Terrestrial Non-Energy Operating Expenditure

What is the annual non-energy operating expenditure per kW_IT for hyperscale AI data centers?

What is the annual non-energy opex per kW_IT for hyperscale AI data centers?

Answer

Non-energy opex for hyperscale AI data centers ranges from $500/kW_IT/year (optimistic) to $1,100/kW_IT/year (conservative), with a central estimate of $750/kW_IT/year. This covers staffing, facility maintenance, property tax, insurance, water, security, and connectivity — everything except electricity and capital amortization.

At the central estimate, non-energy opex is ~8-10% of total terrestrial TCO, making it a secondary but non-negligible cost component. The wide range reflects differences in labor markets, tax jurisdictions, automation levels, and whether the facility requires specialized AI-optimized maintenance (liquid cooling systems, high-density power distribution).

Analysis

Component breakdown

Non-energy opex decomposes into several categories, estimated for a 100 MW hyperscale AI facility:

Key considerations

Staffing is the largest single component. Industry benchmarks show a 12 MW facility requires ~20 staff and a 40 MW facility ~45 staff, but hyperscale facilities over 100 MW achieve much lower FTE-per-MW ratios through automation and standardized systems broadstaff-dc-staffing-levels.1. At the largest scales, operational staffing ranges from ~0.3-1.0 FTE per MW. AI facilities require specialized cooling system technicians but offset this with fewer tenant management functions.

Maintenance scales with infrastructure capex. Thunder Said Energy characterizes maintenance as ~40% of total opex thunder-said-dc-economics.3. JLL reports facility capex of $10.7-11.3M/MW jll-2026-dc-outlook.1. At the standard industry benchmark of 2-4% of facility capex per year (not from any single source cited here), a $12,500/kW_IT facility incurs $250-$500/kW_IT/year. Our central estimate of $250 reflects hyperscale efficiency and newer equipment with lower maintenance needs.

Property tax varies enormously by jurisdiction. Thirty-seven states offer data center tax incentive programs, with abatements typically lasting 10-20 years (up to 30 years in Alabama) [ncsl-dc-incentives.1, abitos-dc-tax-incentives.1]. Five states (Alabama, Iowa, Montana, Nevada, Oklahoma) explicitly offer property tax relief; Nevada provides up to 75% personal property tax abatements for 10-20 years, and Minnesota offers a permanent property tax exemption on equipment. Without abatement, property tax on a $12.5M/MW facility at a 1-2% rate would be $125-$250/kW_IT/year. With typical abatements, effective rates are much lower.

Water costs are declining as the industry shifts from evaporative cooling towers (which consume 1.8L per kWh) to closed-loop liquid cooling systems. AI-optimized facilities with direct-to-chip liquid cooling use minimal water.

Scenario rationale

Optimistic ($500/kW_IT/year): Reflects highly automated hyperscale in a tax-abated jurisdiction with modern liquid cooling. Texas-type economics: low labor costs, no state income tax, property tax incentives.

Central ($750/kW_IT/year): US average hyperscale with moderate tax burden, standard automation, liquid cooling. Consistent with the widely-cited 8-12% non-energy opex share of total TCO.

Conservative ($1,100/kW_IT/year): High-cost jurisdiction (Northeast US, California) with full property tax, premium labor markets, older facility requiring more intensive maintenance, and capacity charges for grid-connected facilities adding ~$100/kW_IT/year.

Validation

At $750/kW_IT/year central, non-energy opex is 7.0% of central terrestrial TCO (~$10,800/kW_IT/year). This is consistent with the Catalyst observation that "energy is only 5 to 15% of an AI focused data center" catalyst-scaling-pathways.1 — if energy is ~8% and GPU depreciation is ~74%, that leaves ~18% for infrastructure amortization (~11%) and non-energy opex (~7%).

Evidence

1. Catalyst podcast notes that "energy is only 5 to 15% of an AI focused data center and chips and maintenance are the rest," implying non-energy costs (capital amortization plus operating costs) account for 85-95% of total cost. Since GPU depreciation alone is ~70-75% of TCO, non-energy opex must be roughly 10-20% of total. — catalyst-scaling-pathways

2. Alpha Matica models a 100 MW hyperscale data center with total CapEx of $3.4B-$5.5B ($34-55/W including IT hardware), with infrastructure-only CapEx of $900M-$1.5B ($9-15M/MW). The article identifies energy as the top opex component but does not provide specific annual opex figures. — alpha-matica-dc-cost-structure

3. Thunder Said Energy models data center opex as "dominated by maintenance (c40%), electricity (c15-25%), labor, water, G&A and other." — thunder-said-dc-economics

4. McKinsey "Beyond compute" estimates total data center investment of $6.7T through 2030, with 25% ($1.3T) going to power generation, transmission, cooling, and electrical equipment. — mckinsey-beyond-compute

5. Broadstaff Global reports hyperscale data center staffing benchmarks: small facilities (1-5 MW) require 8-15 operational staff, medium facilities (5-20 MW) require 15-35 staff, and large facilities (20+ MW) require 35+ staff. A 12 MW facility requires approximately 20 full-time staff; a 40 MW facility employs around 45. Hyperscale data centers over 100 MW operate with fewer people per megawatt due to automation and standardized systems. — broadstaff-dc-staffing-levels

6. The National Conference of State Legislatures identifies five states (Alabama, Iowa, Montana, Nevada, Oklahoma) that explicitly offer property tax relief to data centers. Broader data center tax incentive programs exist in 37 states as of 2025. — ncsl-dc-incentives

7. AbitOs reports that data center tax incentives "typically last between 10 and 20 years," with Alabama offering extensions for up to 30 years. Property tax abatements are "reductions or deferrals on property taxes for data centers, including their buildings and equipment" and are generally administered at the local level, varying widely even within a state. Nevada offers up to 75% personal property tax abatements for 10-20 years; Minnesota offers a permanent property tax exemption on equipment. — abitos-dc-tax-incentives

8. SemiAnalysis reports PJM capacity prices jumped 9.3x from $29/MW-day (2024/25) to $270/MW-day (2025/26), with the 2027/28 auction hitting the federal price cap of $329/MW-day. PJM's Internal Market Monitor found removing all data centers from forecasts would have reduced total capacity payments by $9.33B (64%). This capacity charge represents a non-energy operating cost that scales with power consumption, adding ~$100-$120/kW_IT/year for PJM-connected facilities. [Note: the ~$100-$120/kW_IT/year figure is a derived calculation from the $270-$329/MW-day range, not a figure stated in the source. The previous version incorrectly stated data centers drove "40% of PJM capacity costs in the 2027/28 auction" — the 40% figure is from a different source (volts-pjm-explainer, citing the IMM); the SemiAnalysis article reports a 64% figure for the 2025/26 auction.] — semianalysis-pjm-bills

9. JLL reports shell-and-core costs rose from $7.7M/MW (2020) to $10.7M/MW (2025), a 7% CAGR, with a forecasted 6% increase to $11.3M/MW in 2026. — jll-2026-dc-outlook

10. Cushman & Wakefield's 2025 U.S. Data Center Development Cost Guide surveys costs across 19 US markets, ranging from $9.3M/MW (San Antonio) to $15M/MW (Reno), with Texas markets consistently lowest cost. — cushman-wakefield-dc-cost-2025