The same AI request emits 8.5 times more CO2 in Texas than in Sweden

Discussion overview

Commenters largely accepted the core arithmetic: an AI request inherits the carbon intensity of the grid powering it, so the Texas–Sweden gap says more about fossil generation than some mysterious property of the model. Readers pointed to Sweden’s hydro, nuclear, and wind-heavy supply against Texas’s substantial gas and coal use, while noting that grid emissions also drive much of the water footprint pinned on data centers. Several saw an obvious practical response: schedule latency-tolerant inference where clean power is abundant, or condition data-center growth on clean-energy investment and carbon pricing.

The thread did not settle into simple AI condemnation. One camp argued that per-request comparisons can trivialize a fast-growing workload: autonomous agents, high reasoning settings, and tool use can multiply requests, tokens, and downstream compute dramatically. Others pushed back that the same location-dependent emissions logic applies to every electrical load, and that burgers, fossil plants, and total energy supply offer necessary scale context. A few commenters challenged the broader fixation on individual austerity, while critics replied that avoiding wasteful AI output and protecting ecosystems remain entirely defensible. There was also skepticism about emissions calculators that rely on provider hardware and backend assumptions users cannot independently inspect. The practical consensus was cleaner electricity and better accounting, but not agreement on whether AI demand itself deserves special restraint.

Points of disagreement

The sharpest dispute was whether AI’s footprint should be treated as a distinctive problem or as another consequence of dirty electricity. Critics argued that agentic workflows and expensive reasoning modes can turn tiny per-call figures into enormous demand, including harmful or pointless generated output. The opposing case was that GPUs do not emit carbon on their own: decarbonize the grid, price emissions, and place flexible inference near clean power rather than singling out AI for moral panic.