What 20kW Actually Means
A standard 42U rack in a high-density data centre at 20kW average power density is a reasonable spec for virtualised compute workloads, hyperconverged infrastructure, and scale-out storage. It is adequate for CPU-only inference and for general enterprise IT. It is not adequate for GPU-accelerated training or high-throughput AI inference using current generation hardware.
NVIDIA’s H100 GPU is the current baseline for serious AI compute deployments. A single H100 GPU draws up to 400W under full load. In an 8-GPU server configuration (the standard DGX H100 single node), that is approximately 3.2kW for GPUs alone. Add CPU (2 × Intel Xeon or 2 × AMD EPYC, ~200–400W combined), NVMe storage, networking (InfiniBand HCA, ~40W), and power supply inefficiency, and a single DGX H100 server draws approximately 10–11kW at full utilisation (source: NVIDIA DGX H100 System Architecture Specification, 2024).
In a standard 42U rack, you can fit approximately 6–7 DGX H100 servers (they are 6U each). That is 60–70kW per rack minimum at full utilisation.
A data centre at 20kW per rack average cannot run a GPU cluster in any meaningful configuration. It can run single-GPU workstations. It cannot run the clusters that hyperscalers and serious AI operators are deploying.
The Infrastructure Consequences at 60kW+
The transition from 20kW to 60kW per rack is not a scaling problem. It is a category change in mechanical, electrical, and structural infrastructure.
Cooling at 60kW+
At 20kW per rack, hot-aisle/cold-aisle containment with room-level CRAC/CRAH cooling is effective. The airflow mathematics work: a standard CRAC unit delivering 20–30kW of cooling capacity serves 1–2 racks adequately.
At 60kW per rack, the airflow mathematics do not work. To remove 60kW of heat via air cooling from a standard 42U rack (600mm × 1200mm footprint), you need to move approximately 1,200 cubic feet per minute (CFM) of air across the rack face at a 10°C temperature rise (inlet 20°C, exit 30°C). This exceeds what any conventional CRAC unit can deliver at that rack pitch and air velocity. The physics of air cooling at this density requires velocity levels that create unacceptable acoustic levels, condensation risk, and structural loads on the server chassis.
Liquid cooling is not an upgrade option at 60kW per rack. It is a physical requirement. Direct liquid cooling (DLC) with cold plates attached directly to GPU and CPU die surfaces removes heat at the source, eliminating the air-to-component thermal resistance that makes air cooling ineffective. The facility requires: chilled water plant upgrade; Coolant Distribution Units (CDUs) at each row — typically one CDU per 4–8 racks; piped cooling distribution infrastructure to each rack; and a leak detection system (a chilled water leak in a live GPU cluster causes catastrophic and expensive hardware failure).
Electrical distribution at 60kW+
Standard 42U rack PDUs are designed for 16A or 32A circuit density. At 240V single-phase, a 32A PDU delivers 7.68kW. To deliver 60kW to a rack from standard PDU infrastructure, you need approximately 8 × 32A circuits per rack — which exceeds the physical circuit capacity of most PDU configurations and creates cable management problems at scale.
The industry response has been busway (overhead bus duct) distribution systems. A 400V three-phase busway tap box rated at 125A per phase delivers approximately 86kW per tap at full three-phase utilisation. This is the correct scale for a 60–80kW rack. Busway deployment requires the facility’s electrical distribution to be redesigned from the switchboard outward — existing cable ladder and PDU infrastructure cannot be retrofitted.
Floor loading at 60kW+
A fully configured DGX H100 rack (6 × DGX H100 servers) weighs approximately 750–900kg depending on storage and networking configuration. The cooling infrastructure adds further mass: a CDU servicing 4 racks weighs 200–300kg empty, plus fluid mass. A row of 8 GPU-density racks with associated CDUs creates point load concentrations of 1,200–1,500kg/m² on localised slab areas.
Standard industrial slab design for light manufacturing or warehouse is 5–7kN/m² (approximately 500–700kg/m²). Even a purpose-built data centre designed for high-density at 10–12kN/m² is at or below the structural requirement for a concentrated AI-density deployment. Structural assessment and likely slab reinforcement or post-tensioned slab replacement is required for any existing facility being retrofitted for AI-density.
What This Means for Site Selection
The 60kW per rack threshold changes which sites are viable for AI-compute tenancy. The conventional data centre site selection criteria — grid headroom, fibre, planning risk — remain relevant, but two additional constraints dominate.
Water for liquid cooling
DLC systems do not require large volumes of water directly — the fluid in the CDU circuit is a closed loop. But the CDU exchanges heat to the facility’s chilled water (CHW) plant, which in turn requires a cooling tower or dry cooler to reject heat to the atmosphere. At high density and in water-constrained geographies, the cooling tower’s water consumption becomes a significant operational concern.
A 10MW IT load served by DLC with a conventional evaporative cooling tower will consume approximately 20–30 million litres of water per year (based on ASHRAE 90.4-2019 water usage effectiveness benchmarks). In water-licensed regions of Australia (most of inland NSW and Victoria operate under water sharing plans with licensed extraction limits), this consumption needs to be planned and licensed.
Power supply certainty for continuous GPU workload
GPU training workloads are not interruptible. A training run on a large model may run for days or weeks without interruption. A grid supply interruption that kills a training run wastes hundreds of thousands of dollars of GPU time and the training history to that point.
Sites that depend exclusively on NEM grid supply for AI-density training workloads are exposed to NEM market volatility and unplanned supply interruptions. The site selection and power strategy for a serious AI-density data centre needs to include either 2N power supply (two independent grid connections at transmission voltage) or a BTM generation and storage solution that can island the facility through NEM supply events.
This constraint effectively rules out any site that cannot access either dual-path grid connection or BTM generation within the development economics. In the AU context, dual-path grid connection at transmission voltage typically requires proximity to an existing TNSP substation with a spare transformer position — a configuration that is available in a small number of industrial locations.
The Builder’s Dilemma
The majority of Australian data centre capacity built in the 2010s and early 2020s was designed for 10–15kW per rack. Bringing that capacity to 60kW per rack involves: structural assessment and potential slab reinforcement (AUD 500K–5M+ per data hall depending on area); electrical distribution replacement (busway install, switchgear upgrade: AUD 1–3M per MW IT); cooling plant upgrade and DLC distribution (AUD 500K–1.5M per MW IT).
The retrofit economics are difficult. In most cases, the cost of retrofitting a high-density facility to AI-density exceeds the incremental rental income from the density upgrade. New builds designed for AI-density from the ground up are more economical at scale.
This creates a pipeline problem for Australian sovereign-cloud and hyperscale tenants who need AI-density capacity now: the existing estate cannot serve them efficiently, and the new builds designed to specification are not operational until 2027–2028 at earliest.
The data centre that closes the gap — designed for 60kW+ from the foundation, with liquid cooling infrastructure, busway distribution, structural spec, and a viable water strategy — is the asset the market is actually looking for.
Sources: NVIDIA DGX H100 System Architecture Specification 2024; ASHRAE 90.4-2019 Energy Standard for Data Centers; ASHRAE 2021 Climate Design Data (AU locations); NEXTDC AI-density capability disclosure, Annual Report FY2025; JLL AU Data Centre Report Q4 2025; APC Schneider Electric Busway Application Guide.