Periodic Labs: 1,300 H200s beat frontier models on X-ray evals, 4.1x Megatron throughput

zephyr_z9 · x · 2026-09-16

Periodic Labs details its AI infrastructure: starting from open weights, a peak of just 1,300 H200 GPUs across midtraining and RL produced specialized models that Pareto-dominate frontier models like GPT-6 Astra and Claude Fable 5.1 on X-ray diffraction evals. Their Megatron-derived stack delivers 4.1x training throughput, SGLang contributions made inference 2.5x faster, and cluster sharing with simulation workloads sustains 95%+ utilization — all aimed at discovering better superconductors and magnets.

Related event: Periodic Labs Unveils Neon: 1,300 H200s Train a Science Model Beating GPT-6(22 posts)→

Original post →

More from Infra

Infra channel →