691K H100-hours: community estimates compute behind from-scratch checkpoint trained on just 320 H100s
teortaxesTex · x · 2026-10-06
A checkpoint trained from scratch on just 320 H100 GPUs for three months — a tiny fraction of frontier language and video model compute — prompted teortaxesTex to estimate roughly 691K H100-hours, plausibly 2e24 fp8 FLOPs. He frames it as a warning/research signal: competitive results may be achievable at far lower compute than assumed for frontier training runs.
Related event: Checkpoint Trained on Just 320 H100s Sparks Compute Threshold Debate(2 posts)→
More from Infra
- Learning electronics with Opus: two weeks of experiments distilled into interactive ET-SoC-1 diagrams — yaroslavvb · 2026-10-06
- Claude Code arrives in AWS GovCloud, bringing AI coding to ITAR-regulated workloads — AWS ML Blog · 2026-10-06
- The AI Stack Now Extends to the Power Plant as Google, Amazon, Meta Chase Nuclear — ingliguori · 2026-10-06
- LithosAI Launches LithosBox Millisecond Agent Sandboxes; Hits 727 TPS on GLM 5.3 Flash — JiaZhihao · 2026-10-06
- LithosAI Claims Third #1 Speed Spot: Fastest Inference for GLM 5.3 Flash on Artificial Analysis — JiaZhihao · 2026-10-06
- AWS ships aws-ai-ml skill that turns coding agents into SageMaker inference optimization experts — AWS ML Blog · 2026-10-06