Only 7% of frontier lab compute to pretraining by Q4 2026, but the viral chart is being misread
repligate · x · 2026-10-14
A viral SemiAnalysis chart claims only 7% of frontier lab compute now goes to pretraining, with post-training/RL surging to 55% and inference at 38%. repligate flagged ambiguity: is 7% per-checkpoint or per-lab total, given labs iterate and abandon many post-training runs? voooooogel clarifies the figure comes from SemiAnalysis' Tokenomics hardware-demand forecast covering whole labs, not RL compute on any specific model — pretraining share fell from 67% in 1Q24 to an estimated 7% in 4Q26.
More from Infra
- TokenSpeed serves Qwen3.8-Flash-Next at 685 tokens/s per user on 4x NVIDIA GB300 with NVFP4 — Alibaba_Qwen · 2026-10-14
- $12K of RAM Beats $20K of GPUs: Running 50+ Frontier Coding Agents in Parallel — RishabJainK · 2026-10-14
- Running Coding Agents Locally: RAM Costs Now Exceed Token Costs — RishabJainK · 2026-10-14
- 14-year-old GPU rigid body pipeline revived in CUDA simulates 2.05M dominoes — erwincoumans · 2026-10-14
- AI agents hit monorepo scale: you need Google-style copy-on-write primitives, not Git — rickasaurus · 2026-10-14
- Modal trains RL models on mixed-accelerator clusters, matching homogeneous setups at lower cost — charles_irl · 2026-10-14