Research 2026-08-10

A climate scientist logged eight weeks of his own AI coding assistant. One prompt cost about 600 chatbot questions' worth of electricity

Zeke Hausfather, a climate scientist who has spent years correcting bad estimates of AI's energy use, decided to measure his own. He pulled the token-level logs from his Claude Code sessions between 31 May and 25 July and added them up. The 1,138 prompts he typed in that period triggered more than 14,000 separate model calls and processed 3.2 billion tokens, which he estimates at roughly 170 kilowatt-hours of data centre electricity — a range of 70 to 330 depending on how cache reads are counted. That works out to about 150 watt-hours per prompt he actually typed, against published figures of 0.24 to 0.34 watt-hours for a single chat message. Hence the 600-fold gap. The absolute numbers are less alarming than the ratio: a full year of that intensity comes to about 1.1 megawatt-hours and roughly 370 kilograms of CO₂ on the average US grid, which Hausfather compares to running an electric clothes dryer for a year. He is explicit that the largest uncertainty is the energy cost of reading cached tokens, which no company publishes; he assumes 10 per cent of a fresh token and shows the result across a 1 to 25 per cent range. The estimate is one person's usage, not an industry average.

Why it mattersAlmost every public figure for AI's energy footprint — including the reassuring ones — is built on the cost of a single chat message. That number is now measuring the wrong thing. When you type one instruction to a coding agent, it goes away and makes a dozen or more model calls on your behalf, reads files, runs tools, and reasons through the results, and you are billed for none of that mentally. The gap between what a user perceives as one question and what the data centre actually does has become enormous, and it is growing as agents get more autonomous. Two honest conclusions follow, and they point in opposite directions. Individually this still is not much: a year of heavy agent use costs about what a household appliance does, so nobody needs to feel guilty about a prompt. Collectively, it means the standard estimates of AI's total electricity demand are anchored to a unit of work that is quietly becoming obsolete, and the people building data centres are already planning for the number this article implies rather than the one in the headlines.

#Coding #AI Agents

✓ Verified · 3 sources

Read in the app — free, in 9 languages

Machine learning read the shape of sick brain cells and picked out nine already-approved drugs that calmed them down
2026-08-21 Rumour: the anonymous model that just topped a coding benchmark, for free, is said to be Zhipu's unreleased flagship
2026-08-21 DeepSeek's cheap workhorse can now see — and on agent tasks that need eyes it says it is close to Anthropic's best
2026-08-21 Nvidia is paying $6 billion for the machine that builds a rival's models — and hiring 109 of the people who ran it
2026-08-21 Given four hours and a GPU to improve the way AI is trained, the best agent scored 0.25 out of 1 — and most never tried
2026-08-21

A climate scientist logged eight weeks of his own AI coding assistant. One prompt cost about 600 chatbot questions' worth of electricity

Related stories