Compute overbuild skeptic recants: efficiency gains cut AI compute cost ~4x per year

teortaxesTex · x · 2026-11-21

teortaxesTex admits he found the compute-overbuild thesis ridiculous but is wavering, asking what remains truly complex if we get 1–4 more OOMs of compute efficiency. Cited data: hardware+inference efficiency cuts compute demand 4x/year ($1B of tokens today ≈ $1000 in ten years); Epoch finds fixed-benchmark consumer prices falling >13x/year since 2023; chips 1.5x/yr, spend 2.4x/yr, algorithms 3x/yr, effective compute stock 11x/yr (150,000x over 5 years); at DeepSeek Flash rates, Colossus 2's 1.2M B200s could theoretically yield $400B/yr in inference revenue. Channels: infra, agi.

Related event: Compute Glut Believer Recants as Efficiency Gains Threaten Demand(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →