NVIDIA Is Hitting a Heat Wall. Here Comes the Fix. | Advaith Sridhar, Discovered Materials •DD15

Two best friends from college. One watched Claude go from helping with his code to writing nearly all of it, almost overnight. The other, Akash, was finishing a Stanford PhD in materials science. When they asked the model to work on some unpublished research Akash had spent a year on, it reached similar conclusions in weeks. That was the moment they decided to start a company. In this episode of Deeptech Decoded, Nihal Kurth   / nihalkurth   talks with Advaith Sridhar   / advaith-sridhar  , co-founder of Discovered Materials (YC), about how two people in a garage lab in Oakland made thermal interface materials that, by their own tests (in their lab and at a third-party facility in Livermore), match products Dow, Honeywell and Henkel have sold for over 20 years. And why Advaith believes the bottleneck in semiconductors is no longer just computation, but the materials that have to survive exploding power levels. Advaith says NVIDIA’s next-generation chips are heading toward 4,000 watts, a single chip putting out the heat of roughly four US houses. The materials in chips today were never designed for that. He and Akash are betting that AI agents, combined with real lab iteration, can compress the time it takes to find and qualify new thermal materials. He also explains why he thinks the “self-driving lab” consensus is wrong. Chapters: 00:00 – Intro 03:42 – Origin: Akash and the PhD defense 04:27 – Claude takes over his codebase 05:46 – Claude vs. a year of Stanford research 07:40 – GNoME: +30% materials overnight 10:16 – 5.5 million possible materials 12:05 – Why structure beats composition 14:56 – Materials vs. drug discovery 17:50 – An R&D lab, not a factory 20:25 – Inverse design and AI benchmarks 23:03 – Why lab results aren't repeatable 25:11 – The chip heat wall: Blackwell to Rubin Ultra 29:24 – Thermal paste: 40% of a chip's heat resistance 30:00 – Garage lab: on par with Dow & Honeywell 30:53 – Why the self-driving lab wasn't the win 34:17 – Computation vs. experiment 37:31 – Why they paused thermal paste 41:30 – What they're building next 44:15 – Feedback from chip makers and YC 46:33 – The VC who wanted one slide 52:32 – Staying sane: co-founder, Bay Area, partner 54:26 – Tech stack: Claude Code, Supabase 57:32 – Distribution may beat product 1:01:10 – The 10-year view: chips and AI 1:02:43 – World models and digital twins 1:06:10 – From PM back to engineering Where to find Nihal: LinkedIn →   / nihalkurth   Where to find Advaith: Deeptech Decoded → https://deeptech-decoded.com/guests/a... LinkedIn →   / advaith-sridhar   Discovered Materials → https://discoveredmaterials.com Where to find us: Substack → https://deeptechdecoded.substack.com/ LinkedIn →   / deeptechdecoded   Spotify → https://open.spotify.com/show/5c2T63G... Apple Podcasts → https://podcasts.apple.com/gb/podcast...