Five questions to ask when choosing a data centre for AI

Choosing an AI-ready data centre starts with understanding how the facility will support your hardware, both now and as your requirements grow. Power capacity and cooling capability are useful starting points, but the detail comes from looking at where your racks will sit and how the infrastructure serving them is designed.

That makes the first conversation a practical one: what hardware are you installing, how much power will it draw, and how quickly could the deployment grow? With those answers, you and your provider can work through the five questions below and build a plan that balances today’s budget with tomorrow’s requirements.

1. Can the data centre support your rack density today, and as it grows?

An AI-ready data centre needs to deliver the power and cooling your hardware requires at the rack position you will occupy. The facility’s total megawatt capacity provides useful context, but it is the electrical and mechanical infrastructure serving your racks that determines what you can deploy there.

For example, you may only need 4kW per rack today, with a hardware roadmap that takes you to 20kW or 30kW in the next 12 to 18 months. Rack-scale AI platforms can require substantially more. NVIDIA’s reference architecture specifies up to 142kW for a full GB300 NVL72 rack, a liquid-cooled system integrating 72 Blackwell Ultra GPUs and 36 Grace CPUs.

As density increases, the engineering assessment needs to follow the full power and cooling path. That includes the A and B feeds, power distribution unit (PDU) or busway, upstream uninterruptible power supply (UPS) capacity, cooling, airflow, floor space and surrounding rack loads. Moving from a 6kW rack to a 30kW rack may be possible, but the provider needs to establish what that change involves at your location.

The useful question is how you get from the capacity you need now to the capacity you are likely to need next. Ask your provider to explain any upgrades, costs and lead times, including whether your equipment would need to move. Understanding that path gives you a basis for staged investment, without paying for all of the future capacity on day one.

2. How will the data centre cool your deployment?

Cooling needs to match your hardware specification and the heat load at the proposed rack position. A rack consuming 30kW produces roughly 30kW of heat, so a discussion about increasing power density naturally leads to how that heat will be removed.

Air cooling may suit your initial deployment. As density rises, improved containment, rear-door heat exchangers, direct-to-chip liquid cooling or a hybrid design may become appropriate. Some platforms, including the GB300 NVL72, are designed for liquid cooling from the outset. ASHRAE’s AI data centre guidance also discusses liquid cooling and thermally segmented zones for racks in the 50kW to 100kW-plus range.

If liquid cooling is part of your plan, ask the provider to walk you through how the system connects to the facility. Where will the coolant distribution unit (CDU) sit? How does the customer system interface with the facility loop, and what supply and return conditions are available? The discussion should also cover how much heat can be removed from that part of the hall and how the remaining air-cooled equipment will be supported.

At Macquarie Data Centres, we match cooling design to the location and customer requirement. Our Macquarie Park campus supports air, liquid and hybrid cooling options, while IC3 Super West’s fungible design allows data halls to be tailored for air-, liquid- and hybrid-cooled AI and cloud environments.

That flexibility matters as hardware evolves. A useful cooling plan explains how your current equipment will be supported and what would need to change for the next stage of your deployment.

3. What capacity is available around your first deployment?

Adjacent rack capacity can make expansion easier, particularly where equipment shares cabling, power distribution or cooling infrastructure. If you expect to add racks within six or twelve months, it is worth discussing the space around your first deployment before installation.

Consider a business starting with one rack and adding a second six months later. If the neighbouring position is already occupied, the additional equipment may need to sit elsewhere in the row, another pod or a different hall. That may be workable, but the layout needs to suit the cluster’s network architecture and equipment dependencies.

An early conversation gives your provider more options. You may be able to reserve adjacent space for an agreed period, or the provider may suggest a different starting position with more room to grow. In either case, check that the available power and cooling can support the additional racks as well as the first.

The reservation cost, holding period and any limits on guaranteed availability should be clear. With those details agreed, you can protect a realistic expansion plan while keeping the commitment proportionate to your business needs.

4. What happens if your deployment grows faster than expected?

A data centre’s available IT load, rack space and support for future hardware determine how far your deployment can grow within the facility. These factors deserve a place alongside the initial colocation price, especially for startups balancing limited capital with demand that can change quickly.

A successful product launch or new customer contract may bring forward plans for additional racks and higher-density hardware. If the facility can accommodate that growth, expansion becomes part of the existing deployment plan. If a move is needed, allow for migration engineering, cutover planning and potentially operating both locations during the transition.

Relocation also involves practical work: specialist transport, warranty and insurance checks, new connectivity, re-racking, testing and recommissioning. Security and access reviews may need to be repeated, and the migration plan needs to account for equipment faults or service interruption. At Macquarie Data Centres, our planning guidance reflects why customers prefer to expand around installed equipment where possible. Understanding the facility’s growth limits early helps you weigh the full cost of each option and invest in stages.

5. Who operates the data centre, and who supports your equipment?

An AI-ready data centre needs experienced people behind the infrastructure, with clear arrangements for on-site support, maintenance, incident response and physical security. For your team, that means knowing who can inspect a server, restart a device, check a cable or escort a vendor when help is needed, including outside business hours.

Those day-to-day arrangements sit alongside the facility’s resilience design. Ask how the A and B power paths are separated, how critical plant is maintained and what happens if a UPS, generator, chiller or pump is unavailable. Terms such as N+1, 2N and concurrent maintainability are useful when the provider explains what each means for your service during maintenance or an equipment failure.

At Macquarie Data Centres, we use redundant power and cooling architectures appropriate to each site, backed by 24×7 monitoring and operational support. Our Tier III facilities are designed for concurrent maintainability, allowing critical supporting infrastructure to be maintained without interrupting services to customer equipment.

Our dedicated facility operations and security teams provide 24×7 incident response. Through Intellihands, customers also have access to 24x7x365 remote-hands engineering support, from inspections and device restarts to equipment installation and vendor escorts. For international customers, having that local support can be particularly valuable, so it is worth agreeing the service scope, response arrangements and escalation process before deployment.

Ownership, jurisdiction and access controls also belong in this conversation, particularly for regulated organisations and international customers. Check how the facility and service align with your security and sovereignty requirements, and review the scope of the supporting evidence. Certifications, assessments and ratings serve different purposes; IRAP, for example, is an assessment program rather than a certification.

Our guide to AI data centre certification in Australia explains how to assess ISO 27001, IRAP, the Australian Government Hosting Certification Framework, Uptime Tier and NABERS evidence. Whichever standards matter to your organisation, confirm that the evidence covers the facility and service you are buying.

Bringing the five questions together.

An AI-ready data centre should give you confidence in the deployment you are planning and a clear understanding of how it can grow. That confidence comes from working through the power, cooling, space and support requirements with your provider, including the costs and engineering changes associated with expansion.

The resulting plan might include reserving adjacent racks, choosing a different hall or reviewing the power and cooling design ahead of a hardware upgrade. You do not need to predict every change in AI infrastructure to make a sound decision. You need to know what the facility can support today and how your provider will help you accommodate the changes you can reasonably anticipate.

To explore the options for your deployment, speak with the Macquarie Data Centres team or book a tour of our Sydney data centres. We can work through your hardware requirements, timing and growth plans to help you identify a suitable path forward.

Frequently asked questions.

How do I know if a data centre is AI-ready?

Ask the provider to demonstrate how the power, cooling, available space and operating model support your hardware at the proposed rack position. The assessment should cover your initial deployment and explain how the facility can accommodate foreseeable growth.

What rack density do AI workloads need?

Rack density varies by hardware and configuration. Some deployments may use 4kW per rack, while rack-scale platforms such as NVIDIA’s GB300 NVL72 can require up to 142kW for a full rack. Use your hardware specifications and roadmap to validate the power and cooling required at the proposed location.

Do AI data centres need liquid cooling?

Not all AI deployments require liquid cooling. Air cooling can suit some hardware, while higher-density systems may need direct-to-chip liquid cooling, rear-door heat exchangers or a hybrid design. The right approach depends on the equipment specification and heat load.

What is concurrent maintainability?

Concurrent maintainability allows critical supporting infrastructure, such as power and cooling plant, to be maintained without interrupting services to customer equipment. It helps operators carry out planned maintenance while customer workloads continue running.

Why does data centre jurisdiction matter for AI?

Facility location, ownership and operating arrangements help customers assess how a service aligns with their legal, security and sovereignty requirements. Regulated organisations and international customers should review these factors alongside access controls and the scope of relevant certifications and assessments.