Periodic Labs' Neon: 1,300 H200s and a robotic materials lab to beat GPT-6 Astra on its benchmark

LiamFedus · x · 2026-09-16

Liam Fedus unveils Periodic Labs' "frontier lab for atoms": high-throughput materials labs in Menlo Park create a loop where experiments generate fresh data, models learn from it, and then decide what to try next.

The lab's new model, Neon, was mid-trained and RL'd using only 1,300 H200s plus months of proprietary experimental data, and reportedly surpasses GPT-6 Astra on the team's analysis benchmark. Initial focus: hard materials science problems including superconductors, magnets, and semiconductor materials.

Related event: Periodic Labs trains materials-science model Neon to beat GPT-6 on just 1,300 H200s(50 posts)→

Original post →

More from Companies & People

Companies & People channel →