversity factor.

Indigi Power and Cooling is an Indigenous and Veteran owned critical power and cooling specialist, registered with Supply Nation and ICN Gateway. We work on both halves of this problem, because in high-density environments they are the same problem. Cooling capacity you cannot power is useless, and power you cannot reject the heat from will throttle your GPUs regardless of how much of it you have.

Where your rack density actually sits

Before choosing a cooling architecture, establish the number honestly. The right answer is driven almost entirely by kilowatts per rack, not by the brand of hardware or the size of the room.

The practical dividing line for most Australian sites sits between 20 kW and 40 kW. Below it, a well-contained air solution using in-row units is usually cheaper, simpler and easier to maintain. Above it, you are fighting physics and the economics turn quickly. We size this against your actual measured load rather than the nameplate rating on the hardware, because GPU nameplate figures and sustained draw are rarely the same number.

Cooling architectures for AI and GPU infrastructure

In-row cooling

The air conditioning unit sits in the row, directly between the racks, rather than at the room perimeter. Short air paths mean far less fan energy and much tighter control of supply temperature at the server inlet. Combined with hot or cold aisle containment it will comfortably handle racks up to roughly 40 kW. This is the highest-value upgrade for most Australian sites currently running perimeter cooling, because it does not require any change to the servers themselves. We supply and install the Vertiv Liebert CRV in-row cooling system, which delivers 10 kW to 50 kW per unit with no raised floor required.

Rear door heat exchangers

A liquid-cooled coil replaces the rear door of the rack, capturing heat at the point it leaves the servers so it never enters the room. Passive versions rely on server fans, active versions add their own. This is often the least disruptive route to 40 kW and beyond, because the servers are untouched and the room air handling can stay as it is. It does require chilled water or a coolant loop brought to the rack, which is the real project cost.

Direct-to-chip liquid cooling

Cold plates sit directly on the CPU and GPU packages, and a coolant distribution unit (CDU) manages the loop between the facility water and the technology water inside the racks. This handles the densities that current AI hardware actually produces. Note that direct-to-chip typically removes 70 to 80 per cent of the rack heat, not all of it, so you still need air cooling for the remainder, covering memory, drives and power supplies. Sites that plan for liquid and forget the residual air load end up with a hot room and a puzzled contractor.

Immersion cooling

Servers are submerged in a dielectric fluid. Single-phase immersion circulates the fluid to a heat exchanger, two-phase relies on the fluid boiling and condensing. Thermally it is the most capable option available and it removes essentially all fan energy, but it demands purpose-built tanks, changes how you handle and service hardware, and carries warranty implications with some OEMs. It suits a greenfield AI build far better than a retrofit.

Why AI loads break conventional UPS sizing

This is the part that gets missed. A conventional IT load is close to constant, and UPS and generator sizing has historically leaned on that. AI workloads behave differently in three specific ways, and each one has a consequence for the electrical design.

• Step loads. A training job starting or a synchronised checkpoint across a cluster can move tens of kilowatts within seconds. The UPS must hold output voltage through that step without dropping to bypass, and the generator behind it must accept the block load without an unacceptable frequency excursion.
• Very high peak-to-average ratio. Sizing on average draw will leave you short during peaks. Sizing on peak alone leaves the UPS running at low load factor, where double-conversion efficiency falls away and you are paying for capacity you rarely use. Modular UPS architecture is the usual resolution.
• High harmonic content and poor power factor on some GPU power supplies. Worth measuring rather than assuming, because it changes cable, switchgear and transformer sizing upstream.

In practice this means three-phase online double-conversion UPS with genuine modular scalability, N+1 at minimum, and a deliberate decision about whether the UPS carries the cooling plant as well as the IT load. In a high-density hall, losing cooling is a faster path to an outage than losing IT power. At 100 kW per rack the thermal ride-through, meaning how long you have before an over-temperature shutdown once air movement stops, can be under a minute. Any credible AI facility design puts pumps, CDUs and fans on protected power.

The Australian context

Three things make the Australian AI infrastructure picture different from the American or European one.

Grid connection is the binding constraint in Sydney and Melbourne. Available capacity at the meter is frequently the limiting factor on how much AI compute a site can host, well before floor space runs out. That pushes operators toward efficiency measures that free up existing headroom rather than new connections, which is where Free Cooling, higher supply air temperatures and liquid cooling economics become commercially interesting rather than merely technically interesting.

Ambient conditions vary enormously. A Melbourne site gets far more usable Free Cooling hours than a Brisbane or Darwin one. Liquid cooling changes this calculation substantially, because a direct-to-chip loop can run at much warmer facility water temperatures than an air-cooled hall needs, which widens the window in which you can reject heat without mechanical cooling even in warmer climates.

Compliance and reporting. Electrical work is carried out to AS/NZS 3000, refrigerant work requires ARCtick licensing under the Ozone Protection and Synthetic Greenhouse Gas Management Regulations, and mechanical ventilation hygiene falls under AS/NZS 3666. Supply air temperatures should sit within the ASHRAE TC 9.9 recommended envelope of 18 to 27 degrees Celsius at the server inlet, though AI hardware often tolerates and benefits from the upper end. NABERS for Data Centres is increasingly requested in tenders, and PUE remains the headline efficiency metric operators are held to.

What Indigi does on an AI or high-density project

Indicative pricing for cooling design and installation runs from $8,000 to $15,000 for small rooms, $25,000 to $45,000 for medium installations and $45,000 to $60,000 plus for large projects. High-density and liquid-cooled projects are quoted individually because the coolant loop, plant and containment work dominate the cost. UPS maintenance starts at $720 ex GST for a one-off visit or $1,940 per year ex GST for a standard annual plan.

Frequently asked questions

What is liquid cooling in a data centre?

Liquid cooling uses a fluid rather than air to carry heat away from IT equipment. Because liquids have far higher thermal transfer capacity than air, the coolant can be brought much closer to the heat source. The main forms are rear door heat exchangers, which cool air as it exits the rack, direct-to-chip cold plates mounted on the processors, and immersion, where servers are submerged in a dielectric fluid. AI and GPU workloads are the primary driver of adoption in Australia.

What liquid is used to cool data centres?

Direct-to-chip and rear door systems typically circulate treated water, often with propylene glycol added for freeze and corrosion protection, in a closed loop separated from the facility water by a coolant distribution unit. Immersion systems use a dielectric fluid, meaning one that does not conduct electricity, which is either a synthetic hydrocarbon in single-phase systems or an engineered fluid with a low boiling point in two-phase systems. Plain water is never in direct contact with live electronics.

What are the downsides of liquid cooling?

Higher capital cost, greater design complexity, and a leak risk that has to be engineered out with leak detection, quick-disconnect couplings and negative-pressure loops. It also changes your maintenance model, because servicing a liquid-cooled or immersed server is not the same job as servicing an air-cooled one, and some hardware warranties are affected. Direct-to-chip only removes 70 to 80 per cent of rack heat, so an air path is still required for the remainder. For densities below roughly 30 kW per rack, well-executed air cooling with containment is usually the better commercial decision.

How much power does an AI or GPU rack draw?

Current AI training racks commonly draw 40 kW to 80 kW, and flagship GPU configurations reach 130 kW and beyond. That compares with 3 kW to 8 kW for a conventional enterprise rack. The step change matters more than the absolute number, because it means the cooling architecture, the electrical distribution and the UPS sizing all have to be reconsidered together rather than uprated individually.

Can an existing data centre be retrofitted for AI workloads?

Often yes, up to a point. Adding hot or cold aisle containment and in-row cooling will typically take an existing hall from 5 kW to 8 kW per rack up to 30 kW or 40 kW without structural change. Beyond that you need a coolant loop brought to the racks, which means plant space, pipework and usually a coolant distribution unit, and that is a genuine capital project. The honest first step is a thermal and electrical survey to establish what your existing plant, floor loading and grid connection can actually support.

Should the cooling system be on UPS in a high-density facility?

Yes. In a high-density hall, thermal ride-through after a cooling failure can be under a minute, which is far shorter than the battery runtime protecting the IT load. If pumps, CDUs and fans are not on protected power, the facility will hit over-temperature shutdown long before the UPS batteries are exhausted. Cooling plant on UPS is standard practice in any credible AI facility design.

What UPS topology suits an AI or HPC environment?

Three-phase online double-conversion, modular, at N+1 redundancy or better. Modularity matters because AI loads have a high peak-to-average ratio, and a modular frame lets you match installed capacity to actual load so the system runs in its efficient band rather than lightly loaded. The UPS also has to hold output through large step loads without transferring to bypass, so step load response should be specified explicitly rather than assumed.

Where we work

Related services and equipment

• Liquid cooling solutions including CDUs, rear door heat exchangers and immersion
• Vertiv Liebert CRV in-row cooling, 10 kW to 50 kW per unit
• Vertiv Liebert PEX3 CW chilled water CRAC for large and hyperscale halls
• Server room cooling Australia
• DX, CHW or CDW: which CRAC topology
• CRAC unit maintenance services
• Three phase UPS installation and commissioning
• UPS maintenance plans
• Data centres and IT infrastructure
• Gas detection for battery rooms and BESS

Planning an AI or high-density deployment? Send us your target rack density, node count, room dimensions, existing cooling plant and available electrical capacity. We will come back with a realistic thermal and power position, an architecture recommendation and an installed price. Indigi Power and Cooling is Indigenous and Veteran owned, and registered with Supply Nation and ICN Gateway. Contact Indigi Power and Cooling Server Room Cooling