Hyperscale fitted-out data centers run $10M to $15M-plus per MW of IT load in 2026, depending on cooling design (air vs. evaporative vs. liquid), power redundancy (N+1 vs. 2N), and geography. Powered-shell-only delivery (shell with utility power but without mechanical fit-out) runs roughly half that per MW.
What drives the per-MW number
Cooling design is the single largest swing factor. Air cooling is cheapest per MW but hits a feasibility ceiling at roughly 30 kW per rack. Evaporative cooling adds capex and water consumption but supports higher densities. Liquid cooling (direct-to-chip or immersion) is the most expensive per MW but is required for AI workloads above 50 kW per rack.
Power redundancy is the second largest swing factor. N+1 (one redundant component per critical system) costs less than 2N (full duplication) but carries more downtime risk. Tenants increasingly demand 2N for AI training workloads where downtime carries meaningful retraining cost.
Geography moves the number through labor cost, materials availability, and utility-driven design constraints. Northern Virginia and Silicon Valley typically run at the top of the range; Iowa and Ohio at the lower end.
Sources
- 2024 United States Data Center Energy Usage Report · Lawrence Berkeley National Laboratory (LBNL) · 2024
- Powering Intelligence: Analyzing Artificial Intelligence and Data Center Use of Electricity · Electric Power Research Institute (EPRI) · current