Australia's first large AI data centres are now under construction or contracted. Firmus has GPU capacity under construction in Tasmania and Melbourne as part of Project Southgate. NEXTDC has approval for M4 in Port Melbourne, designed for 1,000 kW racks. OpenAI has signed a memorandum of understanding for NEXTDC's S7 campus in Western Sydney. ABC has reported an Anthropic lease at a proposed 2.16 GW campus in regional Queensland.
This page explains what makes an AI data centre different, lists the Australian projects and sets out the constraints: power, transmission, equipment, cooling, approvals and labour.
What Is an AI Data Centre?
An AI data centre is a data centre designed around large clusters of GPUs or other AI accelerators, used to train and run machine learning models. The difference from a conventional data centre is density. The same floor area draws far more power and produces far more heat. That changes the electrical design, the cooling system, the structure and the network.
Operators use different names. NEXTDC and Firmus both use "AI factory". NEXTDC describes S6 in Artarmon as designed for AI factories, and Firmus calls its Project Southgate sites AI factories. The term has no formal definition. It generally means a facility, or part of one, dedicated to GPU compute at high density.
How AI Data Centres Differ
| Requirement | Traditional data centre | AI data centre |
|---|---|---|
| Compute | CPU-heavy general-purpose servers | GPU or accelerator clusters, often thousands of devices linked together |
| Rack density | Lower; designed around air cooling | Much higher. NEXTDC's M4 is designed for up to 1,000 kW per rack |
| Power density | Lower per m² of data hall | Much higher per m², which needs larger substations and heavier distribution |
| Cooling | Air cooling with CRAH/CRAC and containment | Direct-to-chip liquid cooling increasingly standard; immersion in some designs |
| Network | Standard enterprise and cloud networking | Very high-bandwidth, low-latency fabrics between GPUs |
| Structure | Standard raised or slab floors | Heavier racks and liquid distribution affect floor loading and layout |
| Location sensitivity | Close to users and interconnection points | Training can be further from users; inference still benefits from proximity |
Australian AI Data Centre Projects
| Project | Location | Developer / operator | AI-specific details | Stage |
|---|---|---|---|---|
| Firmus Project Southgate, Tasmania | Tasmania | Firmus | ~36,800 NVIDIA GB300 GPUs when complete, per w.media | Under construction; completion targeted late 2026 |
| Firmus Southgate Melbourne | Hosted in CDC Melbourne | Firmus with CDC | ~18,400 GB300 GPUs under a multi-year hyperscale customer contract | Deployment |
| NEXTDC S7 | Eastern Creek, NSW | NEXTDC | Non-binding MoU with OpenAI for an AI campus and GPU supercluster; closed-loop liquid cooling | Prepare EIS; first phase targeted H2 2027 |
| NEXTDC M4 | Port Melbourne, VIC | NEXTDC | 162 MW; up to 1,000 kW racks; direct-to-chip liquid cooling | Approved Jan 2026 |
| NEXTDC S6 | Artarmon, NSW | NEXTDC | 13.5 MW, designed for AI factories | Operating |
| Macquarie IC3 Super West | Macquarie Park, NSW | Macquarie Data Centres | 47 MW, high-density power and liquid cooling | Opening slated Sep 2026 |
| CDC Perth | Maddington, WA | CDC | Liquid cooling; ultra-high-density compute | Stage 1 targeted 2026 |
| Western Downs Digital Park | Near Dalby, QLD | Zerra DC | 2.16 GW peak; reported Anthropic lease for Building One, subject to FIRB approval | DA lodged Aug 2026 |
| Project Meridien | Kimberley, WA | Gingerah Energy | 240 MW IT initially; described as an AI factory for model training | Environmental studies |
Project Southgate
Firmus announced in October 2025 a partnership with CDC Data Centres to scale Project Southgate to 1.6 GW of AI factories by 2028, across Tasmania, Victoria, NSW and the ACT. It stated an initial investment of A$4.5 billion, scaling to A$73.3 billion. In February 2026, Firmus secured a US$10 billion private debt facility led by funds managed by Blackstone (DCD). Firmus reports a PUE of 1.10 for its design.
OpenAI and NEXTDC
On 5 December 2025, OpenAI and NEXTDC signed a non-binding MoU to develop an AI campus of about A$7 billion at S7, as part of OpenAI's "OpenAI for Australia" programme. The first phase is targeted for the second half of 2027, subject to approvals. The MoU does not commit either party to a lease.
GPU Infrastructure
GPU clusters for training are built as tightly coupled systems. Racks of GPU servers are linked by high-bandwidth networks so that thousands of accelerators can work on one model. Current designs referenced in Australian announcements include NVIDIA's GB300 (Firmus) and NVIDIA's Blackwell and Rubin Ultra architectures (NEXTDC M4).
Practical consequences for the building:
- Rack weight and layout. Liquid-cooled GPU racks are heavier than conventional racks and need coolant distribution to every rack position.
- Power distribution. Higher per-rack loads need higher-capacity busways and larger power distribution units, plus careful redundancy design.
- Network cabling. GPU fabrics need dense optical cabling within and between rows.
- Refresh cycles. GPU generations change quickly, so halls designed today need headroom for higher densities later.
More: Data centre networking
Power Requirements
AI is one of the main reasons forecasts of data centre electricity use keep rising. AEMO's 2026 ESOO forecasts data centre consumption in the National Electricity Market rising from about 5 TWh in 2025-26 to 34 TWh by 2035-36, or about 13% of grid-supplied electricity.
Individual AI proposals are large enough to be significant for a whole state. ABC estimates that the Western Downs Digital Park could draw 43 to 47 GWh a day at full build, roughly a quarter of Queensland's average daily consumption.
Operators are responding in three ways:
- Contracting new renewable generation. Amazon signed nine PPAs in April 2026 totalling 430 MW, eight with co-located batteries (Energy-Storage.news). Firmus says Southgate aims to support up to 5.1 GW of new renewable generation.
- Locating where power is available. This includes Tasmania (hydro), regional Queensland and the Kimberley.
- Adding on-site generation and storage to reduce reliance on firm grid capacity.
More: Data centre power infrastructure
Liquid Cooling
At GPU densities, air cooling alone becomes impractical, and direct-to-chip liquid cooling is becoming the default for new AI halls. Coolant is circulated to cold plates on the processors through coolant distribution units (CDUs), and heat is rejected through chillers or dry coolers.
Australian examples:
- NEXTDC M4. Direct-to-chip, using non-potable recycled water.
- NEXTDC S7. Closed-loop, high-density liquid cooling "without relying on potable water" (per the OpenAI MoU announcement).
- CDC Perth. Liquid cooling with a closed-loop system, and zero water consumption for primary cooling.
- Western Downs Digital Park. Reported as 100% air-cooled, which is unusual for an AI-focused campus.
More: Data centre liquid cooling · Data centre cooling in Australia
Hyperscale Infrastructure
The large cloud providers are also expanding their AI capacity in Australia:
- AWS: A$20 billion from 2025 to 2029 to expand, operate and maintain data centre infrastructure in Sydney and Melbourne (Amazon).
- Microsoft: A$25 billion to the end of 2029, announced 23 April 2026, including expanding its Azure footprint by more than 140% (Microsoft).
Hyperscalers both build their own facilities and lease large blocks from operators such as AirTrunk, CDC and NEXTDC.
Australia's AI Infrastructure Opportunity
Advantages
- Political stability and established legal frameworks, which matter for sovereign and government AI workloads.
- Land and renewable energy resources, particularly in regional areas.
- Strong demand from government, defence, financial services and research for onshore AI compute.
- An established data centre sector with experienced operators and contractors.
Constraints
- Grid connection capacity in the main metro markets is largely committed (Transgrid).
- Distance from the major AI chip supply chains and global model developers.
- Competition for skilled trades with housing, energy and defence.
- Policy uncertainty. Federal legislation on a national data centre framework is expected in early 2027, and state frameworks (NSW, August 2026) are still new.
The Australian Government's Expectations of data centres and AI infrastructure developers, released 23 March 2026, followed a commitment in the December 2025 National AI Plan. They set non-binding expectations on the national interest, energy, water, skills and research access.
Challenges
| Challenge | Evidence |
|---|---|
| Power | AEMO: 225 projects seeking 67 GW of connection capacity, against a forecast of 34 TWh of annual demand by 2035-36 |
| Transmission | Transgrid: more than 10 GW of enquiries within 12 km of Sydney West; 330 kV network increasingly constrained |
| Equipment | In the US market, power transformer lead times averaged 128 weeks and switchgear 44 weeks in Q2 2025 (Wood Mackenzie). Australian buyers compete in the same global supply chain |
| Construction | Multiple gigawatt campuses competing for the same contractors in Western Sydney and Melbourne's west |
| Approvals | NSW SSD applications drawing objections on generators, noise and water; a 75-day pathway for projects meeting the NSW framework |
| Cooling | Water availability; Sydney Water estimates data centres could add up to 20 to 25% to Greater Sydney's water demand by 2035 |
| Skilled labour | Powering Skills Organisation: about 72,000 more electricians and related trades needed nationally by 2030, including about 13,000 attributed to data centres (ABC) |
AI Data Centre Suppliers
AI builds increase demand for particular supplier categories:
- Liquid cooling: CDUs, cold plates, coolant piping
- Electrical: high-capacity busway, HV substations, switchgear
- Networking: optical and GPU fabric cabling
- Energy and BESS: PPAs, on-site storage
- Recruitment: electricians, commissioning engineers, critical facilities staff
Related: Data centre construction · All Australian data centre projects
Frequently Asked Questions
What is the largest AI data centre in Australia?
Among operating or contracted sites, Firmus's Project Southgate, which is under construction in Tasmania and Melbourne, has the largest disclosed GPU deployments. The largest proposal is Western Downs Digital Park in Queensland at 2.16 GW peak, which is not yet approved.
Is OpenAI building a data centre in Australia?
OpenAI has signed a non-binding MoU with NEXTDC to use capacity at NEXTDC's planned S7 campus in Eastern Creek. NEXTDC is the developer. S7 is at the Prepare EIS stage in the NSW planning system.
Why do AI data centres need liquid cooling?
GPU racks concentrate far more heat in a small space than air can remove efficiently. Liquid carries heat much more effectively, so direct-to-chip cooling lets operators run racks at much higher densities.
Selling Into Australia's AI Infrastructure Build?
AI projects involve new specifications, new buyers and short timelines. AISB helps suppliers of power, cooling, networking and construction services reach the developers and contractors behind Australian AI projects.
Sources & Verification
- Firmus: Southgate expansion, Oct 2025. Link
- DCD: Firmus US$10 billion debt facility, Feb 2026. Link
- w.media: Firmus Melbourne contract. Link. OpenAI and NEXTDC MoU, Dec 2025. Link. AEMO 2026 ESOO. Link
- NEXTDC: M4 approval, 19 Jan 2026. Link. S6 Sydney. Link
- ABC News: Western Downs Digital Park, Sep 2026. Link. Project Meridien, Apr 2026. Link
- Department of Industry, Science and Resources: Expectations of data centres and AI infrastructure developers, 23 Mar 2026. Link
- Amazon: AU$20 billion investment. Link
- Microsoft: A$25 billion investment, 23 Apr 2026. Link
See something that needs updating? Submit a correction.
