Discovered Materials, a San Francisco startup barely months old, is offering $1 million, paid out $200,000 at a time, to anyone who can solve five materials-science problems that today’s AI models tried and failed to crack.
The company announced the prize pool on Wednesday, November 5, 2026. Before going public with the challenge, Discovered Materials gave frontier AI lab models internet access and let them work autonomously for days at a stretch on each of the five problems. The models barely moved the needle. Now the company is opening the same five problems to the broader research community, with cash attached to each one.
Discovered Materials came out of Y Combinator’s spring 2026 batch. Its pitch is narrow and specific: use AI models to hunt for new substances that could go into advanced computer chips. That is a tightly scoped problem compared to what the company is now asking the research world to help with, which is explaining physics that nobody, human or machine, currently understands well enough to act on.
Five problems nobody has cracked
The company did not pull these questions out of thin air. It built the list with an advisory group of academics from leading universities, and each one maps to a real gap in materials physics rather than an abstract puzzle. The five are:
- A theory explaining cuprate superconductivity, meaning why certain materials carry electrical current with zero resistance at ultra-low temperatures.
- A numerical method to predict the atomic structure of silicon and silicon-dioxide interfaces, a pairing that shows up in nearly every semiconductor made today.
- A numerical method that maps grain structure to thermal properties, aimed at helping engineers design materials that dissipate heat more effectively.
- A theory predicting why shining light on certain materials changes their behavior, particularly around superconductivity.
- A numerical method to improve Density Functional Theory, the workhorse technique researchers use to model atomic and molecular structures across materials science.
Solve any one of them and the company says there is $200,000 waiting. Solve all five and the total hits $1 million. What Discovered Materials has not published is how a researcher would actually enter, what format a submission needs to take, or what deadline applies. The company has not disclosed an entry process for claiming the money, so for now the offer stands as a target rather than a formal competition with rules attached.
Why these problems and why now
Co-founder Advaith Sridhar framed the challenge as an attempt to give the field something it has never had. Math has the Millennium Prize Problems, a famous list of seven questions with a million dollars attached to each. Materials science has nothing comparable, partly because the problems are harder to pin down in the first place.
“It’s more difficult to specify problems clearly in materials science because it’s not like math,” said Advaith Sridhar, co-founder of Discovered Materials.
That difficulty is exactly what led the company to design the challenges with a verification step built in from the start, rather than leaving it as an afterthought. Sridhar described the sequence as deliberately backwards from how these things sometimes get framed.
“The reason we’re putting out these challenges is because we want them to serve as the goalposts,” Sridhar said, adding that the company’s method is to “come up with the verification and the challenges first, and then we let somebody come up with the ideas that explain the problem.” In other words, Discovered Materials decided how an answer would be checked before anyone, human or AI, was allowed to go looking for one.
That instinct fits a pattern showing up across AI research generally. Labs are increasingly building ways to check a model’s claims independently rather than taking its word for it. Anthropic’s recent move to open-source FlashProve is one example of that push toward verifiable results, built for a different domain but aimed at the same underlying worry: that AI systems are generating output faster than anyone can confirm it is actually right.
Akash Ramdas, Discovered Materials’ other co-founder, was named alongside Sridhar in the announcement but was not quoted directly.
Where AI has succeeded, and where it has not
The timing of this challenge is not incidental. AI models have had real wins this year on notable math problems, and they have made genuine progress in life sciences, where biological systems, for all their complexity, tend to yield to pattern recognition across large datasets. Materials physics has been a different story. The underlying equations governing how electrons and atoms behave in a solid are often too complex to solve directly, and the systems that current models have cracked tend to be the ones with cleaner, more tractable rules.
Stanford professor Felipe Jornada, who helped develop the five challenges, put the gap plainly in a statement accompanying the announcement.
“Across materials science, there are phenomena we observe but cannot explain,” said Felipe Jornada, professor at Stanford University.
Jornada added that explaining those phenomena would improve scientists’ understanding of the underlying physics, not just produce a single useful material. That distinction matters here. A numerical method that correctly predicts silicon and silicon-dioxide interface structures, for instance, would not just answer one question. It would become a tool other researchers could reuse across an entire category of semiconductor problems.
The fact that frontier lab models could not make meaningful headway on any of the five, even with internet access and days of autonomous runtime per problem, is itself a data point. It suggests that whatever capability jump let AI systems start closing in on hard math problems has not yet transferred cleanly into the kind of multi-scale physical reasoning that materials science demands. Predicting a stable crystal structure, or explaining why light exposure shifts a material’s electronic behavior, involves chaining together quantum mechanics, statistical mechanics, and experimental data in ways that do not reduce neatly to the pattern-matching that large models are good at.
A business built on faster materials discovery
Discovered Materials is not a research charity. It is a company that raised a seed round, announced in August 2026, and it partners with chip companies on research, meaning its commercial future depends on actually finding materials that semiconductor manufacturers can use. That gives the prize challenge a dual purpose. It is a genuine attempt to advance a field with chronic blind spots, and it is also a way to recruit outside expertise toward problems that sit directly upstream of the company’s own product.
Sridhar’s own timeline for that product has shifted. Back in August, he said he hoped the company would find a useful material within a year. Now he is more bullish. “Those timelines have accelerated a little bit,” Sridhar said, adding, “I’m more confident now that we will be able to find these industrially useful materials in a timeline shorter than a year.”
That confidence sits a little oddly next to the news that frontier models could not crack any of the five challenge problems. The explanation is likely that finding one commercially useful material is a much narrower task than developing a general theory of cuprate superconductivity or a universal fix for Density Functional Theory. A company can stumble into a useful compound through search and iteration without first solving the deep physics behind why it works. The prize problems, by contrast, are aimed at the kind of foundational understanding that would make future searches faster and more reliable across the whole field, not just inside one company’s pipeline.
That ambition requires serious computing power, and it is part of a broader trend of frontier labs backing their research bets with enormous infrastructure spending. Anthropic’s amended S-1 disclosed compute commitments on a scale that underscores just how much capital is now flowing into the kind of compute-intensive research these materials problems could eventually draw on, whether that compute goes toward training bigger models or running the kind of long, autonomous research sessions Discovered Materials already tested on its five challenges.
What happens next
For now, the five problems sit out in the open with price tags attached and no stated process for collecting the money. Academics working in condensed matter physics, computational materials science, or semiconductor modeling have a new, well-defined target, something the field has lacked. Whether Discovered Materials formalizes submission rules, judging criteria, or deadlines will determine whether this becomes a real competition or stays a standing offer that researchers chase informally, citing the challenge in their own published work without ever filing a formal claim.
The bottom line
Putting a price on problems AI couldn’t solve is a smart way to get free publicity and free research labor pointed at exactly the physics Discovered Materials needs solved for its own roadmap. The lack of any disclosed entry process makes this look more like a marketing moment dressed up as a prize than an actual competition right now, and that gap needs to close fast if the company wants serious academics to treat it as more than a press release. Still, the underlying signal is real: AI’s recent wins in math and biology have not yet translated into materials physics, and naming five specific, verifiable problems is a more useful contribution to the field than another vague claim of AI progress.
The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.
