CRE42 Commercial Real Estate Investment Strategy
Select economic topic or property type
MIT Center for Real Estate

Data Center & Power — Metrics Cheat Sheet

The U.S. data center buildout is generating a torrent of headlines with numbers that are difficult to compare. One announcement quotes megawatts; another quotes total investment. Cost estimates range from $5 billion to $80 billion per GW depending on what is included in the scope—facility construction only, IT hardware (GPUs and servers), or a fully loaded campus with dedicated power generation. Square footage, acreage, and power capacity are used interchangeably as measures of scale. This tool translates between these metrics using a consistent, source-backed methodology so that individual project announcements, regional forecasts, and national projections can be placed in proper context.

Select a facility type and enter a power capacity below to see derived metrics across both scenarios. For detailed methodology, source citations, and sensitivity analysis, see the Data Center Workbook

Data Center Power Utilization Measurement Tool
Configure two scenarios to compare power, cost, physical footprint, and national scale
■ Scenario A
Advanced Settings
■ Scenario B
Advanced Settings
■ Scenario A — AI Training
$45.5B
■ Scenario B — Traditional
$15.0B
Real Estate (Land, Shell, Fit-Out, Networking)
Power Infrastructure (Electrical, Cooling)
IT Hardware (GPUs, Servers, Memory)
Dedicated Power Generation
5.69
TWh / Year
542K
Equivalent Homes
2.0—3.5M
Building SF
450—1,100
Campus Acres
0.14%
% Addition to Current Total U.S. Electricity Usage (4,000 TWh)
3.2%
% Addition to Current U.S. DC Fleet (176 TWh)
4.38
TWh / Year
417K
Equivalent Homes
8.0—10.0M
Building SF
150—400
Campus Acres
0.11%
% Addition to Current Total U.S. Electricity Usage (4,000 TWh)
2.5%
% Addition to Current U.S. DC Fleet (176 TWh)

Metric Definitions & Relationships

MW / GW (Power Capacity)
Instantaneous power draw a facility can support. 1 GW = 1,000 MW. This is nameplate capacity—actual consumption depends on utilization rate. Most headlines report nameplate capacity.
TWh (Annual Energy Consumption)
Energy actually consumed over a year. Conversion: TWh = GW × 8,760 hours × utilization rate. At 65% utilization, 1 GW = 5.69 TWh/year. The U.S. data center fleet consumed ~176 TWh in 2023 (LBNL).
Utilization Rate
Percentage of nameplate capacity actually used on average. LBNL uses 50% as a fleet-wide average across all facility types. AI training facilities run 60–90%. This tool defaults to type-specific rates: 50% (traditional), 60% (hyperscale), 65% (AI training), 55% (AI inference).
Power Density (kW/rack)
Power consumed per server rack. Traditional: 5–10 kW. Hyperscale cloud: 15–30 kW. AI training: 40–130 kW (projected to 250 kW). Higher density = more power in less building space, which is why AI facilities need less SF per MW but more acreage for cooling and power infrastructure.
Cost Scope Ladder
Why $/GW estimates range from $5B to $80B: Facility only (land, shell, electrical, cooling, networking): $10.5–19B/GW. + IT Hardware (GPUs, servers): $15.5–47B/GW. + Dedicated generation (gas, nuclear): $19–52B/GW. Always ask what is included in the scope.
PUE (Power Usage Effectiveness)
Total facility energy ÷ IT equipment energy. A PUE of 1.0 means all power goes to computing; 2.0 means half goes to cooling and overhead. Hyperscale targets 1.1–1.2. Enterprise averages 1.5–1.8. AI liquid cooling is helping drive PUE down even as total power rises.
Reading the Headlines
Recent announcements translated using the methodology above
"Meta announces 5 GW Hyperion campus in Louisiana for $27 billion"
At 5 GW nameplate, this facility would consume ~28.5 TWh/year at 65% utilization—equivalent to powering 2.7 million homes and adding 16% to the current U.S. data center fleet. The $27B price tag covers facility construction only (~$5.4B/GW), consistent with an AI-ready powered shell without IT hardware. Fully loaded with GPUs, total investment would likely reach $150–235B. The 2,250-acre campus is roughly the size of lower Manhattan. Sources: Data Center Knowledge (Feb 2026); McKinsey (Aug 2025)
"Stargate Abilene: $40 billion AI campus with 1.2 GW capacity"
At 1.2 GW, annual consumption would be ~6.8 TWh—comparable to the entire residential electricity use of a mid-sized state. The $40B figure (~$33B/GW) includes Oracle’s NVIDIA GPU deployment, placing it between facility-only and fully-loaded estimates. The 875-acre campus spans 4M SF across 8 buildings. This single project would add ~3.9% to the current U.S. data center fleet and represent ~0.17% of total U.S. electricity consumption. Sources: Data Center Frontier (Aug 2025); CNBC
"Meta breaks ground on $10 billion, 1 GW data center in Indiana"
At 1 GW and 65% utilization, annual consumption would be ~5.7 TWh—equivalent to powering ~542,000 homes and adding ~3.2% to the current U.S. data center fleet. The $10B price tag (~$10B/GW) covers facility construction only—land, shell, electrical, cooling, and $120M+ in local infrastructure (roads, water, transmission). It does not include GPUs or servers. Meta’s Louisiana Hyperion project reveals the hidden layer: sales tax exemptions on “tens of billions” in GPU purchases imply $15–20B+ in IT hardware per GW on top of facility costs. The 1,500-acre campus will include 13 buildings (10 data halls) across 4M SF. Sources: Data Center Knowledge (Feb 2026); Meta Platforms (Feb 2026); Sherwood News (Dec 2025)
"Microsoft invests $3.3 billion in Wisconsin data center campus"
At $3.3B for 315 acres, this represents a facility-only investment—powered shell, electrical systems, and cooling. Without knowing the MW capacity, the $/GW is not calculable, but the cost-per-acre of ~$10.5M reflects the land-plus-infrastructure premium in competitive markets. The actual computing hardware deployed inside will represent additional billions in investment. Source: Programs.com (Jan 2026)

Related Analysis

U.S. Data Center Capital Spending Boom 2025–2030
~$3.0 trillion cumulative—the largest private infrastructure cycle in American history
Data Center Types: From Cloud to AI Factory
Training vs. inference vs. cloud—how facility design, cost, and real estate requirements differ by workload type
Data Center Power Demand vs. U.S. Grid Capacity
U.S. grid demand projected to accelerate as data centers drive 426–606 TWh of new load by 2030

Sources

1. Data Center Cost per GW Model (Feb 2026). Derived from McKinsey framework, C&W cost guide, and 17 industry sources.

2. Lawrence Berkeley National Laboratory (LBNL). “2024 United States Data Center Energy Usage Report.” Dec 2024. Baseline: 176 TWh (2023); 50% fleet utilization standard.

3. McKinsey & Company. “Scaling Bigger, Faster, Cheaper Data Centers.” Aug 2025. GW-scale campus design; $20–30M/MW for AI facilities.

4. Cushman & Wakefield. “Data Center Development Cost Guide 2025.” Land: $5.59/SF avg; 224-acre avg parcel; construction cost by market.

5. Dgtl Infra. “How Much Does It Cost to Build a Data Center?” Jun 2024. $600–$1,100/gross SF; $7–12M/MW facility cost.

6. Epoch AI / Asterisk. “Can We Build a Five Gigawatt Data Center?” 2025. $10B facility + $20B chips = $30B/GW.

7. NVIDIA (Jensen Huang). Earnings call, Aug 2025. $60–80B/GW fully loaded AI factory estimate.

8. U.S. Energy Information Administration (EIA). Total U.S. electricity consumption: ~4,000 TWh (2023); avg household: 10,500 kWh/year.

Complete Methodology, Assumptions & Source Citations (23 sources) →

← Previous Power Demand vs. Grid Next DC Metrics Methodology →