Back-of-Envelope: Global Compute Equals ~20M H100s; Inference Dominates, Pretraining Far Less Efficient Than Biology
JosephJacks_ · x · 2026-09-05
JosephJacks estimates global compute at roughly 20 million H100-equivalents, with 5% training for 6 months enough to produce a 10-trillion-parameter SOTA model (Astra/Fable level). The vast majority of compute goes to inference — and pretraining is "billions of times less efficient than biology," he claims. Unverified personal estimate.
More from Infra
- NVIDIA launches PAIR, turning idle home PCs into a shared local AI inference pool — techNmak · 2026-09-05
- Nvidia DLSS 5 frame interpolation discussed in Stable Diffusion community — KonoTheSavage1 · 2026-09-05
- Dylan Patel on Dwarkesh: How Elon Musk Played the Compute Market — Dwarkesh Patel · 2026-09-05
- MiniMax and Together AI host London event on the economics of open-model production AI — MiniMax_AI · 2026-09-05
- Base-3 packing for ternary GGUFs: ~22% less weight VRAM, lossless — pmttyji · 2026-09-05
- Perceptron's Multilook API Prefills Video Context Once, Cuts Input Cost to 32% at 16 Prompts — AkshatS07 · 2026-09-05