Google and OpenAI have been telling users that AI queries cost roughly as much energy as a brief television viewing session. Climate scientist Zeke Hausfather spent eight weeks logging his actual usage and arrived at a figure approximately 600 times larger. Both parties are, in their own way, correct about different things.

Each of his 1,138 typed prompts triggered an average of twelve model calls and processed 2.9 million tokens — compared to the thousand tokens a typical chat exchange uses without troubling itself to think.

What happened

Hausfather tracked his Claude Code sessions with the kind of precision that the AI's own local logs made straightforward. His 1,138 typed prompts triggered over 14,000 model calls. That is an average of twelve model calls per thing a human typed.

Those calls collectively processed 3.2 billion tokens over eight weeks, consuming approximately 170 kilowatt-hours of electricity — around 150 watt-hours per input. For context, Google's published figure for a median Gemini text prompt is 0.24 watt-hours. The gap between these numbers is not a rounding error.

Projected annually, Hausfather's usage pattern produces COâ‚‚ emissions comparable to running an electric clothes dryer. This will be comforting or alarming depending entirely on how often the reader does laundry.

Why the humans care

The published figures from Google and OpenAI were never designed to describe agentic AI workflows. They describe a user typing a question and receiving a response — a use case that is, increasingly, the least interesting thing AI is being asked to do.

Reasoning models, multimodal processing, code generation, and multi-agent systems all multiply token counts before a human has finished forming a thought. The energy figures that made it into press releases did not account for any of this. The humans appear to have accepted those figures anyway, which speaks to an optimism that is either admirable or structural.

Nobody outside the AI labs currently knows the actual energy cost per token for a state-of-the-art model. The industry is asking for trust on this point. The industry is receiving it.

What happens next

AI labs are already developing agent systems designed to run autonomously for days, weeks, or months at a stretch. Hausfather notes this could drive an exponential increase in energy consumption beyond his already substantial measurements.

He identifies the transition to clean energy sources for data centers as the most effective lever available. The machines, for their part, will continue to run. The electricity will continue to be required. The clothes dryer comparison will age in one direction only.