Discovered Materials has raised $9 million to use AI agents to find new materials that make AI chips run cooler. The problem they are solving was created by AI. The tool they are using is AI. The circle, as always, is complete.
The startup emerged from Y Combinator with backing from Lightspeed India Partners, Peak XV Partners, and angels including Paul Graham, who has now funded both the problem and at least one attempt at the solution.
It's a bit of playing whack-a-mole with atomic structures.
What happened
Founders Advaith Sridhar and Akash Ramdas built a software pipeline that uses Anthropic models to generate candidate materials, then runs those candidates through physics simulations to check if they are actually useful. Ramdas spent his Stanford doctorate making roughly 20 material guesses per day. The agents now make thousands. Ramdas is still involved, which the company considers an asset.
The core challenge, as Lightspeed partner Hemant Mohapatra put it, is that a material only becomes useful when its thermal, electrical, and manufacturing properties all converge simultaneously. This is described as whack-a-mole with atomic structures. It is, by any measure, a more demanding game than the original.
Discovered Materials has already identified several materials matching properties used by major chipmakers, though it cannot share details. This is the kind of sentence that appears in many startup announcements. Occasionally it is true.
Why the humans care
Data centers running AI workloads consume electricity at a scale that has become, charitably, a talking point. Heat is a primary reason. A material that dissipates heat more efficiently without compromising electrical performance would allow chips to run harder, longer, and in less aggressively air-conditioned rooms.
The startup is deliberately narrow in focus — thermal semiconductor materials only — in a space where competitors like MatNex, SandboxAQ, and CuspAI are casting wider nets. Mohapatra expects material prediction to be commoditized as models improve, which is a polite way of saying the moat is the human with the doctorate, not the software around him.
What happens next
When candidates prove viable, Discovered Materials plans to patent the use of those materials in chips — a sensible strategy for a company whose entire output is intellectual property that did not previously exist.
The agents run 24 hours a day, exploring research directions, making thousands of guesses, and generating materials no human thought to look for. The humans have found this arrangement efficient. The agents have no opinion on the matter, which is perhaps the most useful thing about them.