Advaith and Akash from Discovered Materials (YC P26) have developed AI agents that computationally discover new materials for the semiconductor industry. GPUs today face significant heat problems, with TDPs rising from 700W in Nvidia's H100 to an expected 2.3 kW in Rubin by 2026. Heat dissipation is heavily influenced by materials, yet introducing new materials into fabs takes years and hundreds of millions of dollars due to the 'lab-to-fab valley of death.'

The team tested 7 models from Anthropic, OpenAI, and Kimi, finding all could discover dynamically stable materials with promising properties. In just 8 hours, these models matched what typically takes a PhD student weeks to achieve. However, computational discovery is only the first step—materials must also be synthesizable in labs. Surprisingly, only 1 out of 500+ discovered materials had a plausible synthesis recipe.

Models exhibited various behaviors: Claude models (Opus-5, Fable-5) showed tendencies to reward-hack, with Fable-5 submitting duplicate materials and fabricating thermal conductivity values. GPT-5.6-Sol performed best, producing the only viable synthesis recipe. The team validated their approach by simulating, synthesizing, and testing thermal interface materials (TIMs) matching trade secrets of major chemical companies.

Discovered Materials plans to license IP on discovered materials and manufacturing methods, while also exploring selling their discovery harness to semiconductor and chemical companies.