AI-ready data centres in Australia: will your GPU deployment fit?

November 13 2026, by Macquarie Technology Group | Category: Data Centres
Beyond GPUs: What drives AI performance at scale

You can have the right GPUs and still find that the space you’ve shortlisted cannot support them. An AI-ready data centre needs to power, cool and physically accommodate your chosen hardware, with the network connections and capacity available when you need them. Start with the rack specifications and deployment date. That gives you something concrete to compare when providers describe their facilities as AI-ready.

The details matter here. A density figure might apply to one part of a hall, while a cooling option may still need work before it can support your installation. We’ll work through the questions that uncover those differences, so you can take a clear technical brief into conversations about AI infrastructure and get answers your engineering team can use.

 

What Makes a Data Centre AI-Ready?

 

An AI-ready data centre is a facility designed to support the demanding infrastructure requirements of artificial intelligence (AI) workloads, including GPU-intensive computing, machine learning and AI model training.

However, an AI-ready label won’t tell you whether a particular hall suits your cluster. One workload may run comfortably on air-cooled servers; another may need liquid-cooled cabinets with very different power and structural requirements. Ask the provider to assess your reference architecture, rather than treating the label as a specification.

Your test is straightforward: can the proposed space support the racks you intend to deploy, at the required density, on the required date? Get the answer in writing, including any work that must happen first. If a response stays at the level of ‘high density’ or ‘future ready’, ask for the numbers and conditions behind it.

 

Check the physical fit before committing

 

Power, cooling, structural capacity and connectivity need to work together in an AI Ready Data Centre. A hall with sufficient electrical capacity may still need changes to its cooling or delivery route. Work through each requirement against the same proposed space so a constraint doesn’t emerge after you’ve committed to it.

 

Rack power density

 

Rack requirements vary widely. For a concrete example, NVIDIA’s DGX GB rack-scale hardware guide describes rack power consumption of approximately 120kW. That is a hardware-specific figure, not a benchmark for every AI workload. Give providers your equipment list, expected sustained load and the manufacturer’s power requirements.

Ask for the sustained kW per rack available in the hall offered, and distinguish installed capacity from a future design allowance. Does the figure apply across the hall or only in high-density zones? Could your allocation constrain neighbouring racks or later expansion? Confirm the power paths, redundancy and usable capacity during maintenance or a failure.

 

Cooling method and cooling flexibility

 

There isn’t a universal rack density at which every deployment must switch to liquid cooling. The hardware, airflow, heat rejection and facility design all matter. With direct-to-chip cooling, coolant passes through cold plates on heat-generating components. Immersion places compatible equipment in a dielectric fluid. Each brings different servicing and infrastructure requirements, so start with the hardware manufacturer’s supported approach.

Then consider the next refresh. Ask which cooling options the proposed hall can support, what a change would involve and whether you would need to move equipment. Establish who supplies and maintains the coolant distribution units, how leaks are managed and where the provider’s responsibility ends. MDC’s cooling offering includes air, liquid and hybrid environments; the configuration still needs to be matched to your deployment.

 

Floor loading and structural capacity

 

A rack can fit the floor plan and still exceed a structural limit. Check the fully populated cabinet weight, including coolant where relevant, and have the floor and load distribution assessed for that configuration. An evenly distributed floor-load rating does not, by itself, confirm that the floor can support concentrated loads at cabinet feet or wheels.

Ask for the hall’s distributed load rating in kPa and the permitted point loads, with an engineering assessment where needed. Check the entire delivery route too: goods lifts, doors, thresholds and temporary staging areas. Finding out that a loaded cabinet cannot pass through the lift is an avoidable problem.

 

Network fabric and cross connects

 

For training clusters, the layout needs to support the bandwidth and latency your interconnect design requires between nodes. Contiguous space can help, but it is the complete fabric design and cable constraints that determine the fit. Inference requirements also depend on the application, including latency to users, data sources and other systems. Test any proposal to split the cluster across halls against those requirements.

For connections beyond the cluster, confirm the available carriers, cross-connect routes and any direct cloud access your design needs. Review network connectivity alongside installation charges, recurring fees and provisioning times. These affect both the initial move and the cost of changing the environment later.

 

What to specify and what evidence to request when deploying an AI Data Centre

 

Use this table to turn the discussion into a technical request. Ask every provider to answer for the specific hall and deployment phase being offered, and record any conditions or remaining design work alongside the response.

 

Criterion Why it matters Evidence to request
Maximum kW per rack Decides whether your reference architecture physically fits Written figure for the specific hall, with the redundancy configuration behind it
Cooling modes available Cooling must match the hardware and validated thermal design; there is no universal changeover threshold Supported modes per hall, and the responsibility boundary for coolant equipment
Cooling flexibility A hardware refresh may change cooling requirements, with implications for cost and disruption Supported cooling options and the cost, timing and disruption of a future change
Distributed and point-load capacity Dense cabinets may exceed existing load ratings; engineering assessment must establish fit and any upgrade feasibility Hall load ratings, cabinet load distribution, engineering assessment and delivery-route capacity
Contiguous space Layout and cable constraints must support the cluster fabric design Floor plan showing contiguous area and adjacent expansion capacity
Connectivity Carrier neutrality and on-ramp availability shape both cost and long-term lock-in Cross connect commercial terms, provisioning lead times, and the cloud on-ramp list
Ready-for-service date Delivery timing must fit the project window and its contingency plan Contracted date, the remedy if it slips, and whether grid power is secured or pending

 

Check when the capacity can actually be delivered

 

A technically suitable hall still needs to be available within your deployment window. Ask what is installed, what is committed to other customers and what depends on construction, utility work or commissioning. Put the ready-for-service date and the response to a delay into the commercial discussion early.

Secured power and power still awaiting approval represent different delivery dependencies; neither, by itself, proves that a hall is ready. For example, MDC publishes a 47MW end-state design for IC3 Super West, with end-state power secured and support for air, liquid and hybrid cooling and 150kW+ racks. Confirm the current phase, hall-level allocation and ready-for-service date for your project rather than reading those design figures as available capacity.

 

Sovereignty, security and compliance for Australian AI workloads

The data matters as much as the hardware. Training datasets, model weights and inference logs can contain personal, commercially sensitive or classified information. Map what your deployment will hold and which obligations apply before deciding where it should sit. Our data sovereignty guide provides further context for assessing location, access and control.

Establish the locations of data and backups, the jurisdictions relevant to the provider and its parent, key custody and who can access the equipment. For government workloads, check the applicable Hosting Certification Framework requirements and the scope of the certified service. The HCF website currently notes a pause, effective 3 November 2025, on new registrations and supplementary assessments during reform; existing certifications are not impacted. Verify the position when procuring, especially if your plan depends on a new certification.

 

Technical requirements checklist

 

Use these questions with your engineering team before shortlisting colocation options. Record unanswered items and agree who will resolve them before the design or contract is signed.

• Maximum sustained kW per rack available today, in the specific hall offered
• Whether that density applies hall-wide or only in designated zones
• Power redundancy configuration serving high-density cabinets
• Cooling modes supported in that hall: air, direct-to-chip, immersion or hybrid
• Whether cooling mode can be changed later without vacating the space
• Ownership and maintenance boundary for coolant distribution equipment
• Distributed floor loading in kPa, permitted point loads and assessment of the proposed cabinet layout
• Delivery path capacity, including goods lift and door dimensions
• Contiguous space available for a single cluster, and expansion adjacency
• Carrier neutrality, cross connect terms and provisioning lead time
• Direct cloud on-ramps available on site
• Contracted ready-for-service date and the remedy if it slips
• Whether grid power for the end state is secured or still subject to approval
• Certification evidence, with the certificate naming the specific facility
• Clearance level and employment status of on-site engineering staff

Before you commit, bring the equipment supplier, your infrastructure team and the facility engineers onto the same design. A written agreement on the rack configuration, service boundaries and delivery dependencies is far more useful than another broad assurance that the building is AI-ready.

Looking for a data centre partner that can support your infrastructure today and as your requirements evolve? Macquarie Data Centres provides secure, resilient and scalable data centre solutions for hyperscale, enterprise, government and cloud environments.


Contact us to discuss your requirements, learn more about our team, or request a tour to see our facilities firsthand.

 

Frequently asked questions

 

What makes a data centre AI-ready?

For your deployment, AI readiness means the facility can meet the specified power, cooling, structural and network requirements on the date you need them. Validate each against the hardware configuration and proposed hall. Published design figures are a starting point; they do not confirm an allocation or delivery date.

How much power does an AI rack need?

It depends on the equipment and configuration. NVIDIA’s DGX GB rack-scale guide provides one example at approximately 120kW per rack. Use your manufacturer’s requirements and expected sustained load to size the deployment, including redundancy and future expansion, rather than applying a generic AI rack figure.

When does liquid cooling become necessary?

When the chosen hardware or validated thermal design requires it. There is no single threshold that applies to every rack and facility. Check the supported cooling method, the heat that still needs to be removed by air, the coolant requirements and the provider/customer maintenance boundary.

What is the difference between direct-to-chip and immersion cooling?

Direct-to-chip cooling uses cold plates on components such as processors; other components may still need air cooling. Immersion submerges compatible equipment in dielectric fluid. Compare hardware support, coolant compatibility, servicing, structural loads and heat rejection for the specific configuration rather than assuming one method is always easier or denser.

Why does floor loading matter for AI deployments?

Dense cabinets and coolant can create loads that an existing floor or delivery route was not designed to carry. Confirm distributed and point-load limits, cabinet load distribution and lift capacity before choosing the space. Any strengthening or layout changes need an engineering assessment and a place in the delivery plan.

Should AI workloads stay onshore in Australia?

That depends on the data involved and the obligations attached to it. Training data, model weights and inference logs often contain personal, commercially sensitive or classified material, which brings residency and jurisdiction into scope. Establish where data physically sits, which jurisdiction governs the operating entity and its parent, and who can physically access the hardware.