Discovered Materials, a Silicon Valley startup, has raised $9 million in a seed round to build an AI-agent system that runs around the clock to find new materials for cooler semiconductor chips. The company is targeting a growing pain point in AI data centers: the rising power draw and heat output of accelerators. The round was led by Lightspeed India Partners.
Key takeaways
- Seed round: $9 million, closed after the Y Combinator Spring 2026 batch
- Investors: Lightspeed India Partners (Hemant Mohapatra) and Peak XV Partners, with angels including Paul Graham and Gokul Rajaram
- Founders: Advaith Sridhar (AI agents, previously Luma Labs) and Akash Ramdas (materials science PhD from Stanford)
- AI agents run thousands of guesses a day, operating 24 hours a day
- The company released hundreds of new materials and a Material Discovery Bench tracking tool
How the discovery pipeline works
Discovered Materials pairs Anthropic's AI models with custom software that generates candidate compounds, then checks their properties against physics models trained in-house. Instead of running single experiments by hand, a fleet of agents searches the space of possible materials without pause. It is a brute-force?brute-force: A "brute-force" method — systematically checking a huge number of possibilities instead of finding a clever shortcut. approach, but one steered by models that discard dead ends before they reach costly validation.
The discovery pipeline runs in four steps:
We're able to do thousands of guesses a day now by having these agents run 24/7.
That is how Advaith Sridhar, co-founder of Discovered Materials, describes the edge over traditional lab work. The company says its agents have reproduced materials with properties close to those already used by chipmakers, though it has not disclosed specifics.
The race for chip materials
Discovered Materials is not the first to point AI at materials chemistry. Google DeepMind's GNoME model previously flagged hundreds of thousands of hypothetical stable crystals, and Microsoft is developing the generative MatterGen. The difference is focus: the startup narrows the search to one well-defined problem — pulling heat out of chips in AI data centers, where power consumption and heat output keep climbing year over year.
| Player | Tool | Scope and goal |
|---|---|---|
| Discovered Materials | AI agents + physics models | narrow: cooling chips in data centers |
| Google DeepMind | GNoME | broad: hundreds of thousands of hypothetical crystals |
| Microsoft | MatterGen | broad: generative materials design |
The business model rests on patents. The company plans to patent the use of selected materials in GPUs or chip manufacturing processes, then license them to chipmakers — within a year, by its own account. That sets it apart from purely academic efforts that publish results with no path to commercialization.
Why it matters
Heat is now one of the hard limits on AI's growth — the denser the accelerators, the harder they are to cool, which drives up energy costs in data centers. If AI agents really do shrink materials discovery from years to months, the economics of the whole chain shift: a cooler chip means a lower power bill and higher compute density. The key question is whether computationally selected materials survive real validation in a fab. Until independent data exists, the promises remain promises.
What's next?
- The company says it will patent material uses in GPUs and license them to chipmakers within a year of the round
- The next test is independent lab validation of the materials logged in the Material Discovery Bench
Sources
- TechCrunch — Discovered Materials is playing AI whack-a-mole to hunt cooler chips
- Y Combinator — Discovered Materials
