Discovered Materials Screens 4,300 Materials With AI

As AI chips become more powerful, the semiconductor industry is running into a problem that cannot be solved with more computing alone: heat. A young Y Combinator startup called Discovered Materials is betting that artificial intelligence can help find entirely new materials capable of moving that heat away from chips faster.

The San Francisco startup, previously known as Matforge, is building what it calls “AI scientists” for semiconductor materials research. Instead of asking researchers to manually evaluate one potential compound after another, its system uses AI agents to propose materials, run computational tests, reject weak candidates and continue searching. The company says traditional discovery of commercially useful semiconductor materials can take more than 10 years; its ambition is to identify substantially better alternatives within months.

4,300 materials in three days

One of the startup’s early experiments offers a glimpse of how that approach works.

Co-founder Advaith Sridhar said the team adapted an agentic research loop to search for materials with high thermal conductivity. The system repeatedly proposed candidates, ran phonon calculations, checked whether the structures were stable and evaluated their potential as thermal conductors.

Over three days, the AI worked through about 4,300 different materials. According to Sridhar, more than 50% of the verified materials had never previously been made, leaving the team with a collection of relatively unexplored candidates that could potentially warrant laboratory testing.

That does not mean the AI has already discovered a commercially viable cooling material. Computationally promising compounds still need to be synthesized, characterized and validated in the physical world. Discovered Materials says laboratory experiments are part of its intended discovery loop, rather than treating an AI prediction itself as the finished discovery.

The distinction matters. AI can dramatically expand the number of candidates scientists examine, but the difficult jump from a calculated structure to a material that can actually be manufactured economically and integrated into semiconductor packaging remains.

AI racks are getting dramatically hotter

The timing is important because the thermal demands of AI infrastructure are rising at extraordinary speed.

An Nvidia data-center presentation described its GB200 NVL72 systems at roughly 120 kW per rack and said Rubin Ultra systems planned for 2027 could reach around 600 kW per rack. The presentation even discussed the possibility of megawatt-scale racks appearing before the end of the decade. At those densities, conventional air cooling becomes increasingly impractical and thermal engineering becomes a central constraint on computing performance.

The problem extends beyond individual racks. The International Energy Agency reported that global data-center electricity consumption reached roughly 485 TWh in 2025 and projects it to approach 950 TWh by 2030. Electricity consumption from AI-focused data centers alone is projected to triple over that period.

That makes materials capable of moving heat efficiently increasingly valuable. Better thermal conductors could potentially improve heat spreading inside packages, cooling interfaces and other parts of increasingly dense computing systems, although any candidate identified by Discovered Materials would still have to prove it can outperform existing materials under real manufacturing conditions.

A materials scientist and an AI researcher

Discovered Materials is currently a remarkably small operation. Y Combinator lists the company as a two-person startup founded in 2026 by Akash Ramdas and Advaith Sridhar. Ramdas completed his PhD and postdoctoral work at Stanford in semiconductor materials research, while Sridhar studied AI at Carnegie Mellon and previously worked on foundation models and AI agents at Luma Labs and Persona AI.

YC says materials Ramdas previously worked on for nanoscale interconnects have made their way into technology roadmaps at Intel and TSMC, giving the startup unusually direct semiconductor expertise for such a young company.

The company now says it wants to find materials that are as much as 10 times better than existing alternatives and reduce a discovery process measured in decades to one measured in months. That is an ambition rather than a demonstrated commercial result, but the 4,300-candidate experiment shows how radically AI could change the search stage of materials science.

For the AI industry, that search may become increasingly important. The next leap in compute may depend not only on better GPUs or smaller transistors, but also on finding physical materials capable of surviving, and removing, the enormous amounts of heat those chips create.