Traditional data centres vs AI data centres
Most enterprise data centres were built for CPU-driven workloads: databases, web applications, virtualised compute. Rack densities in these environments typically sit between 5kW and 10kW. Air-based cooling using cold-aisle and hot-aisle containment works at those densities. Network fabrics handle bursty, transactional traffic with latency tolerances designed around human-facing application response times.
An AI data centre operates on different assumptions at every layer.
A single NVIDIA H100 GPU draws up to 700W at full load. An 8-GPU server exceeds 5kW on its own. Rack a full AI compute node and you are looking at 30kW to over 100kW per rack in high-density GPU configurations. The power infrastructure, cooling capacity, and structural floor load tolerances required to sustain those densities are simply not present in most traditional colocation environments.
There is also a workload continuity requirement that does not apply the same way to general compute. AI training runs are long, parallel, and sensitive to interruption. A distributed training job that loses power or overheats mid-run may need to restart from the last checkpoint, losing hours of compute time. Reliability requirements are tighter, and the cost of failure is measurably higher.
An AI data centre is not a traditional facility with GPUs installed. It is a different type of infrastructure designed from the ground up for different physics.
Power density and cooling for GPU workloads
Power is the primary engineering constraint in an artificial intelligence data centre.
Standard enterprise racks require a 10A or 20A power feed. A GPU-dense rack needs 60A or more. Facilities supporting AI workloads need high-capacity power infrastructure from site level down: dedicated transformers, redundant uninterruptible power supplies rated for sustained high-draw loads, and sufficient incoming utility capacity to handle the full rack count without derating.
Cooling follows directly from power density. Air cooling is effective up to roughly 15 to 20kW per rack. Beyond that, it becomes thermally inefficient. The volume of airflow required to remove heat from a 30kW or 60kW rack creates velocity and acoustic problems at scale, and the cooling overhead drives up operating costs significantly.
Modern AI-ready facilities use one of two liquid cooling approaches. Direct liquid cooling (DLC) routes coolant through cold plates attached directly to processors and memory modules inside each server. Full immersion cooling submerges entire server boards in dielectric fluid. Both remove heat far more efficiently than air. Both require engineered facility infrastructure: coolant distribution units, precision leak detection, compatible server hardware, and maintenance procedures that do not place adjacent racks at risk during servicing.
Power Usage Effectiveness (PUE) is the standard efficiency metric. A PUE of 1.2 means 20% of total energy goes to facility overhead such as cooling and lighting, with the remaining 80% reaching compute. A PUE of 2.0 doubles the energy cost for the same compute output. For sustained GPU workloads running continuously across large clusters, that ratio has a direct and compounding effect on operating cost.
Network and latency considerations
AI training and inference have different network requirements. Neither maps well to standard enterprise network assumptions.
Distributed training is inter-GPU intensive. As training runs across multiple GPUs and multiple nodes, the GPUs synchronise gradient updates and model weights continuously. The fabric connecting them needs to deliver high bandwidth and low latency simultaneously. InfiniBand at 400Gb/s, or high-speed Ethernet fabrics at 100GbE and above, are the current standard for serious training infrastructure. Even brief latency spikes in the interconnect can cause training jobs to stall, slow, or produce unstable convergence.
Inference is different. Once a model is deployed, the critical network path shifts from the GPU-to-GPU interconnect to the path between the model server and the client application. Consistent, low-latency responses matter to the end user and to the application depending on them. The physical distance between the GPU and the user is a real variable, not an abstraction.
Australian AI workloads served from offshore infrastructure cross oceanic cables and accumulate round-trip latency with no equivalent in a domestic facility. For latency-sensitive inference applications, real-time decision support, interactive AI tools, and financial services processing, that additional latency has a measurable business impact.
Amaze operates facilities in Sydney and Melbourne with direct connectivity to Australian internet exchange infrastructure. Training interconnects and inference response paths remain on Australian soil.
Compliance and sovereignty in AI data centres
An AI data centre is a compliance decision as much as a technical one.
When AI workloads handle sensitive data, customer records, health information, financial transactions, government data, the legal jurisdiction governing that data matters under Australian law. The Privacy Act 1988 and its Australian Privacy Principles establish obligations around data handling and offshore transfer. APRA CPS 234 sets information security requirements for financial services entities and applies to outsourced hosting environments. Defence and government workloads carry additional requirements under the Information Security Manual (ISM) and Essential Eight controls.
The legal exposure from offshore infrastructure is not resolved by pointing to a Sydney-located server rack. The US CLOUD Act allows US authorities to compel US-incorporated companies to produce data held anywhere in their global systems, regardless of where servers physically sit. If the facility operator or its parent company is incorporated in the United States, that legal pathway exists.
Sovereignty in an AI data centre means the operator is an Australian entity, operating under Australian law, with no foreign jurisdiction’s ability to compel access to customer data. Built here, run here, governed here.
Amaze is ISO 27001 certified and operates as an Australian company. No CLOUD Act exposure. Data residency is in Australia. For regulated industries, those are baseline requirements, not features to consider after shortlisting.
What to look for when selecting an AI-ready facility
Use these questions when evaluating a facility for AI workloads.
Power infrastructure. What is the maximum power density per rack? Look for at least 30kW, and 60kW or higher for dense GPU configurations. Is the infrastructure in place now, or is it a planned upgrade? What redundancy level applies: 2N, N+1?
Cooling systems. Does the facility support direct liquid cooling or immersion cooling? What is the rated PUE? Anything above 1.5 is a cost concern for sustained GPU workloads. What monitoring and leak detection is in place?
Network. What port speeds and interconnect fabric are available? Is there direct connectivity to Australian internet exchange points? What are the measured latency benchmarks to major Australian population centres?
Sovereignty and compliance. Is the facility operated by an Australian entity under Australian law? What certifications apply: ISO 27001, Tier 3 or above, SCEC endorsement for government workloads? Does the parent company carry any foreign-jurisdiction legal exposure?
Operational commitments. What are the SLA terms for power and cooling uptime? What on-site support is available, and what are the response time commitments? How is physical access managed, logged, and audited?
Selecting an AI data centre is a long-term infrastructure commitment. The facility determines your compliance posture, your operating cost structure, and your ability to scale AI workloads without replatforming.
Related reading: How sovereign data centres support AI compliance and Migrating AI workloads to the cloud: an Australian guide.
Frequently asked questions
Can a standard enterprise data centre be retrofitted for AI workloads? Rarely, without significant capital investment. Power infrastructure, cooling systems, and structural floor load tolerances for GPU-dense racks are engineered in from the start. Retrofitting a facility built for 5 to 10kW racks to support 30 to 100kW racks typically means rebuilding the power and cooling layers, not upgrading them.
Is air cooling ever sufficient for AI workloads? It can work up to roughly 15 to 20kW per rack, which covers lighter inference workloads. Beyond that, air cooling becomes thermally inefficient, and facilities need direct liquid cooling or immersion cooling to remove heat from GPU-dense racks without excessive energy overhead.
Does a fast network connection guarantee good AI performance? No. Training and inference have different network requirements. Distributed training needs high-bandwidth, low-latency GPU-to-GPU interconnects. Inference needs a consistent, low-latency path between the model server and the end user. A facility can be well-connected for one and poorly suited for the other.
What’s the first thing to verify when a facility claims to be “AI-ready”? Power density per rack, and whether that capacity exists today rather than as a planned upgrade. Marketing claims of AI-readiness are common; confirmed power infrastructure, rated PUE, and cooling method are what actually determine whether a facility can run GPU-dense workloads.
