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.
More from Infra
- Big Tech's Future Data Center Leases Top $1.3T, 70% Added in Last Four Quarters — rohanpaul_ai · 2026-11-02
- Daniel Lemire: AI is exploding demand for boring compute, server CPUs to double — lemire · 2026-11-02
- Physical AI's real story: a 4-layer map of 15 companies behind humanoid robots — pdamodaran · 2026-11-02
- Under 5% of US consumers pay for AI, yet compute demand is already forcing usage cuts — NinaDSchick · 2026-11-02
- Sam Altman: AI query energy cost fell 1,000x-100,000x in 3-4 years — rohanpaul_ai · 2026-11-02
- Marvell and Lumentum to merge, targeting $550B AI connectivity market — BenBajarin · 2026-11-02