Particle.news

Harvard Study Finds U.S. Hyperscale AI Data Centers Use State‑Scale Power and Emit Denmark‑Size CO2

The peer‑reviewed paper shows about half of their electricity came from fossil fuels, prompting COP31 negotiators to launch a voluntary clean‑power and transparency pledge.

Overview

  • The Harvard‑led paper published in Nature Sustainability on Monday found 403 U.S. hyperscale data centers used roughly 68–99 terawatt‑hours of electricity and emitted about 25.7 million metric tons of CO2 for the May 2024–April 2025 period.
  • Researchers built a facility‑level inventory using satellite imagery, permit filings and EPA grid data because operators generally do not disclose site‑level power use or emissions.
  • The study estimates about 54% of attributed electricity came from fossil fuels, with roughly 38% from natural gas and 15% from coal, and shows regional grid mixes in places like Virginia, Oregon, Ohio and Iowa strongly affect carbon intensity.
  • Voters approved 40 of 41 local ballot measures to restrict data‑center growth on Nov. 3, but industry and analysts say most permitted or under‑construction projects will likely proceed while developers shift to on‑site generation or new siting strategies.
  • Federal measurement is nascent: the EIA completed three pilot surveys in 2026 and is designing a voluntary sector survey, and international negotiators at COP31 unveiled the Antalya Artificial Intelligence Pledge urging cleaner power and better disclosure.