TL;DR: data center construction cost per MW in 2026
A useful cost-per-MW model separates what is included before comparing numbers. Shell, powered shell, full fit-out, utility interconnect, cooling, land, soft costs, labor, equipment, and contingency can move independently.
| Cost layer | Best for | Quick list hint |
|---|---|---|
| Land and site control | Early market screening | Include option payments, due diligence, zoning, environmental work, and expansion land. |
| Utility interconnect | Power-risk underwriting | Separate substation, transmission, transformer, network-upgrade, deposit, and schedule risk. |
| Shell and core | Building-cost comparison | Clarify whether structure, roof, envelope, security, offices, and yards are included. |
| Powered shell and MEP | Developer or lease comparison | Separate electrical, mechanical, controls, generators, fuel, UPS, and commissioning scope. |
| Cooling and heat rejection | AI-density premium review | Model air, rear-door, direct-to-chip, immersion, water, and heat-rejection differences. |
| Active IT and tenant fit-out | GPU or server deployment planning | Keep servers, GPUs, storage, network, cabling, cages, and tenant labor separate. |
| Soft costs and contingency | Investment committee review | Include design, permits, financing, insurance, owner costs, escalation, and schedule buffer. |
What should be included in cost per MW?
The denominator should be explicit: critical IT MW, utility service MW, total connected load, or a phase-level number. The numerator should state whether it includes land, utility upgrades, shell, MEP, cooling, backup power, tenant fit-out, active IT, financing, contingency, and owner costs. Many bad comparisons mix a powered-shell estimate with a full AI fit-out estimate.
How does AI density change the model?
AI workloads can increase the cost of power distribution, backup power, cooling, floor loading, controls, commissioning, network, and tenant work. A facility designed for lower-density enterprise racks may need expensive retrofits before it can support GPU clusters. Model density and cooling as explicit assumptions, not as a generic premium.
How should buyers compare build, powered shell, and lease options?
Owned build models put more site, utility, and delivery risk on the buyer but can preserve long-term control. Powered shell can move some delivery risk to a developer while leaving fit-out and operations questions open. Leased capacity can reduce time to service when the capacity is real, but buyers still need to normalize deposits, pass-throughs, renewal risk, expansion rights, and delay remedies.
What evidence should support the estimate?
Require a line-item assumptions register, quantity takeoff, basis of design, utility scope, interconnection status, cooling design, equipment lead times, labor assumptions, escalation basis, contingency policy, tax incentive assumptions, exclusions, and schedule. Keep provider quotes, market reports, and engineer estimates in separate evidence columns so the model can be revised without hiding the source.