Discovered Materials, a $9 Million-Seed YC Startup, Puts $1 Million on Five Problems Its Own AI Agents Could Not Solve

Discovered Materials, a San Francisco startup from Y Combinator's spring 2026 batch, is offering a $1 million prize pool to anyone who solves one of five open problems in materials science, TechCrunch reported on Nov. 5. The company says it first gave frontier AI models internet access and other resources to work autonomously for days on each problem, and the models made little progress.

The prize is a marketing device for a company that sells the opposite claim: that AI agents can find new chip materials. It is also a disclosure that the same agents failed on the hardest cases.

What the five problems are

An advisory group of academics at leading universities helped define them, and Felipe Jornada, a Stanford professor, helped develop them, according to TechCrunch. Advaith Sridhar, a cofounder, said the challenges are meant to act as benchmarks for measuring AI progress in the field.

#	Problem
1	A theory explaining cuprate superconductivity
2	A numerical method for predicting silicon and silicon dioxide interface structures
3	A numerical method mapping grain structure to thermal properties
4	A theory predicting why light changes material behaviour, particularly around superconductivity
5	A numerical method for improving density functional theory

Sridhar said: "It's more difficult to specify problems clearly in materials science because it's not like math, where you have a very specific definition and can get to a yes or no." That is also the hazard for a prize: a judging rule for "a theory" is harder to write than one for a proof.

The $9 million seed behind it

The company announced a $9 million seed round on Aug. 10, led by Lightspeed India Partners, per Dealroom, which names partner Hemant Mohapatra as leading it. Peak XV Partners and the angels Paul Graham, Gokul Rajaram and Thariq Shihipar also invested. TechCrunch's piece confirms the seed but does not give the amount.

RuntimeWire, which covered the launch on Aug. 12, labels the funding claim as company-supplied and names no other investors in the body. The founders are Advaith Sridhar and Akash Ramdas.

How the product works, and what is unproven

Dealroom says a software pipeline uses Anthropic models to generate material leads, then runs simulations on physics models the founders trained to check them. The plan is to patent promising materials, or the process for making chips from them, and license them to chipmakers.

Sridhar told Dealroom that Ramdas made about 20 guesses a day during his PhD and that the company now runs thousands a day, with agents working around the clock in the cloud. The company says it has found several materials that match properties used by major chipmakers, but has not shared details. Both claims are the company's own.

Dealroom also records the limits. Materials that dissipate heat may be too hard to manufacture or have poor electrical properties, and no AI-discovered material has been deployed commercially at scale. Mohapatra's own phrasing: "filtering them correctly and synthesizing them is the bottleneck." RuntimeWire's reading of the company's results agrees that synthesis remains the main bottleneck.

Competitors and the next signal

Dealroom names MatNex, SandboxAQ and CuspAI as pursuing similar goals. TechCrunch says Sridhar now expects an industrially useful material in under a year, a shorter horizon than his August estimate.

The prize adds a public test. If any of the five problems is solved, the company's benchmark gains a data point, and if none is, the $1 million stays unspent while the claim about agents rests on results the company has not published. For what independent measurement says about AI speedups in research more broadly, see Anthropic's study putting the gain at 1.35x to 2.02x and the METR review of the same threshold.