Schneider Electric: Liquid Cooling Can Stretch Capped Grid Links for Australian AI
Schneider Electric argues that cooling architecture will decide how much of a fixed grid connection actually reaches compute in Australian AI data centres. In a blog post by Mark Deguara, the company links the point to tightening grid capacity in New South Wales and Victoria.
- AEMO tracked 9.0 GW of maximum data-centre connection capacity in the formal process at the end of Q2 2026, 52% in NSW and 31% in Victoria.
- Transgrid says about 1.5 GW in Western Sydney is covered by signed agreements, with limited room for more large connections without augmentation.
- In Schneider Electric's 100 MW example, lowering PUE from 1.55 to 1.20 lifts theoretical IT capacity from about 64.5 MW to 83.3 MW.
- RDHx, HDUs and liquid-to-liquid CDUs suit different retrofit, greenfield, colocation and neocloud cases.
- Freed mechanical capacity only helps if the electrical design can deliver it to compute.
Grid capacity sets the ceiling
The post leans on figures from the Australian Energy Market Operator (AEMO). At the end of Q2 2026, AEMO was tracking 9.0 GW of maximum data-centre connection capacity in the formal transmission connection process, across 17 projects. New South Wales accounted for about 4.7 GW and Victoria for about 2.8 GW, or 52% and 31% of the total respectively.
Most of these projects were still at the application stage, so the numbers describe the formal development pipeline rather than committed operating load. In Western Sydney, Transgrid has reported more than 20 GW of data-centre connection enquiries, while about 1.5 GW is covered by signed agreements. According to the post, further large connections there will need additional transmission capacity. AEMO Victorian Planning has also received more than 18 GW of large-load enquiries, largely tied to data centres in metropolitan Melbourne.
Schneider Electric concludes that a secured connection should be treated as a finite resource in facility design.
Why PUE matters under a fixed budget
Power Usage Effectiveness (PUE) compares total facility power with the power used by IT equipment. With a fixed connection, dividing the connection capacity by PUE gives the maximum theoretical power available to IT.
The company gives a simplified example of a 100 MW connection. At a PUE of 1.55, theoretical IT capacity is about 64.5 MW. At 1.20, it rises to 83.3 MW, a difference of 18.8 MW within the same electrical envelope. The post notes that real facilities also have to account for redundancy, utilisation, maintenance states and power quality.
Cooling affects both sides of the ledger. Air movement and mechanical systems consume facility energy, while onboard server fans consume IT energy. According to Schneider Electric, liquid cooling can reduce both, with results shaped by coolant temperature, heat-rejection design, utilisation and the share of equipment that is liquid cooled. It proposes reliable compute per secured megawatt as a capacity measure alongside PUE.
Three cooling architectures, different trade-offs
The post compares three approaches.
- Rear-door heat exchangers (RDHx): they capture hot exhaust air at the rack and pass the heat to liquid. They can cut room-cooling demand and support denser racks, particularly in existing colocation halls. Server fans stay in use.
- Direct-to-chip cooling with Heat Dissipation Units (HDUs): cold plates sit on components such as CPUs and GPUs, and the HDU moves heat from the liquid loop into data-hall air where facility water is unavailable. This suits retrofits, colocation and neocloud deployments, but the existing air-cooling system still carries part of the load.
- Direct-to-chip cooling with liquid-to-liquid CDUs: a Coolant Distribution Unit transfers heat into the facility-water system while keeping the loops hydraulically separate. More heat bypasses room-air infrastructure, and dedicated warm-water systems can further reduce chiller use in greenfield campuses and major expansions where conditions permit.
Chillers and heat rejection remain part of the design, since captured heat must ultimately leave the facility. The company adds that new rack-scale AI systems are raising thermal loads, which makes coordination between chip cooling, coolant distribution and heat rejection more important.
What it means for different operators
Schneider Electric says priorities differ by segment. Hyperscalers are advised to assess IT MW per secured grid MW, cooling-system power, usable GPU capacity and performance per watt. Colocation operators face mixed tenant hardware and staggered adoption, so liquid-to-air HDUs can introduce direct-to-chip equipment before facility water reaches a whole hall, while liquid-to-liquid CDUs can serve larger deployments. Clear specifications for supply temperature, flow, pressure, water quality, redundancy and ownership of the secondary loop are needed.
For neocloud providers, rapid GPU deployment and high utilisation can favour modular direct-to-chip systems and HDUs where facility water is absent. Prefabricated modular IT pods can support repeatable expansion, and the post says cooling overhead weighs more on lifetime power consumption as scale grows.
The central design point is that cooling and electrical planning must go together. Cutting mechanical load only creates usable IT headroom if the electrical architecture, including transformer allocation, can deliver that released capacity to the compute load. For specifiers and developers, the post recommends early modelling of normal, redundant and maintenance states, residual air load, CDU power paths, temperatures, pump power and heat rejection.
Schneider Electric: AI data centres in Australia: How liquid cooling can stretch a capped grid connection
