AI Has Rewritten the Data Centre Spec Sheet
A standard enterprise rack consumes 7-10 kW. An NVIDIA GB200 NVL72 rack consumes 120-140 kW. That single number — a 14x increase in power density — explains why the entire data centre industry is being redesigned around AI workloads.
Hyperscalers committed over $280 billion in aggregate capital expenditure for 2025-2026, the majority directed at AI infrastructure. The facilities housing this investment bear little resemblance to the data centres built even five years ago.
Power Density Requirements
**Current GPU rack power consumption:** - NVIDIA H100 (8-GPU server, air-cooled): 10-12 kW per server, 40-50 kW per rack - NVIDIA H200 (8-GPU server, liquid-cooled): 12-14 kW per server, 50-70 kW per rack - NVIDIA GB200 NVL72 (72-GPU rack): 120-140 kW per rack - AMD MI300X (8-GPU server): 10-13 kW per server, 45-55 kW per rack - Google TPU v5p (multi-chip pod): Variable, 20-80 kW per rack equivalent
**Facility-level implications:** - A 100 MW AI training facility using GB200 NVL72 racks requires approximately 700-800 racks - The same 100 MW with traditional enterprise load (8 kW/rack) would require 12,500 racks - Floor space per MW decreases dramatically: AI facilities need 30-50 sq ft per rack versus 25-30 sq ft for enterprise, but far fewer racks per MW - Result: AI data centres are smaller buildings drawing vastly more power per square foot
**Power distribution design:** - Standard enterprise: 208V or 480V, 30-60A circuits - AI-optimised: 415V or 480V three-phase, 100-200A circuits - Bus duct distribution preferred over traditional whip-and-plug for runs exceeding 50 feet - Power factor correction at the rack level (GPU power supplies have 0.95+ power factor, but harmonic distortion is significant)
Cooling Requirements
Air cooling cannot handle 120+ kW racks. The physics are straightforward: air has a volumetric heat capacity of approximately 1.2 kJ/m³·K, while water's is 4,180 kJ/m³·K — over 3,000 times more efficient per unit volume.
Cooling architecture for AI facilities:
**Direct liquid cooling (DLC) — the current standard:** - Cold plates on GPUs and CPUs, connected to facility chilled water via Coolant Distribution Units (CDUs) - Coolant supply temperature: 30-45°C (allows free cooling in most climates) - Coolant flow rate: 15-25 litres per minute per rack at 120 kW - Captures 70-80% of rack heat; supplemental air cooling handles remaining 20-30% - NVIDIA mandates liquid cooling for GB200 NVL72 — air cooling is not an option
**Rear-door heat exchangers (RDHx) — transitional solution:** - Water-cooled heat exchangers mounted on rack rear doors - Supports 30-60 kW per rack (insufficient for GB200 but adequate for H100 air-cooled) - Can be retrofitted into existing air-cooled facilities - Capital cost: $3,000-8,000 per rack
**Immersion cooling — emerging for highest densities:** - Single-phase immersion supports 150-250+ kW per tank - Eliminates all fans and air handling - Operational complexity (maintenance requires removing servers from fluid) - Vendors: GRC, LiquidCool Solutions, Submer, Asetek
Compare cooling technologies in detail in our cooling guide.
Network Fabric
AI training clusters are network-bound. A 10,000-GPU training job generates petabytes of inter-node communication per hour. Network design is as critical as compute and power.
**Backend network (GPU-to-GPU):** - NVIDIA NVLink and NVSwitch: 900 GB/s per GPU within a node (NVL72 rack) - InfiniBand NDR (400 Gb/s) or HDR (200 Gb/s): Standard for inter-rack GPU communication - Emerging: Ultra Ethernet Consortium's 800 Gb/s standard for AI-optimised Ethernet - Network topology: Fat-tree or rail-optimised topologies with full bisection bandwidth - Latency target: <2 microseconds switch-to-switch
**Frontend network (storage, management, internet):** - 100-400 GbE Ethernet for storage access - Separate out-of-band management network - Standard data centre interconnection for internet and WAN connectivity
**Physical requirements:** - Structured cabling: Single-mode fibre for runs >30m, multimode (OM4/OM5) for shorter distances - Cable management: A 1,000-GPU cluster can require 10,000+ fibre connections. Cable tray capacity is a frequent overlooked constraint. - Cross-connects and carrier access: Evaluate connectivity at target facilities using our facility directory
Floor Loading and Structural
**Weight considerations:** - Standard IT rack: 1,500-2,500 lbs loaded - GB200 NVL72 rack with CDU: 3,000-4,500 lbs - Required floor loading: 250-350 lb/sq ft (versus 150 lb/sq ft standard) - Many existing raised-floor facilities cannot support AI rack weights without structural reinforcement - Slab-on-grade construction preferred for new AI-optimised facilities
**Ceiling height:** - AI racks (particularly NVL72) are 52-58U, requiring 10-12 ft clearance - Overhead CDU piping, cable trays, and fire suppression add 3-5 ft - Minimum clear height: 14-16 ft (versus 10-12 ft for standard data centres)
Facility Design Principles for AI
1. **Power density first:** Design electrical distribution for 50+ kW/rack from day one, even if initial deployment is lower density. Retrofitting power distribution is expensive. 2. **Liquid cooling infrastructure from shell:** Install chilled water piping, CDU pads, and floor drains during initial construction. Adding liquid cooling to an air-cooled facility costs 3-5x more than building it in. 3. **Minimise GPU-to-GPU distance:** AI training performance degrades with network distance. Keep the entire GPU cluster within a single hall where possible. Splitting across buildings adds latency and complexity. 4. **Redundancy trade-offs:** AI training workloads checkpoint progress and can tolerate brief interruptions. Tier II or III reliability (rather than Tier IV) is standard for AI training facilities, reducing capital cost by 15-25%. 5. **Storage adjacency:** Large-scale AI training requires petabytes of high-speed storage. Co-locate storage arrays in the same hall as GPU racks to minimise storage network latency.
Assess any facility or property for AI readiness using the Score Tool. For AI infrastructure advisory, contact our team.