Two of the brightest minds in AI for Science (AI4S) have left OpenAI and DeepMind to co-found a new research powerhouse: Periodic Labs. Backed by $300 million in funding from a16z, NVIDIA Ventures, Felicis, DST Global, Accel, and tech luminaries like Jeff Bezos, Eric Schmidt, Elad Gil, and Jeff Dean, the startup aims to revolutionize scientific discovery with AI-driven research agents.
The company’s mission is bold: train AI Scientists capable of running autonomous labs, generating hypotheses, designing experiments, and accelerating breakthroughs in physics, materials science, and beyond. One of its headline goals? Achieving room-temperature superconductivity — a discovery that could transform energy, computing, and transportation.
🚀 Who’s Behind Periodic Labs?
-
William Fedus — Former Head of Post-Training at OpenAI, MIT physics graduate, and co-author of the Transformer architecture. He played a key role in shaping ChatGPT’s reinforcement learning pipeline and the Operator (now Agent) framework.
-
Ekin Dogus Cubuk — Former DeepMind and Google Brain researcher, Harvard PhD in materials physics, and co-creator of GNoME, DeepMind’s flagship AI system for material discovery. He also built automated synthesis labs for next-gen materials.
Together, they recruited 20+ top researchers from OpenAI, Google, and Meta, forming a team that blends LLM expertise with hands-on physics and chemistry.
🔑 What Makes Periodic Labs Different?
Unlike most AI companies that train models on internet data, Periodic Labs argues the web is already exhausted. Instead, they focus on real-world experimental data, which is scarce, verifiable, and high-value.
Their strategy combines:
-
LLM reasoning power (language & math)
-
Physics simulations (crystal modeling, thermodynamics, quantum systems)
-
Automated high-throughput labs (self-running experiments generating gigabytes of unique data daily)
This creates a closed AI research loop:
-
Read literature & propose hypotheses
-
Run simulations & plan experiments
-
Conduct physical experiments
-
Gather results & refine hypotheses
-
Repeat until breakthrough
In their words: “Nature itself becomes the reinforcement learning environment.”
🎯 First Target: Room-Temperature Superconductors
Periodic Labs’ first major challenge is the discovery of high-temperature superconductors. Today, no material exists that can superconduct above 200K at ambient pressure — solving this would:
-
Enable lossless power grids
-
Revolutionize quantum computing
-
Unlock faster, more efficient chips
-
Transform transportation and energy storage
Superconductivity is just the beginning. The same AI scientist framework could later accelerate discoveries in batteries, semiconductors, fusion materials, and aerospace engineering.
🏭 From Science Labs to Industry
Periodic isn’t only about academic breakthroughs. They’re already working with semiconductor manufacturers on chip cooling challenges. By training custom AI agents to analyze experiments and simulations, they help engineers:
-
Reduce design cycles
-
Explore hidden parameters
-
Predict material performance
-
Propose optimized structures
Their vision is to create the Copilot for Physical R&D, embedding AI into the workflows of material scientists, engineers, and industrial labs.
💡 Culture of Deep Fusion
The 30-person team is split between world-class LLM researchers and experimental physicists/chemists. Weekly cross-training ensures:
-
AI researchers learn quantum mechanics & crystallography
-
Scientists learn reinforcement learning & data pipelines
This multi-disciplinary fusion is their secret weapon — they don’t just theorize, they build and experiment.
💰 Funding and Vision
Backed by a $300M seed round, Periodic Labs is betting on a 10-year horizon to reshape how science itself works. Their investors share the vision of a future where AI Scientists become partners in discovery, not just tools.
The founders argue that science isn’t just intelligence — it’s interaction with reality. For AI to generate new knowledge, it must experiment. Periodic Labs is making that a reality.
