320 H100s for three months: a checkpoint that serves as a compute warning
teortaxesTex · x · 2026-10-06
teortaxesTex highlights a research post stating all its results come from a checkpoint trained from scratch on just 320 H100 GPUs over three months—a tiny fraction of frontier language and video model compute.
He frames this as a warning/research note, estimating 691K H100-hours at roughly 2e24 fp8 FLOPs.
Related event: Checkpoint Trained on Just 320 H100s Sparks Compute Threshold Debate(2 posts)→
More from Infra
- Strata engine boosts RTX 3090 prefill 17x, reigniting the local LLM debate — Iory1998 · 2026-10-06
- Hyperscaler AI capex to jump 92% to $789B in 2026, near $1.2T/year by 2029 — luisdans · 2026-10-06
- PyTorch Conference: IBM to keynote Spyre Accelerator and distributed inference work — PyTorch · 2026-10-06
- llama.cpp v0.6.0 ships MTP speculative decoding for Qwen4Exp and more — vexatious-big · 2026-10-06
- Deep interview: why NVIDIA engineered Nemotron 3 Ultra around speed and long context — yacinelearning · 2026-10-06
- After 2 years, a dev unveils Aether: a local AI 'OS' with structured memory and FailureMesh recovery — Budget_One_8784 · 2026-10-06