AI chips are becoming a cooling problem as much as a computing breakthrough. Discovered Materials, a startup founded by Advaith Sridhar and Akash Ramdas, is betting that artificial intelligence can help solve it by finding new materials for more efficient semiconductors.
The company has raised $9 million in seed funding, led by Lightspeed India Partners after graduating from Y Combinator. Peak XV Partners and angel investors including Paul Graham, Gokul Rajaram, and Thariq Shihipar also participated.
Discovered Materials is targeting chip heat
The startup's core idea is straightforward but difficult to execute: use AI to search through enormous numbers of possible materials, then determine which ones might work in real semiconductor hardware.
Sridhar brings experience building AI agents from his work at Persona AI and Luma Labs. Ramdas contributes a doctorate in materials science from Stanford. Together, they have built a software pipeline that combines Anthropic models with physics models developed by the company.
The AI system generates candidate materials and explores possible research directions. Physics based simulations then help test whether those candidates have useful properties. According to the founders, the system can explore thousands of possibilities each day instead of relying on the much slower pace of manual research.
That increase in search capacity matters because the space of possible materials is enormous. But finding something interesting on a computer is only the beginning.
Finding a material is easier than making it useful
A semiconductor material cannot succeed because of one attractive property. A candidate might dissipate heat efficiently but be difficult to manufacture. Another might be easy to produce but have electrical characteristics that make it unsuitable for a chip.
This creates a difficult engineering tradeoff. Thermal performance, electrical behavior, durability, and manufacturability all have to line up before a new material becomes commercially useful.
Discovered Materials says it has already identified several candidates whose properties compare favorably with materials currently used by major chipmakers. The company has not disclosed enough technical detail to independently assess those claims, however.
The startup has also released examples of hundreds of generated materials and introduced a Material Discovery Bench intended to evaluate how advanced AI models perform on materials research problems.
The real bottleneck may be the laboratory
Discovered Materials is entering a field that already includes companies such as MatNex, SandboxAQ, and CuspAI. Its differentiation is a narrower focus on materials that could address the thermal demands of modern semiconductor hardware.
That focus could become increasingly valuable as AI data centers consume more electricity. Higher computing density means more heat has to be removed from increasingly compact systems, putting pressure on cooling infrastructure and the overall economics of AI deployment.
But AI does not eliminate the physical world. A model can propose a promising atomic structure in seconds. Researchers still have to synthesize it, measure it, refine the formulation, and determine whether it can survive the constraints of industrial chip manufacturing.
This is where the company's founders believe their combination of software and materials expertise could matter. They are not positioning AI as a replacement for laboratory work. Instead, the goal is to use computation to narrow the search before researchers spend time and money making physical samples.
From AI prediction to chip patents
If the company finds materials that prove valuable, Sridhar says it plans to pursue patents covering their use in GPUs or the processes required to manufacture chips with them. The business could then license those technologies to semiconductor companies.
That model gives the startup a potentially attractive position in the chip supply chain. It does not need to manufacture GPUs itself if a newly discovered material becomes important to companies that do.
Still, the commercial track record of AI driven materials discovery remains limited. Promising candidates have emerged across areas such as permanent magnets and semiconductor materials, but widespread industrial deployment has been much harder to achieve.
For Discovered Materials, that gap is the real test. Generating more candidates is becoming cheaper and faster. Proving that one of those candidates can survive the journey from simulation to factory floor is much harder.
If the startup can close that gap, AI materials discovery could become more than a research tool. It could become part of the infrastructure needed to keep the next generation of AI hardware running cool.