What would you build if you could ask a machine to invent the material you need instead of hunting for it?
That question now carries a $2.6 billion valuation. CuspAI, a Cambridge-based startup founded in 2024, closed a $450 million Series B in July with Kleiner Perkins and NEA leading and Jeff Bezos's Bezos Expeditions participating. Co-founder Chad Edwards, a chemist who helped grow Quantinuum, was named to this year's TIME100 AI list.
Its first proof point: generative AI designed 20 candidate materials to strip PFAS "forever chemicals" from drinking water, in six months
The open question for investors: does a designed molecule survive contact with real water?
The pitch fits in one line. Specify the properties you need, and the machine returns chemical compositions that can actually be built, with a synthesis route and a lab that can make them.
Co-founder Max Welling, a University of Amsterdam professor who co-invented the variational autoencoder that underpins modern generative design, puts the ambition plainly. "My dream is that we'll be able to stand in the world and point at certain devices or equipment and say 'that is powered by our materials.'"
A search engine for matter
Materials science still runs on trial and error. Researchers guess a composition, synthesize it, test it, fail, and repeat. The space of possible structures runs into the hundreds of trillions. Humans have explored a rounding error of it.
CuspAI's platform, MIRA, inverts the workflow. A partner types in the property target: conductivity, thermal tolerance, water stability. The model generates candidate molecules, then routes the winners to a lab that can synthesize them. The company says the loop runs ten times faster than conventional R&D.
The customer list reads like a semiconductor industry roll call: ASML, Meta, Samsung, Hyundai. In July CuspAI also announced the AI Materials Foundry, a coalition of more than 45 companies and labs. NVIDIA, Meta, Henkel, Applied Materials, Tokyo Electron and Lam Research are in. Data, compute and lab capacity pool under one roof.
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.— Chad Edwards, co-founder and CEO, CuspAI
That is the actual bottleneck, not a slogan. Better batteries, cheaper carbon capture, denser chips: all of it waits on materials that have not been discovered. The question is whether generative models are the tool that finds them, or just the newest way to burn compute.
The forever-chemical case
Kemira, a Finnish water-treatment company with €2.8 billion in annual revenue, gave CuspAI its clearest test. In May the two announced the first commercial partnership to apply generative AI end-to-end to PFAS remediation material design.
PFAS earned the nickname forever chemicals because the carbon-fluorine bond is among the strongest in organic chemistry. The EPA set national drinking-water limits in parts per trillion in 2024; the EU Drinking Water Directive did the same. The incumbent fix, granular activated carbon, filters the compounds out but does not destroy them. The poison just moves to another stream.
Kemira's brief was specific: design materials that remove PFAS at trace concentrations, using chemistry that is stable in water, cheap to make and easy to scale. CuspAI's model explored roughly 300 trillion candidate structures and returned more than 5,000 designs with full property data for three target molecules: GenX, PFBS, PFOS. The partnership narrowed the field to about 20 priority candidates now moving into testing.
The discovery phase took six months. Traditional routes take years.
As we wrote in August, AI-designed catalysts began rewriting the carbon capture playbook, and Carb Enzero pushed enzyme carbon capture from lab to pilot. The PFAS work takes the same approach at a harder target: chemistry that breaks the molecule down instead of containing it.
Kemira challenged us with finding new solutions to one of the most pressing environmental problems of our time, and in six months our partnership delivered.— Chad Edwards, co-founder and CEO, CuspAI
Kemira's CEO Antti Salminen calls it "a credible path toward a next-generation PFAS remediation product." Read that word carefully.
Credible, not deployed.
These 20 candidates are still designs. None has yet spent a year inside a municipal water plant.
The gap between the model and the water
This is where the caution begins. A structure that works in simulation is not a product. AI-designed chemistry has a short history and familiar failure modes: molecules the model loves but synthesis cannot make, stability that evaporates outside the test tube, cost curves that never reach the plant.
CuspAI's own framing concedes the distance. Kleiner Perkins partner Josh Coyne says the founders "design for materials that can actually be built, not just ones a model can dream up." Synthesis-aware is the operative word. The model is trained to respect what physical laboratories can construct.
What sets Max, Chad, and the team apart is that they design for materials that can actually be built, not just ones a model can dream up.— Josh Coyne, partner, Kleiner Perkins
The field is crowded and well funded. Orbital Materials was spun out by DeepMind alumni. XtalPi sits at a $2.5 billion valuation, Lila Sciences has raised over $550 million, and Periodic Labs, built by former OpenAI and DeepMind staff, is worth $1 billion. CuspAI is not alone in believing AI can close the materials gap. It is simply the best capitalized.
And there are non-AI routes to the same problem. Claros Technologies raised $55 million in July to scale ultraviolet PFAS destruction in industrial facilities, no generative model involved. SBIR-funded teams are engineering dehalogenase enzymes to break carbon-fluorine bonds directly, and an iGEM team of high schoolers ran their own LLM-designed PFAS-degrading enzyme to a bioRxiv preprint in 2024.
The money is flowing because the regulatory clock is loud. Minnesota is spending hundreds of millions on groundwater cleanup. Manufacturers face mounting liability. Whoever ships a PFAS solution that is selective, stable and cheap will own a market regulators have just created.
Edwards and Welling reportedly turned down a billion-dollar acquisition offer months into the company's life. "We know that we're building a generational company," Edwards told Northzone. "To let that go into somebody else's hands or in a different direction would have felt very premature."
The next 18 months separate the thesis from the product. Can MIRA compress research cycles from years to months across 45 partners at once? Or does AI materials discovery repeat the oldest pattern in technology: the gap between what the model promises and what the plant proves?
The investment case was never about the $450 million. It is about whether one of those 20 PFAS candidates survives contact with real water. That is the entire bet.
