Periodic Labs Reveals AI Infrastructure for Scientific RL

On September 15, 2026, Periodic Labs published a breakdown of its internal AI infrastructure for scientific reinforcement learning: peak 1,300 H200 GPUs during midtraining and RL stages, a stack built on Megatron, SGLang, Miles, and Ray, and claimed speedups of 4.1x for training and 2.5x for decoding compared to the baseline.

Compiled by Sergey KostenchukPublished 2026-09-15Updated 2026-09-18

2026-09-15 Coding Anthropic

🔬 Periodic Labs Reveals Infrastructure for Scientific RL

The company published a breakdown of its reinforcement learning stack on September 15, where a single RL rollout takes over an hour: peak 1,300 H200 GPUs, a stack built on Megatron, SGLang, Miles, and Ray. According to the company, training runs 4.1x faster than the Megatron baseline, and decoding for a 1T model has been accelerated from 10 to 25 tok/s.

🌍 Scientific RL with multi-hour rollouts is a distinct class of infrastructure challenges: standard solutions hit a ceiling. Optimizations are already in open upstreams — SGLang PR #24851 and Miles PR #1371, which can be reproduced.

👤 For engineers — a reference: a checkpoint of a trillion-parameter model is rebuilt from Megatron to Hugging Face in 1 minute instead of 30, and the pbox sandbox provides 4.5x lower latency compared to a hosted provider. The figures are self-reported by the company.

Source 1: https://periodic.com/news/ai-infrastructure-at-periodic Source 2: https://news.ycombinator.com/item?id=49717176

Sources