Climate scientist Zeke Hausfather tracked eight weeks of his own Claude Code usage and found that agentic AI consumes roughly 600 times more electricity per prompt than a typical chat query.
His tally from May 31 to July 25, 2026: 1,138 typed prompts triggered more than 14,000 model calls and 3.2 billion tokens — about 170 kilowatt-hours of data center electricity (range 70 to 330). That works out to roughly 150 watt-hours per prompt, about 600 times (250 to 1,200) the energy of a median chat prompt. About 96% of the tokens were cache reads, where the agent re-reads its own context at every step; the visible text was only about 0.4%. A median session consumed around 0.6 kWh — fifty times the energy of charging a phone — while his average working day used about 3.0 kWh, more than two refrigerators.
Hausfather stresses that the reassuringly small per-prompt figures often quoted (about 0.24 Wh for a Gemini prompt, 0.34 Wh for ChatGPT) are not wrong, but increasingly divorced from how AI is actually used: the fastest-growing form of AI use is agentic. The Watershed white paper (Bistline et al. 2026) similarly shows electricity per AI task spanning more than five orders of magnitude, with agentic workflows at 50 to 500 Wh. As chips get more efficient but usage grows faster (the Jevons paradox), the carbon intensity of the electricity is what matters most — and much of the planned U.S. data center capacity intends to run on behind-the-meter natural gas generation.
