Data Center Metrics: Methodology & Sources

This page documents the methodology, assumptions, and source citations behind the Data Center Power Utilization Measurement Tool. The calculator translates between power capacity (MW/GW), energy consumption (TWh), cost ($B), and physical footprint using a consistent, source-backed framework. All underlying data and formulas are available in the Data Center Workbook.

1. Power Conversion: GW ↔ TWh

GW (gigawatts) measures instantaneous power capacity—how much a facility can draw at any moment. TWh (terawatt-hours) measures energy consumed over a year—how much it actually draws, integrated over time. The relationship depends on how heavily facilities are utilized.

Core Formula
TWh/year = GW × 8,760 hours/year × Utilization Rate
50%
Utilization
1 GW = 4.38 TWh/yr
60%
Utilization
1 GW = 5.26 TWh/yr
65%
Utilization
1 GW = 5.69 TWh/yr

LBNL explicitly uses 50% average capacity utilization across all U.S. data centers. Their December 2024 report states that this rate, applied to their 325–580 TWh demand range, translates to 74–132 GW of total power demand. The 50% fleet-wide average reflects the full diversity of the installed base: enterprise data centers often run at 30–40% utilization, while hyperscale facilities operate at 60–70%, and AI training clusters reach 80–90%.

Important: Load factor ≠ capacity utilization. Load factor measures actual consumption vs. peak, while utilization measures actual vs. nameplate capacity. If a facility peaks at 80% of nameplate and its load factor is 90%, true utilization is approximately 72% (0.80 × 0.90). The calculator uses utilization in the nameplate sense, consistent with LBNL methodology.

McKinsey Cross-Check: McKinsey projects “>80 GW” producing 606 TWh, implying ~86% utilization—significantly above LBNL’s 50%. The likely explanation is that McKinsey’s “>80 GW” refers to IT load capacity only, while their 606 TWh figure includes total facility consumption (IT × PUE of ~1.3–1.5). Adjusting for PUE of 1.4 yields total facility load of ~112 GW, which produces 606 TWh at ~62% utilization—much closer to LBNL.

2. Utilization Rates by Facility Type

The calculator assigns different default utilization rates to each facility type. These are based on LBNL fleet data and AI-specific research:

Source / Context Utilization Category Notes
LBNL / DOE (Dec 2024) 50% Overall fleet Used to convert TWh ↔ GW fleet-wide. Covers all DC types.
LBNL / DOE (Dec 2024) 80–90% AI servers LBNL report: “AI servers operate at 80–90% utilization.”
LBNL / DOE (Dec 2024) <60% Non-AI servers LBNL: “Non-AI servers often run below 60%.”
SemiAnalysis (Mar 2024) 80% AI GPU clusters Used 80% for DGX H100 cost modeling. 20,480 GPUs = 28–29 MW.
Congressional Research Service (2025) 93% AI training run Study: 8-hour training on 8 GPUs showed 93% avg utilization during active training.
Google architect (via Tom’s Hardware) 60–70% CSP GPU fleet GPU utilization in cloud service providers = 60–70%. Higher rates shorten GPU life.
Introl / industry (2025) 85–96% Well-deployed AI Properly deployed AI systems achieve 85–96%; poorly planned = 40–60%.
Meta / Llama 3 training ~80%* Training cluster 16,384 H100 GPUs; 466 interruptions in 54 days. High util but frequent hardware faults (~9% annualized GPU failure rate).

Calculator Defaults: Traditional / Cloud = 50% (LBNL fleet average); Hyperscale Cloud = 60% (interpolation); AI Inference = 55% (interpolation); AI Training = 65% (workbook recommendation for AI-heavy incremental demand). All defaults are user-adjustable in the calculator’s Advanced Settings panel.

Direction of bias: Higher utilization → fewer GW needed per TWh → less total capex. Lower utilization → more GW needed → more capex. Our 65% AI default is therefore conservative for cost estimation purposes—it produces a lower capex figure than if we used the LBNL 50% fleet average.

3. Cost per GW Model

The cost model synthesizes estimates from 17 sources to build a bottom-up cost-per-GW breakdown. The model distinguishes three cost layers: (1) facility and infrastructure, (2) IT hardware and chips, and (3) optional dedicated power generation. This separation is critical because most industry figures conflate some or all of these layers, which is the primary reason published estimates range from $5B to $80B per GW.

Facility Cost Breakdown (Excluding IT Hardware)

Component Traditional
($B/GW)
AI-Optimized
($B/GW)
% of Facility
(Trad / AI)
Key Sources
Land & Site Preparation $1.0 $1.5 10% / 8% C&W 2025: $5.59/sqft avg; 224-acre avg parcel; hyperscaler sites $5.38/sqft
Core & Shell (building) $1.5 $2.5 14% / 13% Dgtl Infra: powered shell $105–$235/sqft; AI requires reinforced floors for heavier racks
Electrical Systems (PDU, UPS, switchgear, backup gen) $4.5 $7.0 43% / 37% 40–45% of traditional facility cost (multiple sources); AI needs higher-capacity transformers, 2N+ redundancy
Mechanical / Cooling (HVAC, chillers, liquid cooling) $2.0 $5.0 19% / 26% 15–20% traditional; AI liquid cooling (DTC, immersion) adds 2–3× cost; 60–120 kW/rack vs. 5–10 kW
Building Fit-Out (racks, fire suppression, security) $1.0 $1.5 10% / 8% Dgtl Infra: $100–$200/sqft; AI racks $3.9M avg (2025 est.)
Networking & Connectivity (fiber, interconnect) $0.5 $1.5 5% / 8% 10–15% of AI infra capex per SoftwareSeni; low-latency interconnects critical for GPU clusters
Subtotal: Facility (per GW) $10.5 $19.0 100% / 100% Traditional: $7–12M/MW consensus. AI: $20–30M/MW (McKinsey, Vertiv)

IT Hardware / Chips (per GW)

Component Traditional
($B/GW)
AI Factory
($B/GW)
% of Total
(Trad / AI)
Key Sources
GPUs / AI Accelerators $20.0 0% / 43% Asterisk/Epoch AI: ~$17.5–20B chips per GW; NVIDIA H100 ~$25K each; ~100K GPUs per 100 MW
CPUs, Memory (HBM), Storage $3.0 $5.0 19% / 11% HBM at 5× premium over standard DRAM; SK Hynix >50% market share
Servers & Networking Hardware $2.0 $3.0 13% / 6% Custom racks for AI; InfiniBand/RoCE networking for GPU clusters
Subtotal: IT Hardware (per GW) $5.0 $28.0 32% / 60% McKinsey: 60% of $5.2T AI investment = $3.1T for chips/hardware

Grand Totals

Configuration Cost per GW Notes
Traditional DC — Facility + IT Hardware $15.5B Excludes dedicated power generation
AI Data Center — Facility + IT Hardware $47.0B Excludes dedicated power generation
AI DC + Gas Power Plant $50.5B Fully self-powered AI facility

Calculator category mapping: The calculator groups the six facility components into three display categories for the stacked bar chart: Real Estate (Land + Core & Shell + Fit-Out + Networking), Power Infrastructure (Electrical Systems + Cooling), and IT Hardware (GPUs + CPUs/Memory + Servers). For Traditional DCs: $4.0B + $6.5B + $5.0B = $15.5B. For AI: $7.0B + $12.0B + $28.0B = $47.0B.

4. Cross-Check vs. External Estimates

The model is validated against independently published per-GW estimates:

Source Estimate (per GW) Scope & Notes
Bernstein Research (2025) ~$35B AI facility all-in. Cited in Financial Times.
NVIDIA / Jensen Huang (Aug 2025) $60–$80B Fully loaded AI factory. ~$40–50B is NVIDIA chips/systems. Source: Q2 FY2026 earnings call; CNBC (Sep 2025).
Asterisk / Epoch AI (2025) ~$30B $10B facility + $20B chips. Conservative; excludes dedicated power.
Thunder Said Energy (Nov 2025) $40B AI DC total capex including GPUs at $40K/kW = $40B/GW.
McKinsey implied (mid-scenario) ~$33B $5.2T ÷ 156 GW total AI capacity = $33.3B/GW. Includes existing ~31 GW; incremental 125 GW implies $41.6B/GW marginal cost.
CTVC / Sightline (Aug 2025) ~$5.5B Facility only (median tracked investment: $800M at ~$5.5M/MW installed).
ECL TerraSite-TX1 (announced) $8B Facility only. Houston hydrogen-powered campus.

The $47.0B/GW AI estimate sits near the middle of the $30–$80B range. The variation is almost entirely explained by scope: estimates that include only the facility ($5–$19B) differ fundamentally from those that include IT hardware ($30–$50B) or both IT hardware and dedicated power generation ($50–$80B).

5. Why Cost Estimates Vary So Widely

Published data center cost estimates range from $5 billion to $80 billion per GW. This 16× spread is not error—it reflects four distinct sources of variation:

Scope. Some figures include only the facility (core/shell + MEP). Others include IT hardware. A few include dedicated power generation. The McKinsey $5.2T figure includes all three layers. CTVC’s $5.5B/GW covers only facility shell construction.

Facility type. A colocation shell ($5.5M/MW per CTVC) vs. a hyperscaler turn-key ($12M/MW per Digital Realty) vs. an AI-ready liquid-cooled facility ($20M+/MW per McKinsey) are fundamentally different products.

IT hardware cost. The single biggest variable. A traditional DC might have $5B/GW in commodity servers. An AI factory has $20–$30B/GW in GPUs alone. NVIDIA’s market dominance (~90% AI chip market) creates significant pricing power.

Geography. U.S. averages $9.5M/MW; Europe (Frankfurt/London) ~$14M/MW; Asia varies widely. ConstructConnect reports 47% YoY increase in U.S. per-sqft costs as of August 2025.

6. Equivalent Homes Calculation

The calculator converts annual energy consumption (TWh) to an “equivalent homes powered” figure to provide intuitive scale context.

Formula
Equivalent Homes = (GW × 8,760 × Utilization) × 1,000,000,000 ÷ 10,500

Source: The 10,500 kWh/year per home figure comes from the U.S. Energy Information Administration (EIA) Residential Energy Consumption Survey, which reports average annual household electricity consumption. At this rate, 1 GW at 65% utilization (5.69 TWh) powers approximately 542,000 homes.

Industry Rule of Thumb vs. EIA Data: A common industry convention states that 1 GW powers approximately 1 million homes. This rule typically uses nameplate capacity (no utilization haircut) and a lower per-home assumption (~8,760 kWh/year = 1 kW average draw × 8,760 hours). The calculator uses the EIA-sourced figure and applies the utilization rate, producing a more conservative but methodologically consistent result. At 100% nameplate with EIA’s 10,500 kWh, 1 GW powers ~834,000 homes. The difference between ~834K and the “1 million” rule of thumb reflects the gap between EIA’s measured average (10,500 kWh) and the rounder ~8,760 kWh assumption embedded in the rule of thumb.

7. Physical Footprint: Square Footage & Acreage

Building SF and campus acreage per GW vary significantly by facility type due to differences in power density per rack.

Facility Type Building SF per GW Campus Acres per GW Rationale
Traditional / Cloud 8–10M SF 150–400 acres Lower power density (5–10 kW/rack); more building area needed per MW
Hyperscale Cloud 4–6M SF 200–500 acres Higher density than traditional; campus includes substations, cooling, staging
AI Inference 3–5M SF 250–600 acres Higher density racks (20–60 kW); more power infrastructure per building SF
AI Training 2–4M SF 250–1,000+ acres Highest density (60–130+ kW/rack); massive campus for power gen, cooling, redundancy

Key insight: AI facilities are more power-dense per building square foot (less SF needed per GW) but can require more campus acreage because of the land needed for substations, on-site power generation, liquid cooling infrastructure, water treatment, and buffer zones.

Benchmarks from announced projects: Meta Hyperion (Richland Parish, LA) at 5 GW on 2,250 acres = 450 acres/GW with an estimated 4.5M SF of building area (900K SF/GW). Stargate (Abilene, TX) at 1.2 GW on 875 acres = 729 acres/GW. Microsoft Wisconsin at 315 acres for an undisclosed capacity at $3.3B ($10.5M/acre). Sources: Data Center Knowledge (Feb 2026), Data Center Frontier (Aug 2025).

8. U.S. Fleet & Electricity Baselines

The calculator expresses project-level metrics as a percentage of two national baselines:

Metric Value Source Notes
Current U.S. DC fleet consumption 176 TWh (2023) LBNL / DOE, Dec 2024 LBNL report baseline. Used for “% Addition to Current U.S. DC Fleet.”
Total U.S. electricity consumption ~4,000 TWh (2023) EIA Rounded approximation. EIA 2023 actual was 4,178 TWh. ~4% understatement used for round-number context.
Average household electricity use 10,500 kWh/year EIA RECS Used for equivalent homes calculation.
Hours per year 8,760 Calendar constant 365 days × 24 hours. Leap years (8,784) not material.
On the ~4,000 TWh approximation: EIA reports 2023 total U.S. electricity consumption at 4,178 TWh. The calculator uses ~4,000 TWh as a round approximation for contextual comparison purposes, which understates the denominator by approximately 4%. This means the “% Addition to Total U.S. Electricity” figure is slightly overstated. For precision-sensitive analysis, users should reference the EIA figure directly.

9. Dedicated Power Generation Costs

The calculator’s Advanced Settings allow users to add the cost of behind-the-meter (on-site) power generation. These costs are separate from and additive to the facility + IT hardware totals.

Power Source Cost ($B/GW) Source & Notes
Natural Gas (combined cycle) $3.5 Epoch AI / Asterisk: ~$3.5B to build gas plant for 1 GW DC. Most common near-term option.
Nuclear (existing plant restart) $1.9 Microsoft / Three Mile Island: $1.6B for 835 MW restart = ~$1.9B/GW. Limited sites available.
Small Modular Reactor (SMR) $5.0 Estimate; technology still early-stage. Oracle planning >1 GW with 3 SMRs.
Solar + Battery (firm power equivalent) $4.0 Estimate; requires significant overbuilding + battery storage for 24/7 reliability.

10. Complete Source List

The cost model and calculator methodology are derived from the following 17 primary sources, supplemented by project-specific announcements and government data.

# Source Key Data Points Used Date
1 McKinsey & Company
“The Cost of Compute”
$5.2T AI investment by 2030; 15/25/60% split (builders/energizers/tech); 156 GW AI capacity; $1.7T infra (excl. hardware); $300B hyperscaler capex in 2025 Apr 2025
2 McKinsey & Company
“Scaling Bigger, Faster, Cheaper”
$20–$30M/MW for hyperscaler new builds; $4–$7M/MW retrofit; 25 GW (2024) → 80+ GW (2030) U.S.; training vs. inference power density differences Dec 2025 / Aug 2025
3 Cushman & Wakefield
DC Development Cost Guide 2025
Land: $5.59/sqft avg (2024), $5.38 hyperscaler; 224-acre avg parcel (up 144% since 2022); 23% YoY increase for 50+ acre parcels Dec 2024
4 Dgtl Infra
“How Much Does It Cost to Build a Data Center?”
Facility cost: $7–$12M/MW; $600–$1,100/sqft; Electrical 40–45%, HVAC 15–20%, Fit-out 20–25%; Land 15–20%; Digital Realty: $12M/MW, $1,050/sqft Jun 2024
5 Turner & Townsend
DC Cost Index 2025–2026
Global DC construction cost index; $/W benchmarks across 52 markets; 73% of leaders view sector as recession-proof May 2025
6 Bernstein Research ~$35B per GW for AI facility (all-in). Cited in Financial Times, Medium analysis. 2025
7 NVIDIA (Jensen Huang)
Q2 FY2026 Earnings Call
$60–$80B per GW total AI factory cost; ~$40–50B for NVIDIA chips/systems (revised from initial $50–60B estimate). Source: CNBC Sep 2025; Barclays/Benzinga Oct 2025. Aug 2025
8 Asterisk / Epoch AI
“Can We Build a Five Gigawatt Data Center?”
$10B facility + $20B chips = $30B per GW; $3.5B for dedicated gas power plant per GW 2025
9 Thunder Said Energy $10M/MW traditional capex; $40K/kW ($40B/GW) for AI including GPUs; 135 GW global capacity (2025) Nov 2025
10 ConstructConnect YTD 2025 DC starts: $40B through Aug; avg project cost $499M; avg $/sqft up 47% YoY; $977/sqft Nov 2025
11 Deloitte
“Can US Infrastructure Keep Up with the AI Economy?”
4 GW AI DC in 2024 → 123 GW by 2035; 50,000-acre campuses at 5 GW planned; grid stress #1 challenge Jun 2025
12 Vertiv
“Cost Impact of AI Data Center Design”
$9–$10.50 per watt of IT capacity in N. America; AI power density 2–4 kW → 40+ kW/rack 2024
13 CTVC / Sightline
“The Data Center Report”
Median tracked DC investment: $800M (~$5.5M/MW installed); AI facilities 5–6× more capital intensive Aug 2025
14 ECL (TerraSite-TX1) 1 GW off-grid hydrogen campus near Houston: $8B total; Phase 1: 50 MW for $450M 2024
15 BlueCap Economic Advisors Electrical 40–45% of total; HVAC 15–20%; land 15–20%; AI >$20M/MW Jul 2025
16 Uptime Institute Tier III enterprise: ~$12M/MW in 2010; costs rising post-pandemic; MEP 30–50% of budget Mar 2023
17 SoftwareSeni GPU procurement 35–45% of AI infra capex; $20B chips + $10B facility per GW; networking 10–15% Nov 2025

Additional Sources (Baselines & Context)

# Source Data Used Date
18 Lawrence Berkeley National Laboratory (LBNL) / DOE
“2024 United States Data Center Energy Usage Report” (LBNL-2001637)
176 TWh baseline (2023); 325–580 TWh by 2028; 50% fleet utilization; AI servers at 80–90% Dec 2024
19 U.S. Energy Information Administration (EIA) Total U.S. electricity consumption: ~4,178 TWh (2023); avg household: 10,500 kWh/year 2024
20 SemiAnalysis (Mar 2024) GPU cluster cost modeling at 80% utilization for 20,480 GPU configurations Mar 2024
21 Congressional Research Service 93% average GPU utilization during 8-hour AI training runs 2025
22 Meta / Llama 3 Training Report 16,384 H100 GPUs; 466 interruptions in 54 days; ~9% annualized GPU failure rate; ~80% effective utilization 2024
23 McKinsey & Company
“How Data Centers and the Energy Sector Can Sate AI’s Hunger for Power”
606 TWh U.S. by 2030 (implied baseline: 147 TWh, 2023); $6.7T global capex through 2030 ($5.2T AI-related) Sep 2024

11. Data Center Workbook

All underlying data, formulas, sensitivity tables, and source documentation are available in the Data Center Workbook (13 tabs). Open in Google Drive and download a copy to explore the data and modify assumptions.

Workbook tabs include: PowerGapData, GW ↔ TWh Conversion, US TWh Forecasts, McKinsey Reconciliation, Utilization Research, US Share Research, Cost per GW Model, McKinsey $5.2T Framework, Cost Translation Sources, Cost Translation Methodology, and Source Detail & Citations.

Related Analysis

← Back to Data Center Power Utilization Measurement Tool

U.S. Data Center Capital Spending Boom 2025–2030

Data Center Power Demand vs. U.S. Grid Capacity

Data Centers Landing Page