In the span of nine months, a Cambridge-based startup nobody heard of last year has gone from a $520 million valuation to $2.6 billion. The reason: it might hold the key to solving one of the most stubborn bottlenecks in industrial progress—finding new materials that don't take decades to discover. CuspAI, founded in 2024 by chemist-turned-entrepreneur Dr. Chad Edwards and AI pioneer Professor Max Welling, closed a $450 million Series B round in July 2026, led by Kleiner Perkins and NEA, with a notable check from Jeff Bezos' Bezos Expeditions. Also chipping in were the UK government's Sovereign AI Venture Fund, the Netherlands' Invest-NL, and names like AMD Ventures, Lux Capital, and John Doerr. The total raised now exceeds $650 million, and the headcount has swelled to around 700. But this isn't just another AI hype machine. The real news is the simultaneous launch of the AI Materials Foundry, a global network stitching together Nvidia's accelerated computing, Meta's open-source Universal Model for Atoms (UMA), and over 45 industrial and academic partners—from Samsung and Hyundai Motor Group to Applied Materials and Lam Research. The pitch: an end-to-end pipeline where AI generates candidate molecules, simulates their properties, plans synthesis routes, and coordinates experimental validation, all in months instead of years. "If we don't make progress fast, the next 50 years of industrial progress will be constrained by a single challenge: the world needs materials that don't yet exist," the co-founders said in a joint statement. It's a line that resonates in semiconductor fabs struggling to replace iridium and ruthenium, and water treatment plants hunting for better ways to filter out PFAS chemicals. The Foundry's Four Pillars The Foundry assembles what CuspAI claims is the first production infrastructure for software-led materials discovery: curated experimental datasets (including exclusive AI training rights to the Cambridge Structural Database and Inorganic Crystal Structure Database), Nvidia's HPC firepower, physical lab capabilities, and the domain expertise to interpret results. It sounds almost too good to be true—and not everyone is buying it. On Hacker News, one commenter noted: "$450M at $2.6B valuation for a company with no commercial product yet? This feels like 2021 all over again." Another shot back: "Materials science has been stuck in the 'decades to discovery' paradigm for too long. Even if CuspAI only accelerates discovery by 2x, that's worth billions to the semiconductor industry." Show Me the Molecules The most concrete proof point so far is a six-month collaboration with Finnish chemicals company Kemira. CuspAI's MIRA platform screened 300 trillion possible structures for PFAS removal materials, delivered over 5,000 designs with property data, and narrowed the field to 20 validated novel candidates—now in further development. That's a hit rate of 0.0000000000000067%, which Reddit users were quick to mock: "20 candidates from 300 trillion? Let's see how many make it to commercial deployment." But the counterargument is equally compelling: traditional methods might have taken a decade to get even to 20 plausible leads. Kemira's president, Antti Salminen, said the project offered "a credible path to next-generation PFAS remediation products." And given that the global PFAS filtration market is projected to grow from $2.34 billion in 2026 to $3.28 billion by 2031, there's real money at stake. The Bezos Connection Bezos' involvement isn't random. His physical-AI lab, Prometheus, raised $12 billion in June 2026 to apply AI to engineering and manufacturing, and he's called materials discovery the biggest challenge in that space. CuspAI gives him a front-row seat—and possibly a first-customer advantage—in the race to find the next generation of materials for chips, batteries, and carbon capture. The UK government's engagement is equally strategic. Its Sovereign AI Venture Fund typically writes checks of £1–10 million, and CuspAI is only its fourth equity investment. "This isn't just about chips—it's about supply chain independence," a tech-policy analyst noted on a forum. With rare metals like iridium and ruthenium concentrated in geopolitically sensitive regions, finding alternatives could rewire the semiconductor supply chain. Open Source vs. Secret Sauce CuspAI has released kUPS, a JAX-based toolkit for atomistic simulations, on GitHub—a move that garnered praise from computational chemists. "Finally, a materials discovery company contributing back to the open-source ecosystem," one developer wrote. But the real value, skeptics argue, is locked in proprietary data agreements and experimental validation pipelines. "Open-sourcing kUPS is smart PR, but the secret sauce is behind their CSD license," another commented. That tension—open models versus proprietary data flywheels—is central to CuspAI's defense. Each customer project generates more validation data, making future predictions more accurate and harder for competitors to replicate. It's a high-touch discovery service masquerading as a software platform, with revenue coming from platform access fees, custom projects, and potentially shared IP economics down the line. The Valuation Puzzle To justify a $2.6 billion price tag, you have to believe that materials discovery is on the cusp of an AI-driven revolution. Schrödinger, the 33-year-old leader in physics-based simulations for drug and materials design, trades at just $1.13 billion with $255 million in annual revenue—and it's losing money. CuspAI, with likely minimal revenue today, is valued at more than double that. Either the market is frothy, or the incumbents have been dangerously slow to adapt. Kleiner Perkins partner Josh Coyne, who co-led the round, is betting on the latter: "CuspAI built a search engine that changes that." But as the Zymergen debacle showed—a once-hyped AI-driven materials company that imploded after its flagship product failed at scale—the gap between computationally neat and industrially manufacturable can be a graveyard. CuspAI's emphasis on synthesis-aware generative models, which screen for manufacturability from the get-go, is meant to avoid that fate. The company's chief scientific officer, Professor Aron Walsh of Imperial College, maintains that the AI Materials Foundry can compress design-to-validation timelines from years to months. That'll be put to the test across offices now open in Cambridge, Amsterdam, Berlin, Tokyo, Singapore, and the US. The question isn't whether the AI works—it's whether any of those 20 Kemira candidates, or the next batch, can be made cheaply enough and at scale. Until that happens, CuspAI remains the most richly valued materials R&D project on the planet.