Assessing the Carbon Emissions and Energy Consumption of U.S. Hyperscale Data Centers

Overview

The proliferation of hyperscale data centres (HDCs) in the US, mainly driven by AI adoption, has raised concerns about this industry's environmental footprint. CO2 was attributed using a generation-weighted average of EPA eGRID2023 plant data within each balancing authority. The fleet used about 68-99 TWh (central 82 TWh, about 1.8% of US demand) and was linked to 37-54 Mt CO2. Facility-load scenarios: u = 0.48 gives 67.7 TWh and 36.9 Mt CO2; u = 0.58 (central) 81.8 TWh and 44.6 Mt; u = 0.663 93.5 TWh and 51.0 Mt; u = 0.70 (AI-weighted) 98.6 TWh and 53.8 Mt.

The HDC electricity-weighted carbon intensity was about 545 gCO2/kWh, about 48% above the US grid average of about 370 gCO2/kWh (2023); about 90% of HDC demand sits in balancing authorities above the national average. Attributed supply mix: 53.9% fossil (gas 38.2%, coal 15.2%), 20.9% nuclear, 25.3% renewables. PJM, the largest HDC region at 32.83 TWh, was 60.0% fossil, 33.1% nuclear and 6.8% renewable, with an attributed intensity of about 535 gCO2/kWh. State-level intensities from the central scenario: Virginia 535, Oregon 440, Iowa 645, Nebraska 740, Wyoming 1035 gCO2/kWh; Virginia had 142 HDCs, using 21.0 TWh and emitting 11.5 Mt.

Average facility power capacity was 39.9 MW (median 36 MW), facility-inclusive power density about 1,600 W/m², and median annual consumption about 209 GWh per facility. PUE-implied check: u = u_IT,measured (about 0.59) × climate PUE / design-contingency R (1.30); climate PUEs (Lei & Masanet) ranged from 1.18 (marine: PACW, CISO) to 1.42 (hot-humid: ERCO), giving u from 0.54 to 0.64 (load-weighted 0.59). Using eGRID2022 instead of eGRID2023 raises attributed emissions by about 3% (for example PJM combustion CI 589 vs 536 gCO2/kWh).

Limits: annual average factors miss the correlation between cooling-driven load swings and hourly grid intensity; the authors note their attributional (average) method is not suited to decisions about where to build new load, which need consequential/marginal methods. The electricity estimate depends on capacity and load-coefficient assumptions. It is a preprint.

Key points

1. Facility-load scenarios: u = 0.48 gives 67.7 TWh and 36.9 Mt CO2; u = 0.58 (central) 81.8 TWh and 44.6 Mt; u = 0.663 93.5 TWh and 51.0 Mt; u = 0.70 (AI-weighted) 98.6 TWh and 53.8 Mt; E = facility capacity × 8,760 × u.
2. The HDC electricity-weighted carbon intensity was about 545 gCO2/kWh, about 48% above the US grid average of about 370 (2023); about 90% of HDC demand sits in balancing authorities above the national average.
3. Attributed supply mix: 53.9% fossil (gas 38.2%, coal 15.2%), 20.9% nuclear, 25.3% renewables; PJM (32.83 TWh) 60.0% fossil, 33.1% nuclear, 6.8% renewable, about 535 gCO2/kWh.
4. State-level intensities (central scenario): Virginia 535, Oregon 440, Iowa 645, Nebraska 740, Wyoming 1035 gCO2/kWh; Virginia had 142 HDCs, 21.0 TWh, 11.5 Mt.
5. Average facility power capacity was 39.9 MW (median 36 MW); facility-inclusive power density about 1,600 W/m²; median annual consumption about 209 GWh per facility.
6. PUE-implied check: u = u_IT,measured (~0.59) × climate PUE / R (1.30); climate PUEs from 1.18 (marine) to 1.42 (hot-humid), u from 0.54 to 0.64 (load-weighted 0.59).
7. eGRID2022 instead of eGRID2023 raises attributed emissions by about 3% (PJM combustion CI 589 vs 536 gCO2/kWh); annual average factors miss the correlation between cooling-driven load swings and hourly grid intensity.

Using it in design

• In fleet or facility energy estimation, use capacity × 8,760 × facility-load coefficient (0.48-0.70, central 0.58) and report the range (67.7-98.6 TWh) explicitly.
• Assess carbon intensity by the balancing authority concerned (PJM ~535, Wyoming ~1035 gCO2/kWh); 90% of HDC demand sits in regions above the national average.
• Do not use attributional average factors for deciding where to place new load; use consequential/marginal methods for siting decisions.

Who should read it

Analysts, policy specialists and planners estimating the energy and carbon footprint of US hyperscale data centres.

Document contents

1. Abstract
2. Introduction
3. Methods (inventory of 403 HDCs and satellite validation, facility-load coefficient scenarios, eGRID2023 generation-weighted attribution)
4. Results (energy consumption, carbon emissions, balancing-authority and state-level intensities, supply mix)
5. Sensitivity and validation (PUE-implied check, eGRID vintage sensitivity)
6. Discussion and limitations

Related standards

• GHG Protocol Scope 2 Guidance

Read the document

Publication details

Authors

Gianluca Guidi; Francesca Dominici; Tiziano Squartini; Callaway Sprinkle; Jonathan Gilmour; Kevin Butler; Eric Bell; Scott Delaney; Falco J. Bargagli-Stoffi

Publisher

arXiv preprint (cs.AI), Harvard T.H. Chan School of Public Health

Year

2026

Document number

arXiv:2606.05420

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.48550/arXiv.2606.05420

Licence

CC BY 4.0 — share with attribution

Licence evidence

arxiv.org

Citation