Data Centre Market Navigation

Knight Frank's 2025 data centre capex benchmarks establish baseline construction at $15M per megawatt, reaching $17.25M per megawatt total capex inclusive of acquisition, entitlements, and soft costs (Knight Frank 2025; indicative; construction costs vary significantly by geography, specification, and market conditions). This assumes Tier III or Tier IV specification (99.9% or 99.99% uptime SLA). GPU-focused colocation facilities command 50-75% capex premium over enterprise data centre standards due to power density, cooling infrastructure, and seismic hardening requirements.

A 16-rack GPU cluster consuming 1.9MW peak demands roughly $17.25M-$25.9M real estate capex at greenfield sites. Existing enterprise facilities retrofitted for GPU workloads reduce capex by 35-45% but sacrifice spatial efficiency and thermal distribution optimality. Lease rates for colocation stack at $450-$650 per kilowatt per year in primary markets (North America, Europe), versus $180-$280 for standard enterprise colocation.

Current-generation Blackwell accelerators in NVL72 rack configuration (72 GPUs per 19-inch rack) consume approximately 120 kilowatts per rack at full utilisation. This necessitates dedicated 480V three-phase circuits with 250-ampere capacity per rack, an improvement in power density versus prior-generation architectures. UPS provisioning for 15-minute hold-up (sufficient for graceful shutdown) requires lead-acid or lithium-iron-phosphate battery banks at roughly $18,000-$28,000 per megawatt.

Generator backup for multi-hour outages runs $35,000-$52,000 per megawatt installed. Renewable energy sourcing adds 8-15% capex premium but reduces operating margin via power purchase agreements. Circuit redundancy (N+1 or N+2) for critical loads demands oversized electrical plant, typically 20-30% excess capacity to accommodate equipment diversity factor and future expansion.

Direct liquid cooling is mandatory for current-generation GPU clusters operating at this power density. GB200 NVL72 racks require either immersion cooling (mineral oil or synthetic dielectric fluid) or closed-loop liquid systems delivering 30-40 litres per minute per rack at 35°C supply temperature.

Immersion cooling capex runs $42,000-$68,000 per rack including tank, pump, filtration, and fluid. Closed-loop systems cost $28,000-$38,000 per rack but demand precision chiller integration.

Facility chilled water loops require dedicated 500-ton capacity chillers at $180,000-$240,000 per 500-ton unit. Cooling tower capability for waste heat rejection requires 250-400 tons additional capacity depending on climate zone and ambient wet-bulb temperature. A 32-rack cluster demands approximately $1.8M-$2.4M cooling infrastructure capex. Glycol-based coolants require quarterly flushing and replenishment at $3,200-$4,800 per annum per cluster.

Hyperscaler-owned co-location capacity accounts for 62% of GPU workload. Third-party colocation providers (Equinix, Digital Realty, CoreWeave, Lambda Labs) hold the remaining 38%.

Current facilities marketed as 'GPU-ready' define minimum capability as 50kW power per rack, single-phase cooling, and standard Ethernet. True current-generation capability requires 120kW per rack, direct liquid cooling, and 800Gbps fabric networking.

Fewer than 380 data centre facilities globally meet these specifications as of Q1 2026; capacity is expanding rapidly and this figure should be treated as a point-in-time estimate. Colocation wait times for GPU-capable racks extend 4-8 weeks in primary markets. Operators charging GPU premium rates ($650-$900 per kilowatt per year) realise gross margins of 68-74% on colocation revenue versus 41-51% on standard enterprise colocation. Regional availability: North America leads with 156 qualified facilities, Europe-Middle East-Africa follows at 118 facilities, APAC trails at 106 facilities.

Data centre capex: $15M/MW build, $17.25M/MW total (Knight Frank 2025; indicative;varies by geography and specification); GPU facilities command 50-75% premium

Blackwell NVL72 racks consume 120kW per unit requiring dedicated 480V 250A circuits and N+1 electrical redundancy

Direct liquid cooling mandatory: immersion systems cost $42k-$68k/rack, facility chillers add $1.8M-$2.4M per 32-rack cluster

Fewer than 380 GPU-ready facilities globally (as of Q1 2026;capacity is expanding rapidly); colocation lead times extend 4-8 weeks