Discovered Materials has raised $9 million in seed funding to point swarms of AI agents at a narrow problem: finding semiconductor materials that shed heat faster than the ones chipmakers use today. Lightspeed India Partners led the round after the company came out of Y Combinator’s latest batch, with Peak XV Partners also participating. Individual backers include Paul Graham and Gokul Rajaram, plus angel investor Thariq Shihipar.
The bet is as much about physics as it is about AI. Every watt a chip sheds as heat is a watt that cannot go toward computation, and it is a watt a data center operator has to pay for again in cooling. Rack density, how many accelerators an operator can pack into a given footprint, is now bounded less by how small a transistor can shrink and more by how much heat a rack can dump before components throttle. That puts a materials breakthrough upstream of nearly every scaling claim chipmakers and cloud providers currently make.
The company’s founders are Advaith Sridhar and Akash Ramdas. Ramdas holds a Stanford doctorate in materials science. Sridhar previously worked on agent systems at Luma Labs and Persona AI. Their pipeline runs in two stages: Anthropic models inside a custom harness generate candidate material leads, then a separate set of physics models the founders trained themselves run simulations to check whether those candidates actually hold up.
“[Ramdas] was doing maybe 20 guesses a day during his PhD,” Sridhar told TechCrunch. “We’re able to do thousands of guesses a day now by having these agents run 24/7 on the cloud, exploring research directions that he gives them.” The company published examples of hundreds of candidate materials alongside a new Material Discovery Bench, a tool meant to score how frontier models perform at this specific search task.
Discovered Materials joins a crowded field. MatNex, SandboxAQ, and CuspAI are all pursuing comparable AI-driven materials search. Discovered Materials is wagering that narrowing its focus to chip thermal problems, rather than materials science broadly, will let it compete with better-funded rivals. The company claims it has already identified several materials whose properties match ones major chipmakers currently use, though it declined to share specifics.
Hemant Mohapatra led the round for Lightspeed India Partners. He described the trade-offs bluntly: “It’s a bit of playing whack-a-mole with atomic structures. A material is only useful in the real world if all of them converge at once.” A compound that dissipates heat well may be nearly impossible to manufacture at scale, or its electrical properties may fail once built into an actual chip. Mohapatra does not think generating candidates is the hard part anymore; he expects that function to get commoditized as models improve, leaving filtering and synthesis as the real bottleneck.
That caution is earned. AI-driven materials and drug discovery has a long history of promising leads that took years, or never arrived, at commercial deployment. MatNex’s magnets that skip rare-earth elements, and semiconductor materials that Panasonic and Citrine Informatics developed, remain unshipped at scale. Insilico Medicine’s Renterosib is still only the first AI-discovered drug to reach a Phase II trial. Sridhar concedes the same limit applies here: turning a candidate into a real chip material still requires wet lab validation, “a process that cannot be sped up.”
Sridhar says the company plans to patent the materials it finds and license them to chipmakers, with a goal of having patentable candidates within a year. For any team modeling rack density into a 2027 capacity plan, the binding constraint may arrive from a materials lab well before it arrives from a fab.
TechCrunch’s Tim Fernholz first reported this story on August 10, 2026.
