Former OpenAI VP and DeepMind Scientist Launch AI‑Powered Science Startup with $300M Funding

In spring 2025, former OpenAI research vice‑president Liam Fedus announced his departure, emphasizing a strategic interest in applying AI to scientific research. Around the same time, DeepMind scientist Ekin Dogus Cubuk left DeepMind after leading a large‑scale crystal‑structure generation project.

During a conversation in San Francisco, the two recognized that generative AI could reshape how experiments are designed and executed, but also noted that existing internet‑scale data pools were reaching their limits and that LLMs alone could not discover breakthroughs such as room‑temperature superconductors.

Motivated by this insight, they founded Periodic Labs, a research company whose mission is to create an AI‑driven scientific platform that not only analyzes data but also designs experiments, operates laboratory instruments, and discovers new materials.

The platform is built on a three‑layer “science stack.” First, an autonomous robotic lab performs precise powder synthesis, material mixing, and high‑temperature furnace operations. Second, high‑fidelity physics simulations powered by AI rapidly evaluate reactions in a virtual environment. Third, a large‑language‑model research assistant parses literature, generates hypotheses, and interprets experimental results. The loop follows virtual inference → physical execution → data feedback, turning every experiment—successful or failed—into training data for the next iteration.

This approach builds on prior AI breakthroughs such as AlphaFold, molecular generative models, and the A‑Lab platform described in Cubuk’s 2023 Nature paper, which synthesized 41 new compounds in 17 days, demonstrating the feasibility of AI‑driven experimentation.

In September 2025, Periodic Labs announced a $300 million seed round, the largest ever for an AI startup at that stage. The round was led by Andreessen Horowitz with participation from a16z, DST, Nvidia Ventures, and angel investors including Jeff Bezos, former Google CEO Eric Schmidt, and DeepMind figure Jeffrey Adgate. Investors framed the investment as a chance to compress decades of research into a few years.

Following the funding, the company recruited more than 20 top researchers from Meta, OpenAI, and DeepMind—among them the inventor of transformer attention, an OpenAI operator developer, and a Microsoft MatterGen creator. An advisory board led by Nobel laureate Carolyn Bertozzi adds expertise from superconductivity and materials science.

Periodic Labs’ initial scientific target is the discovery of high‑temperature superconductors, aiming to replace the extreme cooling and pressure requirements of existing materials. The team is also collaborating with a chip manufacturer to use AI agents for optimizing thermal‑management materials in semiconductor devices.

The founders argue that traditional research overlooks “negative results,” yet these failures are crucial for training AI scientists. By continuously generating a “science experience database” from both successful and failed experiments, Periodic Labs seeks to give AI models a richer understanding of physical reality than text‑only datasets can provide.

As the autonomous lab repeats thousands of experiments and the AI model processes hundreds of thousands of data points, the authors anticipate a scientific revolution driven by silicon‑based intelligence.