New Gas-Powered Data Centers Could Emit More Greenhouse Gases Than Whole Nations

· ai · Source ↗

TLDR

  • WIRED reviewed air permits for 11 US data center campuses tied to OpenAI, Meta, Microsoft, and xAI, finding combined potential emissions of 129 million tons of CO2 equivalents per year.

Key Takeaways

  • Behind-the-meter gas power is the core mechanism: developers bypass grid queues by building private plants that run near-constant load, unlike grid-connected plants that cycle down.
  • Permitted emissions are worst-case maximums, but energy researcher Jon Koomey argues data centers will run closer to permitted limits than traditional plants since demand is flat and continuous.
  • Turbine scarcity caused by the AI buildout is pushing some operators toward less efficient models, compounding emissions per unit of compute.
  • Stargate-linked gas projects in Texas and New Mexico alone could emit over 24 million tons annually; the Fermi campus near Amarillo is permitted for 40.3 million tons across two plants.
  • Major hyperscalers are simultaneously making emissions-reduction pledges: Meta’s three Ohio behind-the-meter plants at half capacity would offset over 10% of its stated four-year emissions reductions.

Hacker News Comment Review

  • Commenters split sharply on the comparison framing: the Morocco/Norway/Connecticut benchmarks were called meaningless without per-unit-of-economic-output context, with one commenter calculating the full list at roughly 1.9% of US emissions and arguing productivity gains could net-reduce carbon intensity.
  • A structural observation gained traction: inference workloads are actually well-suited to intermittent renewables since latency tolerance is high, making the gas buildout an economic and regulatory failure rather than a technical necessity.
  • The political economy angle dominated more than the emissions math: low-opportunity regions mute local resistance, the current administration limits regulatory bandwidth, and companies face no carbon pricing signal to favor intermittent clean sources.

Notable Comments

  • @usrusr: argues AI inference could overbuild on intermittent renewables by design, making the gas choice a market-structure failure, not an engineering one.
  • @josefritzishere: “Our species is going to go extinct over samurai cat memes.”

Original | Discuss on HN