AI chips run too hot, and that single thermal problem drives billions of dollars in electricity and cooling costs across every data center on the planet. Discovered Materials, a Y Combinator startup founded by Stanford materials scientist Akash Ramdas and AI engineer Advaith Sridhar, just raised $9 million in seed funding to use swarms of AI agents to find new materials that could make integrated circuits dramatically more efficient.
The irony is hard to miss. AI created the overheating problem by demanding denser, faster chips that push thermal limits further with every generation. Now this startup is using AI to search for the atomic structures that could solve it. Lightspeed India Partners led the round, with Peak XV Partners and angel investors Paul Graham, Gokul Rajaram, and Thariq Shihipar also participating.
On their website, the company writes in bold words:
Advances in compute have driven most of technological progress in the last 50 years. That engine stalled in the mid-2010s, and we intend to restart it.
The headroom is enormous. Chips today are at least 10,000x less power efficient than the human brain, and our goal is to close this gap by accelerating material discovery.
The speed advantage over traditional materials research is where the story gets interesting for anyone watching the semiconductor supply chain. During his PhD at Stanford, Ramdas tested roughly 20 material candidates per day through manual experimentation. Discovered Materials now runs AI agent swarms 24/7 on cloud infrastructure, generating thousands of guesses daily. The agents use Anthropic models in a custom harness to produce material leads. Proprietary physics models the team trained separately then simulate whether those candidates actually perform as predicted.
The company released examples of hundreds of new materials alongside its “Material Discovery Bench,” a benchmark designed to track how frontier AI models handle materials science challenges. Discovered Materials says it has already identified several materials matching the properties that major chipmakers currently use, though it cannot share specifics while patent processes are underway.
The engineering challenge is what Lightspeed partner Hemant Mohapatra described as “playing whack-a-mole with atomic structures.” A material that reduces heat generation might be impossible to manufacture at chip scale. One that dissipates heat beautifully might compromise electrical performance. Every property improvement in one dimension tends to break something else. The candidate is only useful if thermal conductivity, electrical behavior, manufacturability, and stability all converge simultaneously.
“[Ramdas] was doing maybe 20 guesses a day during his PhD,” Sridhar said. “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.”
Mohapatra expects the prediction side of materials discovery to become commoditized as frontier models improve. What differentiates Discovered Materials, he argues, is Ramdas’ deep domain expertise combined with the ability to rapidly synthesize and validate candidates in a physical laboratory. The founders have already validated several new materials in wet labs, moving them beyond computational predictions into real-world confirmation.
The business model targets semiconductor IP licensing rather than manufacturing. When they identify commercially valuable materials, Sridhar says the company will patent either the material’s use in GPUs or the fabrication process that makes the chip possible, then license those patents to chipmakers. He expects patentable discoveries within the next year.
The broader AI materials discovery space remains commercially unproven despite enormous scientific promise and growing investor enthusiasm. No AI-discovered material has reached commercial deployment at scale yet. Competitors including MatNex, SandboxAQ, and CuspAI have launched similar efforts targeting different material categories. The closest parallel in drug discovery is Insilico Medicine’s Rentosertib, the first generative-AI-discovered drug to reach Phase III clinical trials in July 2026.
Sridhar acknowledges the fundamental bottleneck that no amount of computational power can bypass entirely.
“A lot of this will involve actually going into wet labs and making things as well,” he told the media “And this is the process that cannot be sped up.”
The atoms still move at the speed of chemistry, not the speed of inference.
