Terrestrial Data Center Infrastructure Cost

Answer

The all-in infrastructure cost per kW_IT for a terrestrial AI data center -- encompassing shell-and-core construction, power distribution, cooling systems, and site infrastructure, but excluding GPU/server hardware -- falls in the range of $8,000-$20,000/kW_IT (equivalently $8-20M/MW), with a central estimate of $12,500/kW_IT ($12.5M/MW) for a US-based hyperscale AI facility in 2025-2026.

This decomposes approximately as:

• Shell and core (building, site, land): $2,500-$4,000/kW_IT
• Electrical infrastructure (switchgear, transformers, UPS, generators, PDUs): $4,000-$8,000/kW_IT
• Cooling systems (liquid cooling CDUs, piping, heat rejection): $1,500-$4,000/kW_IT
• Building fit-out and other systems (fire suppression, security, networking infrastructure): $1,000-$3,000/kW_IT

The optimistic value of $8,000/kW_IT reflects the lowest-cost US markets (Texas) for traditional or lightly AI-optimized facilities. The conservative value of $20,000/kW_IT captures fully AI-optimized facilities with liquid cooling, 2N redundancy, and behind-the-meter power infrastructure in high-cost markets.

Evidence

Shell-and-core construction costs

E1. [evidence] JLL 2026 Global Data Center Outlook (jll-2026-dc-outlook) reports shell-and-core construction costs increased from $7.7M/MW in 2020 to $10.7M/MW in 2025, with a 2026 forecast of $11.3M/MW (6% increase). This figure covers the building structure and basic site infrastructure only, and explicitly excludes tech fit-out. JLL states that AI tech fit-out can add up to $25M/MW on top of shell-and-core.

E2. [evidence] Turner & Townsend DCCI 2025 (turner-townsend-dcci-2025) reports a 5.5% YoY cost increase for traditional cloud/air-cooled data centers, measured in $/W across 52 global markets. The index identifies a 7-10% construction cost premium for AI-ready facilities over traditional ones. Most expensive markets: Tokyo $15.2/W, Singapore $14.5/W, Zurich $14.2/W. US markets: Silicon Valley $13.3/W, New Jersey $12.9/W, Charlotte $9.5/W.

E3. [evidence] Cushman & Wakefield US Data Center Development Cost Guide 2025 (cushman-wakefield-dc-cost-2025) found costs across 19 US markets ranging from $9.3M/MW (San Antonio) to $15M/MW (Reno), with an average of $11.7M/MW. Texas markets (Dallas, Austin, San Antonio) consistently recorded the lowest costs. This figure covers the full development cost (shell, MEP, fit-out) but excludes IT equipment, land acquisition, and soft costs.

E4. [evidence] ChinaTalk US vs China DC cost comparison (chinatalk-dc-cost-comparison) states US data center construction costs are $8-12M/MW, with a midpoint assumption of $10M/MW. Chinese data centers cost $5.5-6.5M/MW. These figures cover construction only, separate from hardware.

Cost breakdown by component

E5. [evidence] Dgtl Infra cost breakdown (dgtl-infra-dc-cost-breakdown) decomposes total data center development costs ($7-12M/MW) as follows: Electrical systems 40-45% of total (the single largest component); HVAC/mechanical/cooling ~20%; powered shell (land + building) 17-21%; building fit-out and other systems 20-25%. MEP systems collectively consume up to 50% of total budgets.

E6. [evidence] Dgtl Infra (dgtl-infra-dc-cost-breakdown) reports per-sqft costs: land $25-75/sqft, building shell $80-160/sqft, data center improvements (electrical, HVAC, fire suppression, fit-out) $520-900/sqft. Total development: $600-1,100/sqft.

E7. [evidence] TrueLook DC Construction Cost analysis (truelook-dc-construction-costs) confirms electrical systems at 40-45% of total, with cooling systems commanding 43.2% of mechanical infrastructure spending in 2024 (showing significant growth). Air cooling infrastructure costs $1.5-2M/MW; liquid cooling $3-4M/MW.

AI-specific infrastructure premium

E8. [evidence] JLL (jll-2026-dc-outlook) states AI tech fit-out adds up to $25M/MW. This covers power distribution at higher densities, liquid cooling infrastructure (CDUs, piping, heat rejection), rack-level power delivery (busbars, 480V distribution), and reinforced flooring for heavier AI racks.

E9. [evidence] Turner & Townsend (turner-townsend-dcci-2025) identifies a 7-10% construction cost premium between traditional and AI data centers in the US, reflecting higher complexity of AI-supporting facilities. This is the premium on the shell-and-core/base build, separate from the much larger fit-out premium.

E10. [evidence] Introl GB200 NVL72 deployment guide (introl-nvl72-deployment) reports that facility upgrades to support a single NVL72 rack cost $5-10M: reinforced flooring, 480V power distribution, liquid cooling infrastructure, and expanded network capacity. A single NVL72 rack draws 120 kW, implying retrofit infrastructure cost of $42-83/W_IT for brownfield conversions.

E11. [evidence] Alpha Matica analysis (alpha-matica-dc-cost-structure) models a 100 MW hyperscale data center with total CapEx of $3.4B-$5.5B ($34-55/W including IT hardware). Initial construction cost (infrastructure only) ranges from $900M-$1.5B for a 100 MW facility, implying $9-15M/MW infrastructure cost.

Cooling system costs

E12. [evidence] Introl liquid cooling analysis (introl-liquid-cooling-cost) reports complete liquid cooling infrastructure costs $3-4M/MW for new builds. Air cooling costs $1.5-2M/MW. Retrofit to liquid cooling costs $2-3M/MW with 40% energy savings. Building new 100 kW-capable rack infrastructure costs $200K-300K per rack.

E13. [evidence] Introl CDU cost analysis (introl-cdu-cost-analysis) prices CDUs at $75K-150K per 500 kW unit. Piping installation runs $50-100 per linear foot. Cold plates and manifolds add $5K-10K per server. The CDU market is projected to grow from ~$1B to $7.7B at 33% CAGR.

E14. [evidence] Introl liquid vs air cooling (introl-liquid-cooling) reports PUE values: air cooling 1.4-1.8; liquid cooling 1.05-1.15; immersion cooling 1.02-1.03. The PUE difference directly impacts the ratio of total facility power to IT load, and therefore the effective infrastructure cost per kW_IT.

Power infrastructure costs

E15. [evidence] McKinsey "Beyond compute" (mckinsey-beyond-compute) estimates 25% ($1.3T) of total $6.7T global data center investment through 2030 goes to "energizers" -- power generation, transmission, cooling, and electrical equipment. With projected 219 GW demand, this implies ~$5,900/kW in power/cooling infrastructure alone at the macro level.

E16. [evidence] McKinsey "Cost of compute" (mckinsey-cost-of-compute) breaks down the $6.7T investment: servers ~$3.5T, electrical/mechanical ~$0.8T, power generation ~$0.4T, storage ~$0.8T, remaining in other categories. The $0.8T electrical/mechanical for ~125 GW implies ~$6,400/kW for electrical and mechanical systems specifically. Note: the $42M/MW implied average includes IT hardware and is not infrastructure-only.

E17. [evidence] Thunder Said Energy (thunder-said-dc-economics) models standard data center capex at ~$10M/MW. AI data centers with heavy compute can reach $40,000/kW total capex, with over half being GPUs. This implies AI infrastructure-only capex of up to ~$15-20M/MW.

E18. [evidence] SemiAnalysis "Datacenter Anatomy Part 1: Electrical Systems" describes the standard power distribution path: MV switchgear, step-down transformers (to 415V AC), diesel generators, ATS (automatic transfer switches), UPS systems with 5-10 min battery storage, and PDUs. Microsoft uses standardized 3 MW generator and 3 MVA transformer pods for modularity and procurement efficiency.

Macro-level cost validation

E19. [evidence] xAI Colossus (xai-colossus-expansion): 2 GW facility, 555,000 GPUs, ~$18B total. Building permit filed for $659M. Land acquired for $80M. If $18B is primarily GPU cost (~$9M/MW in GPU costs per the source summary), then infrastructure cost is additional. The $659M building for a fraction of the 2 GW campus, plus the $80M land, suggests infrastructure costs in the range of $3-6B for the full 2 GW facility, or roughly $1.5-3M/MW -- but this is likely low because xAI used repurposed industrial infrastructure and behind-the-meter gas turbine power (not included in standard facility cost comparisons).

E20. [evidence] The Catalyst podcast (catalyst-scaling-pathways) notes that "only 5-15% of DC cost is energy" and that chip supply chain is the dominant cost. This implies infrastructure cost is a secondary but significant fraction of total -- consistent with the pattern that GPU hardware dominates total data center capex.

Regional variation and cost trajectory

E21. [evidence] Turner & Townsend (turner-townsend-dcci-2025): YoY cost increase slowed from 9.0% (2024) to 5.5% (2025), suggesting cost pressure is moderating but still positive. The 7% CAGR from 2020-2025 (JLL data) shows sustained upward trajectory.

E22. [evidence] Cushman & Wakefield (cushman-wakefield-dc-cost-2025): Texas is the lowest-cost US market at $9.3-10M/MW. High-cost markets (Reno, Silicon Valley) reach $13-15M/MW. This is a ~60% spread from lowest to highest US market.

E23. [evidence] ChinaTalk (chinatalk-dc-cost-comparison): Chinese construction costs ($5.5-6.5M/MW) are roughly 40-50% lower than US costs due to cheaper labor, faster construction timelines, and lower regulatory burden.

E24. [opinion] JLL forecasts 6% cost increase for 2026 to $11.3M/MW average shell-and-core globally. The trajectory suggests continued but decelerating cost inflation as supply chains stabilize but AI density requirements continue to escalate.

Analysis

Decomposing infrastructure cost from total project cost

The central challenge in this question is separating infrastructure cost from GPU/server hardware cost. Many widely cited $/MW figures bundle everything together:

• McKinsey's $42M/MW average (mckinsey-cost-of-compute: $5.2T for 125 GW) includes servers ($3.5T), storage ($0.8T), and electrical/mechanical ($0.8T). Stripping out IT hardware leaves approximately $6,400/kW_IT in electrical and mechanical infrastructure, plus additional shell-and-core costs.

• xAI's $9M/MW figure (xai-colossus-expansion) is explicitly GPU-only cost ($18B for 2 GW). Infrastructure is additional but hard to isolate from public filings.

• JLL's $11.3M/MW (jll-2026-dc-outlook) is infrastructure-only (shell-and-core), making it the cleanest benchmark -- but it excludes the AI-specific fit-out which adds substantially.

Building the infrastructure-only cost stack

Using the component percentages from Dgtl Infra (E5) and Cushman & Wakefield (E3), and calibrating against JLL shell-and-core data (E1):

For a standard US hyperscale DC ($10-12M/MW all-in infrastructure):

For an AI-optimized facility, add:

• Liquid cooling premium: +$1.5-2M/MW over air cooling (E12)
• Higher-density power distribution (480V busbars, enhanced UPS): +$1-3M/MW
• Reinforced flooring for 1,500+ kg racks: included in fit-out
• AI premium on base build: +7-10% (E9)

This brings the AI-optimized total to roughly $12-18M/MW for a greenfield US facility.

JLL fit-out paradox

JLL reports shell-and-core at $10.7-11.3M/MW, then states AI fit-out adds "up to $25M/MW" (E8). If taken literally, that would imply $36M/MW total infrastructure -- which seems high relative to other sources. The $25M/MW likely includes rack-level power distribution equipment, networking infrastructure, and possibly some IT-adjacent hardware (server chassis, cables) that blurs the infrastructure/IT boundary. The fit-out figure may also represent extreme cases (very high density, full liquid cooling, 2N+1 redundancy). For our purposes, the infrastructure-relevant portion of AI fit-out is likely $5-10M/MW above standard fit-out.

Behind-the-meter power

A growing trend is the inclusion of on-site power generation (gas turbines, solar, batteries) in the facility cost envelope. McKinsey's "Beyond compute" estimates 25% of total investment goes to power generation and electrical infrastructure (E15). The Catalyst podcast notes 30% of planned US DC capacity includes BTM power (latitude-btm-traction). When BTM generation is included, infrastructure costs can increase by $2-5M/MW depending on technology (gas turbines cheaper, nuclear significantly more expensive).

Our central estimate excludes BTM power generation capex, treating it as a separate line item analogous to a grid connection charge. If BTM were included, the upper range would shift to $22-25M/MW.

Regional spread

Proposed values

• Optimistic: $8,000/kW_IT ($8M/MW) -- Reflects the floor achievable in lowest-cost US markets (Texas) for large hyperscale facilities with modest AI optimization, taking advantage of economies of scale, favorable labor costs, and streamlined permitting. Also roughly corresponds to non-AI-premium global average.

• Central: $12,500/kW_IT ($12.5M/MW) -- US average for an AI-optimized hyperscale facility with liquid cooling, covering shell-and-core ($4,000-5,000/kW), electrical infrastructure ($4,500-5,500/kW), cooling systems ($2,000-3,000/kW), and fit-out ($1,500-2,500/kW). Consistent with Cushman & Wakefield's $11.7M/MW average plus the 7-10% AI premium from Turner & Townsend.

• Conservative: $20,000/kW_IT ($20M/MW) -- Represents fully AI-optimized facilities in high-cost markets with full liquid cooling, 2N redundancy, behind-the-meter power conditioning (but not generation), and premium construction timelines. Captures the high end of the alpha-matica $9-15M/MW infrastructure range plus AI fit-out premiums. Consistent with reports of AI facilities "easily doubling" standard costs to $20M/MW or more.

Cost trajectory outlook

Infrastructure costs have increased at ~7% CAGR from 2020-2025 (E1), but the rate is moderating (5.5% in 2025 vs 9.0% in 2024 per E21). Several countervailing forces:

Upward pressure: AI density requirements (liquid cooling, reinforced floors, higher-amperage power distribution), supply chain constraints (transformer lead times at 128 weeks per powermag-transformer-shortage), labor shortages in skilled trades, and increasing behind-the-meter power requirements.

Downward pressure: Economies of scale at GW+ campuses, modular/prefabricated construction techniques (xAI built in 122 days), standardized pod designs reducing procurement costs, and competition among hyperscalers driving efficiency innovation.

Net assessment: Infrastructure costs likely continue rising at 4-6% annually through 2027-2028, then stabilize as supply chains catch up and modular construction techniques mature. The AI premium may narrow as liquid cooling becomes the default rather than a premium option.