Zeke Hausfather estimated that his eight weeks of Claude Code use required about 170 kWh of data-center electricity, with a methodological range of 70–330 kWh. His 1,138 typed prompts triggered more than 14,000 model calls and processed 3.2 billion tokens. The figures describe his personal workload, not a universal rate for Claude Code.
What the eight-week estimate says
Hausfather’s central estimate was 170 kWh across his eight-week workload. Depending on the calculation method and assumptions, the estimate ranged from 70 to 330 kWh.
His estimated average was 3.0 kWh per day, with a range of 1.2–5.9 kWh. The heaviest day reached about 11 kWh. Those daily figures also describe Hausfather’s own Claude Code use.
Why one prompt can trigger many model calls
Claude Code can gather context, take actions, use tools and check its results in repeated steps. A typed prompt can therefore lead to multiple model calls as the system works through a task, rather than a single input followed by one response.
What the per-prompt figure means
Dividing Hausfather’s 170 kWh central estimate by his 1,138 typed prompts gives about 150 Wh per prompt. That average belongs to this particular workload: its prompts triggered more than 14,000 calls and processed 3.2 billion tokens.
The token mix also matters. Cache reads—previously processed context reused during a session—accounted for about 96% of tokens and an estimated 66% of electricity under the central assumption that cache reads use 10% of the energy of fresh input. The electricity figures are estimates derived from session activity, so the assumptions used to convert tokens into energy affect the result.
How workload and model choice affect the estimate
The amount of work behind a prompt changes the estimate: a task involving repeated context processing and tool calls can require a very different workload from a short exchange. Hausfather estimated that smaller models used about five to seven times less energy per token than frontier models for simple tasks.
