Periodic Labs, a newly-launched startup co-founded by former OpenAI researcher Liam Fedus and ex-Google DeepMind scientist Ekin Dogus Cubuk, has pulled in an unprecedented $300 million seed round to build what it calls the next frontier of scientific discovery: AI systems that don’t just analyse data, but generate and test hypotheses in physical labs.
A Bold Vision
Periodic Labs says the time is right to move beyond internet-trained models toward AI that intersects real-world experiments. According to Cubuk, the convergence of reliable robotics for powder synthesis, powerful simulation tools for physical systems, and advanced large-language-model (LLM) reasoning capabilities — helped along by Fedus’s past work at OpenAI — set the stage for this venture.
Seed Round Highlights
The funding round, led by Andreessen Horowitz (a16z), also saw involvement from Accel, DST Global, NVIDIA’s venture arm NVentures, and angels such as Jeff Bezos, Eric Schmidt, and Elad Gil. The round reportedly values the company at roughly $1 billion pre-money according to one source.
Why It Matters
Rather than relying solely on internet-derived text and code, Periodic Labs aims to generate new scientific data by giving AI models control of robotic labs and experimental workflows. As one commentary puts it: “Until now, scientific AI advances have come from models trained on the internet. But despite its vastness, it’s still finite.” Their first target domain is materials science — notably the search for next-generation superconductors and advanced compounds that could reshape power, computing, and manufacturing.
The Team & Strategy
Fedus, known for his role at OpenAI in developing ChatGPT and large-scale models, teamed up with Cubuk, who led materials/chemistry research at Google DeepMind and published work on robot-powered labs that synthesised novel compounds. Their hire list already includes elite AI, materials science and robotics talent, each week participating in cross-discipline lectures to maintain a tight integration between AI modelling and physical experimentation.
Challenges Ahead
While the vision is ambitious, the startup acknowledges that turning robotic labs and AI scientists into consistent breakthroughs is a high-risk, deep-tech play. Many experiments will “fail” — but in their view, those failures are valuable training signals. The next few years will test whether this model delivers tangible results and scales beyond proof-of-concept.