Discovered Materials

AI chips face a severe heat problem, with GPUs handling heat fluxes higher than a space shuttle nose cone re-entering the atmosphere, and each new generation producing more heat than the last. This heat drives massive datacenter power and water consumption for cooling. The materials used in chips today are at their performance limit, yet discovering and qualifying a new material for use in a fab currently takes years and hundreds of millions of dollars. As a result, most promising material ideas die in the gap between a science experiment and real fab deployment.

Discovered Materials builds AI agents designed to compress months of interdisciplinary materials science work into days, aiming to accelerate the lab-to-fab timeline for new semiconductor materials. During their Y Combinator batch, the team used these agents to simulate, synthesize and test thermal interface materials matching the performance of products that major chemical companies have sold and guarded as trade secrets for over 20 years. Alongside this work, they are open-sourcing Material Discovery Bench, a benchmark built with experts from IBM, IMEC, Stanford and Cambridge to measure frontier AI model ability on real-world materials discovery problems. The benchmark includes multiple verifiers for grading model performance in both simulation and experiment, with more planned.

Advances in compute drove most technological progress for 50 years but stalled in the mid-2010s. Chips today are at least 10,000x less power efficient than the human brain, leaving enormous headroom that faster material discovery could help close.