AI agents waste 96% of their energy re-reading their own memory

AI coding agents may consume as much electricity as two household refrigerators running around the clock — and nearly all of that energy goes toward work the user never sees. That's the finding from climate researcher Zeke Hausfather, who spent eight weeks logging every token his AI assistant processed and discovered a structural inefficiency baked into how these tools operate.

The hidden loop

Hausfather tracked his use of Anthropic's Claude Code over eight weeks, recording 3.2 billion tokens processed and roughly 14,000 model calls. Total estimated energy consumption: around 170 kWh, though his own methodology puts the uncertainty range at 70–330 kWh. Daily use swung between 1.2 and 5.9 kWh — the upper end comparable to two standard fridges running all day.

The breakdown is where things get striking. According to The Climate Brink (Hausfather), 96% of all tokens processed were cached reads — the agent re-reading its own accumulated context at each of those ~14,000 steps. Every time it runs a command, reads a file, or calls a tool, it loops back through everything it already knows. Only 0.4% of tokens produced output the user actually sees.

Cached tokens are cheaper to process than fresh ones, but the sheer volume of re-reads still drives the majority of energy cost. Hausfather compares it to someone re-reading a book from page one before writing each new sentence.

The gap between benchmarks and reality

Google claims a standard Gemini text prompt costs around 0.24 Wh. OpenAI's Sam Altman has cited 0.34 Wh per query for ChatGPT. Hausfather's agent workload works out to roughly 600 times higher per interaction — a gap that exists because published benchmarks describe single prompts, not the looping, multi-step processes that agentic tools actually run.

Anthropic has not published per-call energy figures, so all estimates here rely on published token pricing rates and researcher inference — not official disclosure.

On an annualized basis, Hausfather puts the CO2 footprint of his usage pattern at roughly equivalent to an electric clothes dryer, or about half the carbon footprint of driving an EV 18,500 km in California.

Disputed, but not dismissed

The numbers aren't settled. AI sustainability researcher Boris Gamazaychikov called the analysis a "good attempt" but flagged use of outdated data, with his own competing study expected later in 2026. That study could shift the picture significantly in either direction.

What the research does clarify is that the energy debate around AI has mostly been fought on the wrong battlefield. Single-prompt efficiency figures tell consumers little about what agentic workloads actually cost — and as agent use scales up across enterprise software, that gap is only going to widen.