Jeff Bezos, founder of Amazon, has invested in another AI company. This time, his bet is on whether AI can help humanity find the next "silicon." On July 20, UK-based AI materials company CuspAI announced the completion of a $450 million Series B funding round, valuing the company at $2.6 billion. The round was led by Kleiner Perkins and New Enterprise Associates (NEA), with participation from Bezos Expeditions, the UK government-backed Sovereign AI Fund, AMD Ventures, and Samsung Ventures. While a $2.6 billion valuation is not unprecedented among AI-native companies, it is notably high for a vertical science company, especially one founded less than two years ago. If mainstream AI competition revolves around large models and agents, CuspAI represents a different path: can AI move from processing existing information to helping humans discover the unknown?
Humanity has always been searching for the next material. If the development of industrial civilization were condensed into one sentence, it would be a history of constantly finding new materials. Bronze opened the Bronze Age, steel powered the Industrial Revolution, silicon created computers, and lithium drove the development of new energy vehicles. Behind every major industrial transformation, there is almost always a key material supporting it. Today, humanity is still searching for the next "silicon." Over the past decades, the semiconductor industry has relied on silicon to shrink transistor sizes and continuously improve computer performance. But as manufacturing processes approach physical limits, the industry is increasingly looking toward new semiconductor materials, interconnect materials, and advanced packaging for the next breakthrough. The new energy industry faces similar challenges. Compared to the widely used liquid lithium batteries, solid-state batteries are considered promising for improving safety, energy density, and range. However, from Toyota and QuantumScape to CATL, the industry has yet to see large-scale commercialization. This year alone, multiple review papers on solid electrolyte discovery have appeared, including "Breaking Bottlenecks in Solid Electrolyte Discovery with Large Artificial Intelligence Models" and "Machine Learning Pipelines for the Design of Solid-State Electrolytes." Catalysts, aerospace, advanced manufacturing, and environmental protection face similar difficulties. To remove PFAS (per- and polyfluoroalkyl substances), "forever chemicals" that hardly degrade naturally, researchers continue to search for new adsorbent materials that are more efficient and cost-effective. The growing computing demands of data centers also make thermal management materials a key variable affecting energy consumption. More and more industries are finding that engineering capabilities are advancing rapidly, but material breakthroughs are becoming slower. The reason is not hard to understand: finding a new material is not like searching for an answer in a database. Scientists typically need to propose a hypothesis, design the material structure, perform computational simulations and experimental validation, and then adjust the plan based on results. One failure means starting over. Moreover, the possibilities in the material world are almost infinite. The performance of a material depends not only on elemental composition but also on atomic arrangement, crystal structure, defect states, and manufacturing processes. Even tiny changes can lead to vastly different outcomes. Faced with such a vast search space, it is increasingly difficult for humans to rely on experience and experimentation to find truly valuable new materials one by one. Thus, a new idea emerged: the time-consuming part of materials R&D is not just experiments, but also finding the few candidates worth testing among countless possibilities. And that is precisely what AI excels at. It is against this backdrop that a batch of new companies focused on AI-driven materials discovery have begun to attract capital attention. CuspAI's funding round is the latest example. Capital is starting to bet on a new possibility: the next "silicon" might be born from an AI-driven materials R&D system.
Why is Bezos betting on this company? CuspAI is a UK-based AI materials company founded in 2024, headquartered in Cambridge. The company was co-founded by two founders with complementary backgrounds. CEO Chad Edwards previously co-founded quantum computing company Cambridge Quantum Computing (CQC) and drove its merger with Honeywell's quantum business to form Quantinuum. Co-founder Max Welling is a prominent scholar in machine learning, a former distinguished scientist at Microsoft Research, and a co-author of the classic paper "Auto-Encoding Variational Bayes," with significant influence in generative models. In the past, materials R&D typically started from an existing material. Scientists would design a new structure, then test its properties through computational simulation and experiments; if results were unsatisfactory, they would adjust the structure and start over. CuspAI's MIRA attempts to change this process. It adopts an "inverse design" approach: researchers do not need to first think about "what material should be designed," but instead define the target—such as higher conductivity, better heat resistance, or lower manufacturing cost—and then AI generates material structures that might meet these requirements in reverse, predicting their properties to help scientists screen the most promising candidates for validation. In other words, it does not replace scientists in conducting experiments, but aims to let scientists spend more time validating the most promising materials rather than searching for a needle in a haystack. However, to date, CuspAI has not delivered a stunning track record. It has not published representative new materials like Google DeepMind's GNoME, nor has it announced that an AI-designed material has achieved industrialization. Currently, one of CuspAI's closest cases to commercial validation is helping Finnish chemical company Kemira find new materials to remove PFAS pollutants. The company says MIRA screened candidates from approximately 300 trillion potential material structures, narrowing down to 20 for further development, but these candidates are still in the validation stage, with no commercial results announced yet. So the question arises: why can an AI company that has not yet delivered decisive results achieve a $2.6 billion valuation? One detail is worth noting. On July 20, CuspAI announced not only the $450 million Series B round but also the establishment of AI Materials Foundry. The round was led by Kleiner Perkins and NEA, with participation from Bezos Expeditions, AMD Ventures, Samsung Ventures, and others. Meanwhile, AI Materials Foundry has gathered over 45 partners, including Nvidia, Meta, Applied Materials, and Hyundai Motor Group. This list spans AI, semiconductors, advanced manufacturing, automotive, and materials industries. Some provide computing power and chips, some have manufacturing capabilities, some have real industrial needs, and others can undertake subsequent experiments and industrial validation. The significance of the list is not just how many partners CuspAI has attracted, but more importantly, it shows capital a market far larger than single-material R&D. A traditional materials company's value usually depends on whether it can make a certain material and win orders; CuspAI aims to cover the entire R&D chain from material requirement generation, candidate generation, performance prediction, validation, to industrial application. Once this model works, all industries that rely on material progress—chips, automobiles, batteries, chemicals, aerospace—could become its customers. This is the imagination behind the $2.6 billion valuation: CuspAI has the potential to transform from a company that finds new materials into a shared R&D platform for multiple industries. AI Materials Foundry, and the participation of Nvidia, Meta, Applied Materials, Hyundai, and others, provides a realistic anchor for this imagination. It shows that CuspAI is organizing models, computing power, industrial needs, and experimental capabilities into a collaborative network, and it also makes investors believe they have the opportunity to participate in building this infrastructure. If materials R&D gradually evolves into a new collaborative model—where AI generates candidate materials, industries propose needs, computing power completes simulations, and laboratories validate results—then competitiveness will no longer be determined solely by one company's R&D capability, but by who can enter this system and collaborate with more participants to complete materials R&D. The $2.6 billion prices the possibility of CuspAI becoming a materials R&D platform. This valuation is still based on an unfulfilled future. But in the AI materials field, capital is willing to pay in advance precisely because of the huge market that could be covered once this chain runs smoothly.
Has AI already found new materials? CuspAI is not the first company to try to use AI to reconstruct materials R&D. Before it, Google DeepMind had already proven that AI can expand materials search to scales previously unimaginable; MatNex began designing materials around specific industrial needs; Orbital further pushed materials models into real industrial scenarios. These players have chosen different paths and collectively answered a question: how far has AI materials discovery actually come? Google DeepMind is the most prominent representative. If AlphaFold brought AI into life sciences, then GNoME (Graph Networks for Materials Exploration) is DeepMind's important attempt in materials science. In 2023, Google DeepMind released the GNoME project. The system used graph neural networks to predict the stability of crystal structures, discovering over 2.2 million potential crystal structures, of which about 380,000 were predicted to be stable materials, nearly ten times the number of known stable materials. This included 528 potential lithium-ion conductors, about 25 times the number from previous similar research, providing more candidates for next-generation batteries. Traditional materials databases are mainly accumulated from past calculations and experiments, and researchers typically search for patterns within the range already recorded by humans. GNoME demonstrated another possibility: AI can actively explore crystal structures not yet in databases based on learned material patterns. But predicting that a material is stable does not mean it can be manufactured in reality. To verify that these predictions are not just in computers, DeepMind subsequently collaborated with Lawrence Berkeley National Laboratory. The latter's A-Lab system can use algorithms to generate experimental plans, control robots to complete material mixing, heating, and testing, and successfully synthesized over 40 new materials. However, GNoME mainly addresses "which crystal structures might be stable." Stability does not equal practicality, and being synthesizable does not mean having better conductivity, magnetism, heat resistance, or lower manufacturing cost. There is still a gap of performance validation, process development, and large-scale manufacturing before entering industrial production. If DeepMind answers "can AI find new materials," then MatNex is more concerned with another question: can AI design materials according to industrial needs? MatNex, formerly Materials Nexus, was founded in 2020 and is a deep-tech company incubated from the University of Cambridge. Similar to CuspAI, Materials Nexus adopts an "inverse design" route: based on target performance requirements, use AI to find new materials that might meet the requirements, rather than relying on traditional experimental trial-and-error. One of the company's earliest directions was rare-earth-free permanent magnets. Rare-earth permanent magnets are widely used in new energy vehicles, motors, and wind power, but they heavily depend on specific rare earth resources and supply chains. In 2024, MatNex announced that its AI platform screened and designed the rare-earth-free permanent magnet material MagNex from over 100 million candidate combinations. The material went from design, synthesis, to testing in just 3 months, significantly shortening the traditional industrial materials R&D cycle. The company estimates that MagNex could cost about 20% of traditional rare-earth magnets while reducing material carbon emissions by 70%. This case goes a step further than merely predicting a material—it revolves around a clear industrial need and completes the process from setting goals, algorithm screening, to experimental synthesis and performance testing. The value of AI is thus not just "discovering more possibilities," but also helping R&D teams quickly find a set of answers worth manufacturing. But MagNex being successfully synthesized does not directly equate to completed commercialization; it proves AI can shorten early R&D cycles, but whether it can extend this speed to large-scale production remains to be seen. Another typical player in the AI materials track also comes from the UK. Orbital Materials was founded in 2022, headquartered in London, and later renamed Orbital Industries. Orbital initially chose a path similar to DeepMind's "AI for Science": training foundation models that understand atomic structures, material properties, and physical laws, then using these models to predict material performance and screen candidate structures. In 2024, the company released the materials simulation model Orb and made it open, aiming to lower the barrier for researchers to conduct atomic-level materials simulations. But as business progressed, it gradually shifted focus from materials discovery to industrial applications, designing specific solutions around energy, carbon capture, and AI infrastructure. One effort is using AI to design porous CO₂ adsorbent materials for direct air capture, attempting to deploy them in data centers. The company says that since establishing its laboratory in 2024, the performance of this material has improved about 10-fold. Orbital is concerned not just with whether it can find better-performing materials, but also how these materials enter specific equipment and industrial scenarios. Materials models thus become the starting point for solving practical problems, not the final product. This also makes Orbital increasingly like an AI industrial company: materials discovery remains its core capability, but what it ultimately sells may not be just models or materials, but a set of solutions designed around specific industrial problems. From Google DeepMind to MatNex, Orbital, and CuspAI, several different development paths have emerged in the AI materials field. DeepMind tries to expand the material space humans can explore; MatNex advances AI design to synthesis and testing around clear needs; Orbital embeds materials models into specific scenarios like carbon capture and data centers; CuspAI hopes to organize models, computing power, experiments, and industrial partners into a R&D platform usable by multiple industries. These advances show that AI materials discovery is not entirely conceptual. AI can already predict crystal structures not previously recorded, has helped researchers synthesize new materials, and is even beginning to shorten design and testing cycles around specific industrial needs. But no matter which path they choose, they all must face the same challenge: AI finding a possible material does not mean humanity already has a usable material. From computational prediction to experimental synthesis, from experimental samples to stable manufacturing, to large-scale application, each step may eliminate many candidates. The natural world will not change validation rules because of model capabilities. Therefore, AI's clearest value at present is still to shorten the search process, leaving limited experimental resources for more promising directions. It can make "finding a needle in a haystack" faster, but it cannot yet guarantee that the needle picked up will ultimately support a new industry. The next "silicon" may be found by AI first, but what truly defines the "next silicon" is still the real world.