AI data centers will materially raise U.S. electricity demand and associated emissions where marginal generation is fossil-fuel–dominated, while creating regionally varying pressures on water, noise and land; energy-efficiency, cooling changes and operational strategies can reduce harms but only if aligned with electricity system, water availability and land-use policies.

Main Finding

AI-specialized data centers materially change the environmental and economic footprint of the U.S. data center sector, but their impacts are determined less by facility design alone and more by the electricity, water, and land-use systems in which they are embedded. Electricity consumption (and thus marginal generation mix and transmission constraints) drives the dominant climate and air-quality impacts; local effects (water use, noise, land-use change, backup diesel emissions) vary strongly by location and infrastructure context. Technological and operational mitigations (efficiency, cooling choices, timing) can reduce impacts, but their effectiveness is system-dependent, and current regulatory/market frameworks often fail to internalize these location- and time-specific externalities.

Key Points

• Scale and growth
  • AI workloads (training/inference on GPU/accelerator clusters) increase power density and heat load relative to traditional workloads.
  • Projections (Shehabi et al., EPRI × EIA scenarios) suggest U.S. data centers could account for roughly 6.7–12% of U.S. electricity consumption by 2028 and 9–17% by 2030, implying tens of GW of concentrated new load.
• Emissions and electricity system dependence
  • The dominant source of GHGs and many air pollutants from data centers is electricity consumption (indirect emissions), so impacts depend on the marginal generators that meet incremental load (not system-average mixes).
  • Coal and diesel (backup gens) have high marginal pollutant intensities; natural gas generally has lower emission intensities but outcomes are sensitive to methane leakage.
  • The authors emphasize marginal emission factors (time- and location-varying) as the appropriate metric for incremental impacts.
• Backup power and local air quality
  • Diesel generators used for backup/testing produce localized NOx, PM, CO and other pollutants; these contribute to community exposures and are a non-trivial local externality.
• Water and cooling trade-offs
  • Cooling choices shift water vs electricity trade-offs: evaporative/wet solutions raise on-site water use (high WUEsite); dry cooling reduces water use but increases electricity consumption. Liquid cooling adoption for AI clusters increases on-site heat rejection needs and can raise water demand in some designs.
  • Water impacts are regionally heterogeneous and depend on local water availability and regulatory constraints.
• Noise and community/ecological impacts
  • Continuous cooling equipment noise and intermittent generator testing produce community noise (reported "hum"/drone); measured site values can reach levels associated with annoyance and sleep disturbance. Systematic exposure studies are limited.
• Land-use and infrastructure pressures
  • Large, concentrated loads affect transmission needs, local grid reliability, and land-use patterns; siting decisions feed into transmission expansion and tariffs for large customers.
• Mitigation and policy
  • Efficiency, advanced cooling, operational timing (time-shifting, demand response), and cleaner marginal generation can reduce impacts, but their effectiveness depends on grid mix, transmission constraints, and local resources.
  • Current regulatory and market frameworks (permits, tariffs, planning) often do not capture time/location-specific externalities from data center loads; integrated policy approaches are recommended.

Data & Methods

• Data sources
  • Electricity consumption and projections: Shehabi et al. (2024), EIA (2025–2026), EPRI (2024–2026).
  • Emission factors and pollutant intensities: EPA (various years), EIA (2020a), NETL (2011), IPCC/Forster et al. for GWP.
  • Heat rates and generation performance: EIA (2026b) to convert fuel-based emissions to electricity-basis (kg/GJ → kg/GJ_e).
  • Noise and exposure metrics: EPA (1974), WHO (2018), OSHA/NIOSH (occupational limits), site measurements from literature (Miljković, Hamilton, JLARC, Gour et al.).
  • Water-use and WUE literature: Lei et al. (2025), Li et al. (2025), National Academies (2025).
• Analytical approaches
  • Emphasis on marginal rather than average emission factors to estimate incremental emissions from added demand; marginal factors are treated as time- and location-sensitive because incremental load is typically met by specific marginal generators.
  • Conversion of fuel-based pollutant and GHG factors to electricity basis via representative heat rates and plant efficiencies (coal steam, gas combined cycle, diesel IC assumptions).
  • GWP evaluated on 20-year and 100-year horizons to show sensitivity of climate impact rankings (coal > diesel > natural gas in CO2e terms, with methane making natural gas relatively closer).
  • Use of scenario projections (applying data center share estimates to EIA projected electricity sales) to bound future electricity demand from data centers.
  • Noise assessment using standard metrics (Leq, Lden, Lnight) and occupational limits; water analysis using site- vs source-level WUE distinctions and cooling-technology trade-offs.
• Limitations noted by authors
  • Many local impacts (noise, ecological effects, site-specific water constraints) lack systematic, large-sample measurement studies; results are therefore contextual and sensitive to local infrastructure and regulatory settings.
  • Effectiveness of mitigation depends on grid dispatch patterns and marginal mixes, which vary across regions and over time.

Implications for AI Economics

• Externality valuation must be spatially and temporally granular
  • Social cost calculations for AI data center electricity should use marginal emission factors (location/time-specific) rather than national averages; aggregate estimates can misstate true climate and local air-pollution externalities.
• Siting and locational comparative advantage
  • Economic costs and environmental impacts of AI data centers differ by location because of local grid mix, transmission constraints, water availability, and land-use conflicts. Optimal siting should internalize these heterogeneities (affects firms' site selection and social planner decisions).
• Grid and market consequences
  • Concentrated large loads can change generation investment and transmission planning needs; utilities and regulators must account for load-induced changes in peak demand, capacity value, and locational marginal prices.
  • Time-varying tariffs, demand charges, and locational pricing can better align private incentives (operation/scheduling of AI workloads) with social costs.
• Operational and procurement strategies
  • Firms can reduce external costs through operational measures (time-shifting training runs to low-emission hours), advanced cooling (trade-off water vs electricity), and procurement (PPAs, storage) — but the private value of these measures depends on market design (price signals, penalties/credits for emissions).
• Regulation of backup and local impacts
  • Localized diesel backup emissions and noise are potential sources of community-level externalities; stricter permitting, emissions controls, or incentives for lower-emitting backup (batteries/hydrogen) could lower local damages but will change capital and operating cost trade-offs.
• Investment and policy design
  • Policymakers should adopt integrated policies that coordinate electricity system planning, water resource management, land-use/zoning, and environmental permitting to avoid cost-shifting and negative local outcomes.
  • Markets/regulations that fail to price location- and time-specific externalities can lead to misinvestment (e.g., siting loads where marginal generation is carbon-intensive), under-provision of grid upgrades, and higher social costs.
• Research and data needs
  • Better empirical measurement of marginal emission factors at fine temporal/spatial resolution, systematic noise and water-use exposure studies, and modeling of how large flexible loads interact with dispatch and investment would improve economic assessments and policy design.

Key takeaway: economics of AI data centers cannot be separated from electricity-system and local-resource contexts—pricing, siting, and operational policies must be designed to reflect the time- and place-specific externalities these facilities impose.