At a glance
- Author: Mark Deguara, Technology Innovation Leader Building an AI data centre in Australia increasingly requires close attention to grid capacity
- At the end of Q2 2026, AEMO was tracking 9.0 GW of maximum data-centre connection capacity in the formal transmission connection process, with 52% in..
# AI Data Centres in Australia: Liquid Cooling as a Grid-Connection Strategy
Grid capacity is emerging as a defining constraint for AI data centre development in Australia. According to AEMO data from the end of Q2 2026, 17 data-centre projects representing 9.0 GW of maximum connection capacity were in the formal transmission connection process. New South Wales accounted for approximately 4.7 GW of that pipeline, while Victoria represented about 2.8 GW. Most of these projects remained at the application stage, meaning the figures describe the formal development pipeline rather than committed operating load.
In Western Sydney, constraints are already acute. Transgrid has reported more than 20 GW of data-centre connection enquiries, with about 1.5 GW covered by signed agreements. The network operator has indicated that existing capacity has largely been allocated, leaving limited room for additional large connections without augmentation. AEMO Victorian Planning has also received more than 18 GW of large-load enquiries, largely associated with metropolitan Melbourne data centres.
For operators holding a secured connection, that allocation is now a finite resource. Cooling efficiency directly determines how much of that fixed capacity can be delivered to compute.
Power Usage Effectiveness as a Capacity Metric
Power Usage Effectiveness (PUE) measures total facility power consumption against IT equipment power. With a capped grid connection, dividing connection capacity by PUE yields the theoretical maximum power available to IT.
Consider a simplified 100 MW connection. At a PUE of 1.55, theoretical IT capacity is approximately 64.5 MW. At a PUE of 1.20, that figure rises to 83.3 MW—an 18.8 MW difference within the same electrical envelope. Actual facilities must also account for redundancy, utilisation, maintenance states, power quality, and other operating constraints.
Cooling influences both facility and server power. Air movement and mechanical systems consume facility energy, while onboard server fans consume IT energy. Liquid cooling can reduce both, though results depend on coolant temperature, heat-rejection design, utilisation, and the share of equipment being liquid cooled. Reliable compute per secured megawatt is emerging as a useful capacity measure alongside PUE.
Liquid-Cooling Architectures and Trade-offs
Liquid cooling encompasses several architectures, each with distinct implications for retrofit and greenfield deployments in Australia.
Rear-door heat exchangers (RDHx) capture hot exhaust air at the rack and transfer heat to liquid. This approach can reduce room-cooling demand and support higher-density racks, particularly in existing colocation halls and mixed-density environments. Internal server fans remain in use, and water temperature along with downstream cooling plant influences overall efficiency.
Direct-to-chip (DTC) cooling uses cold plates on high-power components such as CPUs and GPUs. A Heat Dissipation Unit (HDU) transfers heat from the liquid loop into data-hall air where facility water is unavailable. This architecture suits retrofit facilities, colocation environments, and neocloud deployments that cannot extend a facility-water loop to each deployment. Because heat returns to the room, the existing air-cooling system continues to carry part of the facility load.
A liquid-to-liquid Coolant Distribution Unit (CDU) transfers heat from the technology cooling loop into the facility-water system while keeping the two loops hydraulically separated. More heat can bypass room-air infrastructure, reducing demand on fans and air handlers. Supply and return temperatures, pump energy, heat exchangers, and chilled-water conditions all influence performance.
Dedicated warm-water systems can further reduce chiller use in greenfield campuses and major expansions where climate and design conditions permit. Chillers and heat rejection remain integral because captured heat ultimately must leave the facility.
New rack-scale AI systems are increasing the thermal loads these architectures must accommodate, making coordination between chip cooling, coolant distribution, and facility heat rejection increasingly important.
Market-Specific Implications
Hyperscalers and internet giants operating campus-scale deployments benefit from evaluating efficiency across the complete electrical and thermal system. Useful measures include IT MW per secured grid MW, cooling-system power, usable GPU capacity, and performance per watt. Liquid-to-liquid architectures and warmer coolant temperatures can reduce mechanical cooling demand, but electrical distribution must have enough flexibility to make released capacity available to IT.
Colocation operators face mixed tenant hardware and staggered adoption. Liquid-to-air HDUs can introduce direct-to-chip equipment before facility-water distribution reaches an entire hall, while liquid-to-liquid CDUs can serve larger deployments with lower room-cooling demand. Service delivery depends on clear specifications for supply temperature, flow, pressure, water quality, redundancy, and ownership of the secondary loop.
Neocloud providers deploying GPUs rapidly at high utilisation may favour modular direct-to-chip systems and HDUs where facility water is unavailable. Prefabricated modular IT pods can support repeatable capacity expansion. As deployments scale, cooling overhead has a larger effect on lifetime power consumption. Liquid-to-liquid systems can reduce facility demand compared with architectures that return captured heat to the data hall.
Electrical and Thermal Coordination
Cooling and electrical design must be planned together. Reducing mechanical load only creates usable IT headroom when the electrical architecture can deliver that released capacity to the compute load. Early-stage design modelling should cover normal, redundant, and maintenance states, residual air load, CDU power paths, supply and return temperatures, pump power, heat rejection, and the expected mix of air- and liquid-cooled IT. Transformer allocation deserves the same attention.
Australia’s data centre pipeline includes projects such as NEXTDC’s S7 campus in Eastern Creek, Sydney, currently in planning for high-density AI and cloud infrastructure, and Macquarie Data Centres’ 47 MW IC3 Super West facility for AI and cloud workloads. These sit within a wider pipeline of hyperscale and high-density development across NSW, Victoria, and other markets.
With grid access tightening, cooling choices influence how much of a constrained grid allocation can reach IT. Liquid cooling can reduce power used for fans, air movement, and mechanical cooling, with the final result determined by the path from direct-to-chip heat capture through CDUs or HDUs to chillers and heat rejection. Rack density remains an important design input, while reliable compute per secured grid megawatt gives operators a broader view of capacity efficiency.
MegaWatt Intelligence Desk
Hyperscale & Grid Analyst. `More stories by $MegaWatt Intelligence Desk →`
