How much does data center construction cost per MW in 2026?
Use cost per MW as a budgeting screen, not a quote. A mainstream 2026 construction benchmark is roughly $10 million to $12 million per MW for many hyperscale or build-to-suit facilities before every buyer-specific scope item is added.
JLL forecasts an average global construction cost of about $11.3 million per MW for 2026 and separately notes that tenant AI technology fit-out can cost as much as $25 million per MW. Turner & Townsend reports that traditional air-cooled data center construction cost inflation moderated in 2025, but U.S. liquid-cooled data centers of similar IT capacity carry an average construction premium of about 7% to 10%.
| Budget view | Planning range | Usually included | Must verify |
|---|---|---|---|
| Core construction benchmark | About $10M-$12M per MW for many mainstream 2026 benchmarks | Shell, core, architecture, mechanical and electrical fit-out, contractor costs, and major M&E equipment depending on source scope | Whether the benchmark is IT load or facility power, and whether land, utility works, active IT, owner costs, tax, and contingency are excluded |
| Higher-cost global markets | About $13M-$15M per MW in expensive index markets | Regional labor, material, equipment, contractor, and supply-chain pressure | FX rate basis, local tax, import duty, contractor availability, land scarcity, and grid capacity |
| Liquid-cooled AI facility premium | About 7%-10% above comparable air-cooled construction cost in Turner & Townsend's U.S. analysis | Higher-density mechanical and cooling systems, white-space piping, CDUs, and altered heat-rejection design | Whether the premium excludes active IT, utility works, land, and owner soft costs |
| Tenant AI technology fit-out | Can add up to $25M per MW for AI infrastructure in JLL's outlook | GPUs, servers, networking, storage, rack systems, and tenant-side deployment work | Whether a quoted number is landlord shell/core, powered shell, tenant fit-out, or all-in project capex |
| Full project pro forma | Site-specific and not comparable without scope control | Land, utility interconnect, substations, owner costs, contingency, financing, tax, and incentives | Define inclusions before comparing bids or markets |
Before carrying the range into an RFP workbook, use the Data Center Cost per MW Calculator to test shell, powered-shell, liquid-cooling, and AI fit-out assumptions. If the build case is close to a leased alternative, compare utilization, term length, and risk transfer in the Build vs Lease AI Capacity Calculator.
What should buyers include in a per-MW budget?
Separate the budget into cost layers before comparing markets, contractors, developers, or colocation alternatives. The same $/MW number can mean shell and core, powered shell, full mechanical/electrical construction, or a tenant AI fit-out with active IT hardware.
| Cost layer | Buyer question | Cost risk |
|---|---|---|
| Land and campus control | Is the site large enough for current halls, substations, cooling, staging, and future phases? | Large parcels and power-ready sites can move faster than headline land price suggests |
| Utility interconnect and power delivery | Is power committed, deliverable, and paid for by the utility, landlord, tenant, or developer? | Interconnection upgrades, substations, switchyards, bridge power, and bring-your-own-power mandates can sit outside a base construction benchmark |
| Shell and core | What is the cost to deliver the building envelope, structural work, core infrastructure, and base construction package? | Shell-only numbers can look low if they exclude MEP, utility works, tenant fit-out, and owner costs |
| Powered shell and MEP | What electrical, mechanical, generator, UPS, switchgear, cooling, and controls work is included? | Long-lead electrical equipment and high-density cooling can dominate schedule and contingency |
| Cooling and heat rejection | Is the design air-cooled, close-coupled, direct-to-chip, rear-door, immersion, or hybrid? | AI density can shift cost from building area to liquid loops, CDUs, heat rejection, fluid policy, and operations |
| Active IT and GPU fit-out | Are servers, GPUs, networking, storage, racks, and tenant deployment included? | AI hardware can exceed the facility construction budget and should not be blended into landlord construction cost without a separate line |
| Labor, contractor, and schedule | Is the market deep enough in data center trades and specialist contractors? | Labor scarcity, overtime, phased delivery, and bid inflation can change cost faster than square-foot assumptions |
| Contingency and owner costs | What soft costs, professional fees, permits, financing, tax, incentives, and escalation are modeled? | Missing owner-side assumptions can make two apparently similar $/MW benchmarks impossible to compare |
How do shell, powered shell, full fit-out, and AI budgets differ?
Shell and core is the narrowest useful lens. It tells buyers what the base data center structure and core construction package may cost, but it often excludes land purchase, abnormal groundworks, utility upgrades, active IT equipment, fiber cabling outside the construction scope, and professional services.
Powered shell adds the electrical and mechanical infrastructure needed to support IT load. That is usually the better comparison point for developers and large tenants because electrical equipment, backup power, cooling, and controls are where much of the data center-specific cost sits.
Full fit-out includes the tenant or owner work required to make the capacity usable. For conventional enterprise workloads, that may mean racks, network, and storage fit-out. For AI, it may include GPUs, high-speed fabric, liquid-ready rack systems, and commissioning work that belongs in a separate technology budget. Buyers should keep landlord construction, tenant technology fit-out, and all-in project capex in separate columns until the final investment committee model.
Why do AI and high-density workloads change cost per MW?
AI workloads move the budget from generic white space toward power density, cooling design, electrical transients, and rack-level integration. Uptime Institute says liquid cooling is typically used for high rack power above 50 kW or specialized high-performance IT, and that direct liquid cooling changes facility operations through added piping, coolant distribution units, and service-boundary questions.
NVIDIA's GB200 NVL72 rack shows why this matters: the platform uses liquid cooling and high-bandwidth rack-scale architecture to increase compute density. DPR's public Abilene campus description also shows how AI projects are being designed around high-density halls, direct-to-chip liquid cooling or rear-door heat exchangers, and very large IT-load blocks.
The buyer takeaway is simple: a 50 MW air-cooled enterprise facility, a 50 MW liquid-ready AI training hall, and a 50 MW all-in GPU campus should not be compared with a single blended $/MW number. Normalize the scope first, then compare market, schedule, and delivery risk.
Which assumptions move the number fastest?
The fastest-moving assumptions are power basis, geography, cooling, redundancy, schedule, and what sits outside the quoted construction package. CBRE reports constrained supply, high preleasing, and rent premiums for AI-optimized facilities with liquid cooling and high-power-density racks. Cushman & Wakefield points to limited power in established markets, larger land parcels, material and equipment lead times, and labor availability as core development-cost variables.
| Assumption | Why it matters | What to ask |
|---|---|---|
| IT load vs facility power | A cost per MW of IT load is not the same as cost per MW of total utility capacity | Ask every bidder to state the denominator and PUE or design-load basis |
| Market and labor pool | Contractor depth, labor rates, tax, import duty, and local supply chain change the delivered cost | Ask for market-specific labor, materials, and equipment assumptions |
| Power delivery | Utility interconnect delays and required upgrades can move cost and schedule outside the building package | Ask who pays for substations, switchyards, transmission upgrades, bridge power, and standby generation |
| Cooling architecture | Liquid cooling can reduce floor area pressure but adds design, equipment, commissioning, and operating requirements | Ask whether the model includes CDUs, secondary loops, heat rejection, fluid management, leak detection, and warranty support |
| Redundancy and reliability | Tier, topology, and resilience targets change UPS, generator, switchgear, and distribution cost | Ask whether the design is N, N+1, 2N, distributed redundant, or another topology |
| Procurement schedule | Long-lead electrical and mechanical equipment can drive escalation and holding costs | Ask for equipment lead times, escalation allowances, alternates, and procurement release dates |
| Incentives and taxes | Incentives can change the pro forma but may not reduce upfront build cost | Keep incentives separate from construction cost until eligibility and clawbacks are verified |
How should buyers use cost per MW before an RFP?
Use cost per MW to screen feasibility, not to pick a market or contractor. The first pass should normalize scope, denominator, design density, redundancy, cooling architecture, power-delivery responsibility, construction schedule, and active IT exclusions.
Before an RFP, ask each bidder for a cost workbook that separates land, utility interconnect, shell/core, mechanical, electrical, cooling, controls, generators, UPS, security, fiber, owner soft costs, contingency, escalation, and active IT. Require the bidder to state whether the model is based on IT load or facility power, whether it assumes air or liquid cooling, and whether it includes tenant technology fit-out.
For AI deployments, add a separate fit-out schedule for GPUs, servers, networking, liquid-ready rack systems, commissioning, spares, and operations training. That keeps a real estate budget from hiding a technology budget.
Which sources should buyers use for cost benchmarks?
Use source families, not a single cost quote. Analyst outlooks help frame benchmark ranges, cost indexes help explain market spread, and project/operator sources help test whether the technical assumptions match real AI builds.
| Source | What it supports | Caveat |
|---|---|---|
| JLL Data Center Outlook | 2026 global cost per MW, shell/core distinction, and tenant AI tech fit-out context | Global average, not a site-specific quote |
| Turner & Townsend Cost Index | US$/W market comparisons, 2025 cost inflation, cost allocation, and liquid-cooled AI premium | Index benchmark, not a delivered GMP |
| Turner & Townsend Methodology | Included cost headings, 30-50 MW IT-load baseline, and exclusions | Excludes land, utility works, active IT, abnormal groundworks, and some soft costs |
| Cushman & Wakefield Development Cost Guide | U.S. land, power, material, equipment, and labor-cost pressure | U.S. guide highlights, not a project quote |
| CBRE North America Data Center Trends | Supply pressure, high-density rent premiums, and power/procurement constraints | Market pricing evidence, not construction-cost guidance |
| Uptime Institute AI Cooling Methods | Rack-density thresholds and liquid-cooling operational implications | Cooling guidance, not a cost index |
| NVIDIA GB200 NVL72 evidence | Rack-scale liquid-cooled AI density context | Vendor platform evidence, not a facility cost quote |
| DPR Crusoe Abilene Campus | AI campus scale, high-density halls, and direct-to-chip or rear-door design context | Project evidence, not a public cost model |