Ex-OpenAI Exec Slams Goldman Sachs Token Demand Forecast, Cites 100x Cost Drop

ChrSzegedy · x · 2026-08-05

Goldman Sachs recently projected that global token consumption will surge from 5 quadrillion currently to 120 quadrillion by 2030 (a 24x increase), arguing this will significantly boost hyperscaler earnings.

Former OpenAI core researcher Christian Szegedy strongly refuted this prediction. He pointed out fundamental flaws in the forecast: LLM inference compute costs will decrease dramatically (by at least 100x), while available compute will increase by at least 10x. Furthermore, he argued that "token" is an ill-defined and terrible unit of measurement for assessing future compute economics.

Original post →

More from AGI Musings

AGI Musings channel →