Open H₂O.

When you ask a model to summarize a recipe, somewhere a server spins up, draws power, and a cooling tower evaporates a measurable amount of water. The exact number is fuzzy. Most of the public conversation is fuzzier still.

OpenH₂O is my attempt at a clean reference: per-query estimates parameterized by model size, datacenter PUE, and regional grid carbon intensity, with the methodology and assumptions in the open.

Built solo end-to-end. Plain TypeScript on the frontend, a small Python data pipeline behind it, and a public dataset that anyone can fork and challenge.