“AI-ready” vs. purpose-built AI data centres: choose wisely

November 9 2026, by Macquarie Technology Group | Category: Data Centres
image-intellihands-service-1

A data centre may claim to be “AI-ready” after adding liquid cooling or increasing rack power, but those changes alone don’t prove it can support a large GPU deployment. Retrofitted facilities are still bound by the site’s architecture. Higher rack density matters little if the facility’s electrical systems were never built to support AI workloads. 

A purpose-built AI data centre is designed around AI demands from the outset, with power, cooling, and network requirements engineered alongside the facility. For a GPU-cloud operator, that distinction can mean the difference between a cluster that performs at scale and a deployment held back by the infrastructure it relies on. 

Infrastructure shortfalls put your GPU investment at risk 

GPU-cloud operators build their investment case around the revenue their planned cluster can deliver. If an AI-ready data centre can’t sustain that cluster at its intended scale, it caps what that investment can achieve. That leaves the operator exposed to: 

  • Already-purchased GPUs sitting idle because available power can’t support the full cluster, tying up capital in stranded compute 
  • GPUs running below their intended performance due to cooling-induced thermal throttling, making it harder to meet customer performance commitments
  • Restricted growth, with the operator unable to add the data centre capacity needed to expand existing customer contracts or take on new business

A provider may promise to resolve the facility’s constraints in time for later deployment phases, but that timeline is hard to guarantee. A seemingly straightforward rack-density upgrade can expose limits in cooling capacity, network bandwidth, and upstream power distribution, widening the scope of work and putting the agreed rollout schedule at risk. 

Those setbacks come at a cost. A 2026 Carnegie Endowment financial model estimated that a three-month delay could reduce an illustrative 100 MW US AI data centre’s life-cycle value by approximately US$178 million.

AI-ready vs. purpose-built: what buyers should verify 

Your GPU deployment needs high-density data centre infrastructure that can sustain your intended cluster at peak operating load. A provider’s headline capacity estimates and optimistic future plans don’t establish that capability. 

Use these checks to distinguish retrofitted colocation facilities from a data centre built for AI from the ground up. 

Finding a purpose-built data centre: What to look for

Criteria “AI-ready” facility Purpose-built for AI data centre What to verify
Power Adapts electrical systems originally built for lower rack loads; constrained by the site’s incoming power capacity Designs incoming power capacity and electrical distribution to handle 100+ kW racks across the full GPU cluster Request electrical designs and commissioning records confirming the power available to your full cluster at the required rack density
Cooling Retrofits rear-door heat exchangers (RDHx) or direct-to-chip cooling to cool GPUs at the rack level, but still relies on a legacy cooling system to carry heat away from the wider facility Integrates native liquid cooling and facility heat rejection as one system to handle sustained, high-density AI loads Compare cooling specifications against your full cluster’s heat output; verify that calculations cover sustained peak load across the cluster, rather than average utilisation
Structure and layout Capacity to accommodate heavy GPU racks is defined by inherited floor-load limits and site layout Engineers the load-bearing structure for heavy GPU racks and cooling equipment from the outset Confirm the proposed hall can support your rack weights and full cluster footprint; ask whether any further structural work is required

Why AI workloads break traditional data centre solutions

High-density AI puts legacy infrastructure under loads far beyond those of conventional enterprise racks. If the facility can’t sustain those loads, the GPUs can’t deliver the capacity the operator planned. 

Legacy electrical systems face a step change in rack demand

Traditional facilities were built to house CPU-based racks consuming 5–10 kW. AI can draw over 10 times that power. As an illustrative example, HPE’s NVIDIA GB200 NVL72 consumes 132 kW per rack.

Concentrating that much power in a single rack places exceptional demand on a facility built for traditional CPU loads. From the substation supplying the site to the rack power feeds and cooling pipework, every part of the facility must be engineered to handle AI-scale power and heat loads. The weakest component sets a hard ceiling on the scope of the deployment. 

AI workloads demand sustained power across the full cluster

Traditionally, colocation facilities could plan capacity for mixed enterprise workloads on the assumption that different workloads would peak at different times. 

Large-scale AI training workloads require GPUs to operate simultaneously, creating sustained demand across the cluster. Data centre infrastructure must be designed to accommodate these operating requirements reliably. 

These demand patterns require a different approach to capacity planning. AI colocation facilities must be able to support the GPU cluster at peak operating load without compromising performance. 

Floor-loading limits sink high-density AI deployments 

High-density AI concentrates enormous weight into a comparatively small physical footprint. The aforementioned HPE GB200 NVL72 weighs over 1.47 tonnes per rack – significantly more than conventional CPU racks – with a base of just 0.64 m² (600 mm x 1,068 mm).

A data hall can therefore have space for the full cluster but lack the floor-loading capacity to support its racks at the planned density. This forces the operator to sacrifice planned rack density or delay deployment until the provider completes structural floor-load reinforcements, both of which add cost and complexity to the rollout.

The architecture behind purpose-built AI colocation 

Rack-level upgrades can’t erase architectural limits. Presenting retrofitted capabilities alone as proof of full-cluster operational readiness is misleading marketing. 

Purpose-built data centres combine high-capacity power feeds and direct-to-chip cooling with a fully integrated facility built to support the rack density, electrical distribution, and cooling that AI demands. This lets the data hall be configured around the planned deployment without overcoming infrastructure designed for low-density workloads. 

Designed for 47 MW of IT capacity, Macquarie Data Centres’ IC3 Super West supports rack densities of 150 kW+ and a floor-loading capacity of 22 kPa, providing the infrastructure required for high-density AI deployments. 

Our new data halls incorporate redundant cooling infrastructure designed to support direct-to-chip and liquid-to-liquid cooling at scale.

Build your investment case on infrastructure you can trust 

The facility behind your GPUs dictates the commitment you can offer to customers. Before committing to a colocation facility, confirm the capacity allocated to your cluster, when it will be available, and any outstanding work required to deliver it. 

The strongest foundation for your GPU investment is an AI data centre purpose-built for AI deployments. Macquarie Data Centres is Australia’s most trusted data centre operator. With more than 25 years of experience, we support global hyperscalers, neocloud operators, and 42% of Australian Federal Government agencies with secure, resilient infrastructure for their critical workloads. 

Learn more about who we are or speak with our AI colocation specialists to request a facility blueprint and an AI colocation quote for your deployment.