AI Data Centres in Australia: What Changes for Power and Cooling
AI Infrastructure · High-Density Cooling · Critical Power
An AI data centre is a facility, or a part of one, built to run GPU and accelerator servers for training and running AI models. The big difference is density: AI racks typically draw 30 to 100 kW or more, compared with 5 to 15 kW for a traditional server rack. That changes how UPS, PDUs and cooling must be designed, and above roughly 40 to 50 kW per rack, air cooling alone usually cannot keep up. Indigi Power and Cooling designs and supports high-density power and cooling across Australia and is Supply Nation and ICN Gateway registered.
Search interest in AI data centres in Australia has climbed sharply through 2026, driven by hyperscale announcements, sovereign AI projects and businesses bringing GPU servers in-house. But AI infrastructure is not only a story for billion-dollar campuses. Universities, hospitals, engineering firms, mining companies and government agencies are installing a handful of GPU servers in existing server rooms, and quickly discovering that the room, the UPS and the air conditioning were never designed for that load.
This guide explains what makes an AI data centre different, how rack density drives the choice of cooling, what changes on the power side, and what to check before you put GPU servers into an existing Australian server room.
What is an AI data centre?
An AI data centre is designed around accelerated computing: servers packed with GPUs or other AI accelerators that process huge volumes of data in parallel. A single modern AI server can draw as much power as an entire rack of conventional servers did a decade ago. When those servers are packed into a rack, the result is a heat and power load far beyond what traditional raised-floor, perimeter-cooled data halls were built to handle.
There are broadly two types of AI workload. Training builds models and runs GPUs flat out for days or weeks, producing the highest sustained densities. Inference runs trained models to answer queries; it is often lower density per rack but more widely distributed, including in edge and on-premises server rooms.
How much power does an AI rack use?
According to the AFCOM State of the Data Center report, average rack density has grown from around 6 kW nine years ago to about 16 kW today, while AI workloads demand 30 to 40 kW or more, and only around one in five operators say they are ready to support the 50 to 70 kW per rack that many AI deployments need. The latest GPU platforms push some liquid-cooled racks well beyond 100 kW.
| Rack type | Typical power per rack | Heat to remove |
|---|---|---|
| Office comms or network cabinet | 1 to 3 kW | Low |
| Traditional enterprise server rack | 5 to 15 kW | Moderate |
| AI inference / small GPU cluster | 15 to 40 kW | High |
| AI training rack | 50 to 130+ kW | Very high -- liquid cooling needed |
Almost every watt that goes into a rack comes out as heat. A 40 kW rack needs roughly 40 kW of cooling, which is the equivalent of more than a dozen large domestic split systems pointed at a single cabinet.
AI data centre cooling: matching the method to the density
There is no single right cooling system for AI. The right answer depends on how many kilowatts are in each rack, how many racks there are, and whether you are building new or retrofitting. The table below is a practical guide.
| Cooling method | Practical rack density | Where it fits |
|---|---|---|
| Perimeter CRAC / CRAH units | Up to about 10 to 15 kW | Traditional server rooms and data halls |
| Hot or cold aisle containment + CRAC | Up to about 20 kW | Cheapest first step to lift density in an existing room |
| In-row cooling with containment | About 20 to 40 kW | Small GPU clusters and inference racks |
| Rear door heat exchanger | About 20 to 50+ kW | Retrofitting individual high-density racks without rebuilding the room |
| Direct-to-chip liquid cooling | 50 to 130+ kW | AI training racks and the latest GPU platforms |
| Immersion cooling | 100 kW+ per tank | Specialist and purpose-built high-density deployments |
Most real AI deployments are hybrid. Direct-to-chip liquid cooling typically captures the majority of the heat from GPUs and CPUs, but the remaining heat from memory, drives, power supplies and networking still has to be removed by air. That means precision air cooling does not disappear in an AI data centre; it is redesigned to work alongside liquid. Our liquid cooling solutions and CRAC precision cooling units are often specified together for this reason.
What changes on the power side
Cooling gets most of the attention, but AI changes the electrical design just as much.
- Three-phase power to the rack. A 40 kW rack cannot run on a couple of 10A circuits. High-density racks commonly need 32A three-phase feeds, often two per rack for A and B redundancy, with high-capacity rack PDUs to match.
- Bigger, more efficient UPS systems. AI loads push UPS capacity up quickly. Modular three-phase UPS systems that scale in blocks, running in high-efficiency modes, help control both capital cost and energy losses.
- Dynamic load behaviour. GPU training workloads can swing power draw rapidly as jobs start and stop. UPS and generator systems need to handle those step changes without instability.
- Cooling must be on protected power. In a low-density room, a short cooling outage gives you minutes before temperatures become critical. In a 50 kW rack it can be seconds. Coolant pumps, CDUs and fans in high-density designs are often placed on UPS so heat removal continues while generators start.
- Grid capacity. In Australia, grid connection capacity is increasingly a limiting factor for new AI sites. Every kilowatt saved on cooling and conversion losses is a kilowatt available for compute, which is why efficient cooling and low PUE (power usage effectiveness) now have a direct commercial value.
The Australian climate factor
Australia's hot summers and humid north make data centre cooling harder than in much of Europe or North America. High ambient temperatures reduce condenser and chiller efficiency and cut the number of hours available for free cooling. Liquid cooling helps here, because it can run at higher water temperatures, which widens the window for efficient heat rejection even in Brisbane, Darwin or Perth summers. Designing for local peak conditions, not average ones, is essential.
Putting GPU servers in an existing server room: a checklist
Most Australian organisations starting with AI will not build a new data centre. They will add a few GPU servers to an existing room. Before you do, check the following:
- Total power draw: confirm the GPU servers' real kW load and whether your circuits, switchboard and UPS have spare capacity.
- UPS rating and runtime: check that the UPS kW rating, not just kVA, covers the new load with headroom, and that battery runtime is still adequate.
- Cooling capacity: compare the room's total cooling capacity with the new heat load, and whether it can reach the rack where the GPUs will sit.
- Hot spots: a 20 kW rack in a room designed for 5 kW per rack will create hot spots even if the total cooling is enough. Thermal imaging and temperature sensors help confirm this.
- Airflow management: blanking panels, containment and brush strips are low-cost and can lift effective cooling capacity significantly.
- Supplemental cooling: if the room cannot cope, in-row units or a rear door heat exchanger on the GPU rack are usually faster and cheaper than replacing the room's cooling.
- Future growth: if AI use is likely to grow, plan pipework, power and space for liquid cooling now rather than retrofitting later.
Our AI and high-density GPU infrastructure service covers this assessment end to end, and our server rack cooling guide explains the rack-level options in more detail.
AI power and cooling support across Australia
Indigi Power and Cooling supports high-density and AI deployments nationally from our Brisbane, Sydney and Melbourne hubs, covering QLD, NSW, ACT, VIC, SA, WA, NT and Tasmania. We assess existing server room cooling and power, design supplemental and liquid cooling, and supply and maintain the UPS, PDUs and precision cooling that keep GPU infrastructure running.
Planning GPU servers or an AI cluster? We can assess your room's power and cooling headroom before the hardware arrives. Book a high-density readiness assessment.
Frequently asked questions
What is an AI data centre?
An AI data centre is a facility, or part of one, designed to run GPU and AI accelerator servers for training and running AI models. It differs from a traditional data centre mainly in power density, with racks drawing 30 to 100 kW or more, which requires high-capacity power distribution and advanced cooling, often liquid cooling.
How are AI data centres cooled?
AI data centres use a mix of methods matched to rack density. Up to about 40 kW per rack, containment, in-row cooling and rear door heat exchangers can work. Above roughly 50 kW, direct-to-chip liquid cooling is usually required, with precision air cooling removing the remaining heat. Immersion cooling is used in some specialist deployments.
How much power does an AI rack use?
AI inference and small GPU racks typically draw 15 to 40 kW, while AI training racks commonly draw 50 to more than 100 kW. By comparison, a traditional enterprise server rack draws about 5 to 15 kW, and the industry average is around 16 kW per rack.
Can I put GPU servers in my existing server room?
Often yes, for a small number of GPU servers, but check the power, UPS and cooling capacity first. A single high-density rack can create hot spots and overload circuits in a room designed for 5 kW racks. Supplemental in-row cooling or a rear door heat exchanger is usually the quickest fix.
Do AI data centres still need CRAC units?
Yes, in most cases. Even with direct-to-chip liquid cooling, part of the heat from memory, storage, power supplies and networking is released into the air and must be removed by precision air cooling such as CRAC, CRAH or in-row units. AI data centres are typically hybrid air and liquid cooled.
Related services and products
- AI and high-density GPU infrastructure
- Liquid cooling solutions
- CRAC units and precision cooling
- Server room cooling
- Server rack cooling guide
- CRAC installation and commissioning
|
Is your infrastructure ready for AI? Tell us what GPU hardware you are planning and we will assess the power and cooling you need. Indigi Power and Cooling is an Indigenous and Veteran-owned critical power and cooling specialist, Supply Nation and ICN Gateway registered, supporting sites Australia-wide. Contact Indigi Power and Cooling |
