Periodic Labs: specialized models beat GPT-6 on science evals with 1,300 H200s
vwxyzjn · x · 2026-09-16
Periodic Labs details its AI infrastructure: specialized models trained from open weights Pareto-dominate frontier models like GPT-6 Astra and Claude Fable 5.1 on X-ray diffraction evals, using only a peak of 1,300 H200s across midtraining and RL. Key points:
- Models now analyze experiments in high-throughput labs, searching for better superconductors and magnets.
- Optimizations span training throughput, inference speed, custom sandboxing, GPU memory efficiency, and scientific tool execution; the cluster sustains 95%+ utilization by sharing capacity with scientific simulation workloads.
- Built on Megatron, SGLang, Miles, and Ray, heavily modified for scientific RL where rollouts run for hours while training steps take minutes. The stack delivers 4.1x Megatron training throughput on the same GPUs, and their SGLang contributions made inference 2.5x faster.
The goal: make every GPU-hour go further and shorten the idea-to-experiment loop.
Related event: Periodic Labs Trains Materials Model Beating GPT-6 with Just 1,300 H200s(28 posts)→
More from Infra
- Dev slams agent sandbox pricing as 20x+ the cost of a $6/month always-on VPS — Aryvyo · 2026-09-16
- Signal65 benchmarks NVIDIA Vera CPU at 1.64x per core vs x86 on agentic workloads — ryanshrout · 2026-09-16
- McKinsey: Sovereign AI TAM to hit $500-600B by 2030, 30-40% of AI demand — Beth_Kindig · 2026-09-16
- Intel CEO Lip Bu Tan tells AI Infra Summit: embrace AI, don't fear it — karlfreund · 2026-09-16
- Brad Gerstner: AI buildout hinges on revenue keeping its steep growth curve — markjeffrey · 2026-09-16
- Reported: NVIDIA hardware delivers 40% more throughput at the same power — karlfreund · 2026-09-16