AI agents use 600× more energy than chat prompts, study finds

Climate scientist Zeke Hausfather undertook a detailed analysis of his own AI usage over eight weeks to measure the real energy cost of AI agent systems, challenging the low figures published by tech companies. Using Claude Code, which stores complete local logs of every session including exact token counts, Hausfather tracked 1,138 typed prompts that triggered over 14,000 model calls—an average of twelve calls per prompt. Each prompt processed an average of 2.9 million tokens, far exceeding the roughly one thousand tokens typical of a simple chat exchange. His total usage over the eight-week period processed 3.2 billion tokens, of which a striking 96 percent were cache reads, meaning the agent re-read its entire accumulated context at each of the 14,000 steps. The text output Hausfather actually saw on screen accounted for just 0.4 percent of all tokens processed.

Hausfather estimated his total data center electricity consumption at 170 kWh over eight weeks, with an uncertainty range of 70 to 330 kWh. This translates to roughly 150 Wh per prompt—about 600 times the median Gemini chat prompt's 0.24 watt-hours and far exceeding OpenAI CEO Sam Altman's estimate of 0.34 Wh for an average ChatGPT query. To ground this in everyday terms, Hausfather's median Claude Code session consumed 0.6 kWh, fifty times the electricity needed to charge a phone. His average day hit 3.0 kWh (range: 1.2 to 5.9 kWh), more than two refrigerators' daily draw. On his most intensive day, when several parallel agents ran an extensive geodata analysis, he estimates consumption reached 11 kWh—more than one-third of an average U.S. household's daily electricity use.

Scaled to a full year, Hausfather's agent-based usage would consume about 1.1 MWh of data center electricity (0.4 to 2.2 MWh uncertainty range), roughly one-tenth of what an average U.S. household uses annually. Based on the average U.S. electricity mix, that corresponds to about 370 kg of CO₂ equivalents per year—slightly more than running an electric clothes dryer for a year and about half of a round-trip economy flight from San Francisco to New York. Hausfather contextualizes this as simultaneously a large emissions source and a modest portion of his total carbon footprint—roughly two percent of an average American's yearly emissions and about eight percent of annual emissions from a typical U.S. gas-powered car.

Hausfather's core argument rejects both guilt-driven restraint and technological pessimism. Personal sacrifice by the small group of heavy AI users "is not going to bend any curves," he writes, though he notes that routing simple tasks to smaller models (which use five to seven times less energy per token than frontier models) does make practical sense. The decisive lever, he argues, is the carbon intensity of the electricity grid itself. If the same workload ran on largely clean power, the carbon footprint would drop by about 90 percent. He identifies a policy opening: AI companies bring enormous capital and unusual urgency to infrastructure decisions. If that investment flows into clean energy, grid expansion, and advanced technologies like geothermal or nuclear power, the AI boom could leave the grid cleaner than it found it. Nearly three-quarters of planned on-site power generation for U.S. data centers runs on natural gas, leaving substantial room for that reallocation.