AI adoption in US adds ~900k tons of CO₂ annually, study finds

Original Article ↗ Hacker News Discussion ↗

Scale of AI CO₂ emissions

  • The cited 900k tons/year is framed as ~0.02% of U.S. emissions; several commenters see this as non-trivial but small in national context.
  • Others argue the study likely underestimates: they cite newer estimates of AI data center use (tens of TWh/200+ PJ in 2024) that would imply emissions one–two orders of magnitude higher than the paper’s 28 PJ projection.

Comparisons and framing

  • Many criticize headlines that give an absolute number without context, calling it misleading or agenda-driven.
  • Comparisons are made to household/car emissions, air travel (1B tons/year), streaming video, lawn equipment (30M tons in the U.S.), and gaming GPUs—often to argue AI’s share is relatively small.
  • Some say the meaningful comparison class is other industrial/commercial uses (shipping, metals, etc.), not households.

Methodology and data quality

  • The paper is described as “guesses multiplied together,” relying on:
    • old energy and data-center baselines (2016–2019),
    • GPT‑3-era inference assumptions, and
    • speculative adoption/productivity gains per industry.
  • Critics question assuming national-average grid carbon intensity and note the model’s low implied power (0.9 GW) conflicts with known individual AI projects (e.g., single 4.5–10 GW facilities).

Energy mix and infrastructure debates

  • Long thread on fossil vs solar/wind vs nuclear:
    • Fossil fuels seen as extremely energy-dense but with unacceptable atmospheric side effects.
    • Solar is abundant but diffuse; transmission and storage (hours vs needed weeks) are cited as major cost and reliability problems, using Australia/South Australia as examples.
    • Some push nuclear for firm capacity; others argue “baseload” is ill-defined in highly renewable grids.
  • Short-term, AI data centers often use gas (including dedicated plants), raising concerns about local pollution, “mobile” turbine loopholes, and price impacts.

CO₂, efficiency, and rebound

  • Some argue AI saves time and thus emissions (faster search, coding, document conversion).
  • Others counter that under current economic incentives, higher productivity tends to increase total energy use; the hours “saved” are reallocated to other activity, adding net CO₂ (a Jevons-paradox-style view).

Economic and social impacts of data centers

  • Concerns: cost pass-through to residents, parallels to crypto mining, and minimal local employment vs large capex.
  • Counterpoints: new infrastructure, tax base, and job creation justify projects; if AI proves profitable, market risk falls on companies. Skeptics compare the boom to the dot-com bubble.