Pathfinder / funder / opportunities / AI energy forecasts may be missing large-scale inference demand
This appears to be a publication/title rather than a standalone organization. The item argues that if AI is deployed at labor-substitution scale, inference energy demand could exceed training-based and current data-center-growth-based estimates, with token use per unit of useful work as a key uncertainty.
No founder records yet.
This EA Forum post (May 2026) presents original, quantified modeling showing AI inference energy demand could dwarf training costs if models require 10M+ tokens per human-workday-equivalent task, a variable currently poorly measured and highly sensitive to labor-replacement scenarios. The author (engineering/energy background, new to AI compute) builds a public Guesstimate model anchoring 8-hour workday-equivalents to METR time-horizon data, revealing that 20% labor substitution by 2035 might consume 9000+ TWh under conservative token-efficiency assumptions. The analysis directly addresses AI-transition infrastructure constraints (power availability, siting, nuclear deals) that current training-focused forecasts appear to underweight, aligning with Funder's interest in underexplored risk vectors and rigorous quantitative work on AI trajectory dynamics.