A Startup Is Using AI to Find New Materials That Keep Computer Chips From Overheating

Discovered Materials, a startup using AI to discover new materials for computer chips, has raised a $9 million seed round led by Lightspeed India Partners, with participation from Peak XV Partners and angel investors Paul Graham, Gokul Rajaram, and Thariq Shihipar TechCrunch. The company, founded by Advaith Sridhar and Akash Ramdas, came out of Y Combinator and is also open-sourcing a tool called Material Discovery Bench, which tests how well AI models can discover new materials.

Lightspeed India Partners partner Hemant Mohapatra led the round. The founders met over a decade ago at IIT-Madras. Ramdas earned a doctorate in materials science from Stanford, where his work on tiny wires inside chips became Stanford Engineering's most popular story of 2025. Sridhar previously worked on AI agents at Persona AI and Luma Labs TechCrunch.

The idea is simple to describe, if hard to do: use AI agents to search through the enormous range of possible chemical compounds and find ones that can keep chips cooler and use less power. The stakes are real. Today's GPUs generate heat at about 140 watts per square centimeter, which is more heat than a space shuttle nose cone endures during re-entry into Earth's atmosphere. The company also notes that today's chips are at least 10,000 times less power efficient than the human brain.

During its three-month Y Combinator program, the company simulated, created, and tested heat-transfer materials that matched the performance of secret formulas held by the world's largest chemical companies for over two decades. These thermal interface materials sit between a chip and the metal part that draws heat away from it, filling tiny gaps so heat moves out efficiently. Going from computer simulation to making the actual material to testing it, all within three months, is what caught investors' attention.

Material Discovery Bench, built with experts from IBM, IMEC, Stanford, and Cambridge, tests how well today's most advanced AI models do on real materials problems. It includes multiple checks for grading model performance, giving researchers a shared standard for evaluation.

The broader context here matters. The chip industry's heat problem has gotten worse as AI systems require more and more computing power. As individual chips push past 1,000 watts and the biggest data centers now run hundreds of thousands of these chips, the thin layer of material that carries heat away from each chip is no longer a minor part. It has become a serious bottleneck. Finding new materials has traditionally been a slow, trial-and-error process controlled by large chemical companies with decades of experience. The claim that an AI-driven approach can compress that timeline into months and match closely guarded trade-secret formulas deserves scrutiny. The company has published its benchmark, which gives others a way to check the claims.

What separates credible claims from hype in AI for materials will be whether others can reproduce the results. Open-sourcing Material Discovery Bench is a meaningful step in that direction. If AI models can be tested on standardized, real-world materials problems with multiple independent checks, the field gets a shared yardstick. That matters because materials science does not have the fast feedback loops that software does. An AI can suggest a new compound, but proving it works requires actually making the material and testing it under real heat and electrical loads. The benchmark's use of multiple verifiers suggests the company understands this gap.

The investor lineup also signals something about how the market sees this. Lightspeed India and Peak XV bringing institutional money, plus Graham's angel investment, places Discovered Materials at the crossroads of deep materials science and the AI-agent investment trend. Ramdas's Stanford background in tiny chip wiring and Sridhar's experience building autonomous AI agents at Persona AI and Luma Labs create a founder pairing that fits the problem: one knows the materials science, the other knows how to build AI systems that work on their own.

The heat-management layer of a computer chip has not historically attracted this kind of founder or this kind of money. The fact that it is doing so now reflects the pressure that AI computing demand is putting on every part of the hardware stack, from silicon manufacturing to packaging to data center cooling. If AI can speed up materials discovery by even a fraction of what Discovered Materials claims, the effects go beyond one company. New heat-transfer materials that improve efficiency even slightly, used across hundreds of thousands of chips, could meaningfully cut data center power consumption. The open benchmark will determine how quickly the field can verify and build on these claims.