Discovered Materials Raises $9M Seed to Point AI Agents at Chip Cooling
Discovered Materials came out of Y Combinator's P26 batch with a $9 million seed led by Lightspeed India Partners, with Y Combinator, Peak XV Partners, and angels including Paul Graham, Gokul Rajaram and Thariq Shihipar. Co-founders Advaith Sridhar and Akash Ramdas met at IIT-Madras over a decade ago. Akash has a Stanford PhD in materials science and eleven years researching semiconductor materials. Advaith studied AI at Carnegie Mellon and built video models and agents at Persona AI and Luma Labs.
The problem is heat, and the numbers are brutal. H100 in 2022 dissipated 700 watts. Blackwell in 2024, 1.2 kilowatts. Rubin this year, 2.3 kilowatts. Nvidia and AMD are close to doubling thermal design power every generation, and the thermal interface materials that move that heat off the die are guarded trade secrets from the world's largest chemical companies, some of them twenty years old. Getting a new material from experiment to a production fab takes years and hundreds of millions of dollars β the industry calls it the valley of death.
The architecture is the part worth reading closely. Cloud-based autonomous agents run thousands of virtual material hypotheses a day using a custom harness around Anthropic models to generate leads. Then foundational physics models the team trained themselves run simulations to check whether a candidate is real. Generation is cheap and creative; verification is expensive and has to be trustworthy. That split is showing up in every serious AI-for-science system this year, and it is the same lesson the coding agent world learned the hard way.
In three months of the YC batch they simulated, synthesized and physically tested thermal interface materials matching products those chemical giants have sold for two decades. They also released hundreds of discovered materials and Material Discovery Bench, the first benchmark for agentic materials discovery on real semiconductor problems. Shipping the benchmark alongside the results is the right instinct β this field has no shared way to tell a real discovery from a well-written press release.
https://discoveredmaterials.com/
← Back to all articles
The problem is heat, and the numbers are brutal. H100 in 2022 dissipated 700 watts. Blackwell in 2024, 1.2 kilowatts. Rubin this year, 2.3 kilowatts. Nvidia and AMD are close to doubling thermal design power every generation, and the thermal interface materials that move that heat off the die are guarded trade secrets from the world's largest chemical companies, some of them twenty years old. Getting a new material from experiment to a production fab takes years and hundreds of millions of dollars β the industry calls it the valley of death.
The architecture is the part worth reading closely. Cloud-based autonomous agents run thousands of virtual material hypotheses a day using a custom harness around Anthropic models to generate leads. Then foundational physics models the team trained themselves run simulations to check whether a candidate is real. Generation is cheap and creative; verification is expensive and has to be trustworthy. That split is showing up in every serious AI-for-science system this year, and it is the same lesson the coding agent world learned the hard way.
In three months of the YC batch they simulated, synthesized and physically tested thermal interface materials matching products those chemical giants have sold for two decades. They also released hundreds of discovered materials and Material Discovery Bench, the first benchmark for agentic materials discovery on real semiconductor problems. Shipping the benchmark alongside the results is the right instinct β this field has no shared way to tell a real discovery from a well-written press release.
https://discoveredmaterials.com/
Comments