Bezos pursues next-generation silicon investments, signaling Amazon's chip-supply vertical integration strategy.

On July 20, British AI materials company CuspAI announced the completion of $450 million in Series B funding, valuing the company at $2.6 billion.

The round was led by Kleiner Perkins and New Enterprise Associates (NEA). Investors include Bezos Expeditions, the investment arm of Jeff Bezos, the UK government-backed sovereign AI fund, and organizations including AMD Ventures and Samsung Ventures.

A $2.6 billion valuation is not unprecedented among AI-native companies, but it is exceptionally high for a vertical science company—one that has existed for less than two years.

While mainstream AI competition centers on large language models and agents, CuspAI represents a different track. Bronze launched the Bronze Age, steel built the Industrial Revolution, silicon created computers, and lithium enabled electric vehicles. Over past decades, the semiconductor industry has continuously shrunk transistor size by relying on silicon, driving persistent performance gains. But as processes approach physical limits, industry focus increasingly shifts toward new semiconductor materials, interconnect materials, and advanced packaging in search of the next breakthrough.

The renewable energy industry faces a similar challenge. Compared to lithium-ion batteries widely in use today, solid-state batteries are seen as potentially delivering simultaneous gains in safety, energy density, and range. Yet from Toyota to QuantumScape to CATL, the industry has yet to see large-scale commercialization.

This year alone, the academic world published multiple review papers on solid electrolyte discovery, including "Breaking Bottlenecks in Solid Electrolyte Discovery with Large Artificial Intelligence Models" and "Machine Learning Pipelines for the Design of Solid-State Electrolytes."

To eliminate PFAS (perfluoroalkyl and polyfluoroalkyl substances)—persistent pollutants that are virtually non-biodegradable—researchers continue to search for new adsorbent materials with higher efficiency and lower cost. The data center industry's growing computational demands have also made thermal management materials a critical variable affecting energy consumption.

The reason is straightforward: finding a new material is not like searching a database. Scientists typically begin with a hypothesis, then design material structure, conduct computational simulation, run experiments, and adjust based on results. One failure means starting over.

A material's performance depends not only on its elemental composition but also on atomic arrangement, crystal structure, defect state, and manufacturing process. Even tiny variations can produce drastically different results.

Facing such an enormous search space, humans increasingly struggle to rely on experience and experimentation alone to find truly valuable new materials. Material development is time-consuming not only in the lab but in identifying which candidate compounds are worth testing among countless possibilities.

CuspAI's Series B is a recent case in point. Capital is beginning to bet on a new possibility: the next "silicon" might emerge from an AI-driven materials development system.

The company was co-founded by two complementary leaders. CEO Chad Edwards previously co-founded quantum computing company Cambridge Quantum Computing (CQC) and oversaw its merger with Honeywell's quantum business to create Quantinuum. Co-founder Max Welling is a prominent machine learning scholar, former Distinguished Scientist at Microsoft Research, and co-author of the influential paper "Auto-Encoding Variational Bayes," with significant impact in generative modeling.

Historically, materials development started from an existing material. Scientists designed a new structure, then tested its properties through computational simulation and experiment. If results were unsatisfactory, they adjusted and started over.

CuspAI adopts an "inverse design" approach: rather than scientists first thinking "what material should I design," they first define targets—such as higher conductivity, greater heat resistance, or lower manufacturing cost—then have AI reverse-engineer possible material structures that might meet those requirements and predict their properties, helping scientists identify the most promising candidates to verify.

In other words, it does not replace experimental work but aims to let scientists spend more time validating the most promising materials rather than searching blindly through countless possibilities.

CuspAI has not, like Google DeepMind's GNoME, announced representative new materials, nor claimed that any AI-designed material has reached commercialization.

One of CuspAI's most commercially validated cases is helping Finnish chemical company Kemira find new materials to remove PFAS contamination. The company reports that its model screened candidate materials from approximately 3 trillion potential structures, narrowing them to 20 for further development—though these candidates currently remain in the verification phase with no announced commercial results.

On July 20, CuspAI announced not only the $450 million Series B funding but also the formation of the AI Materials Foundry. The round was led by Kleiner Perkins and NEA, with participation from Bezos Expeditions, AMD Ventures, Samsung Ventures, and others. Simultaneously, the AI Materials Foundry has gathered over 45 partners, including NVIDIA, Meta, Applied Materials, and Hyundai Motor Group.

Some provide compute and chips, some control manufacturing capabilities, some possess real industrial demand, and some can manage follow-up experiments and industrial validation.

The list's significance lies not only in the number of partners CuspAI has assembled but in what it signals to capital: a market far larger than individual material development.

A traditional materials company's value typically depends on whether it can produce a certain material and win contracts. CuspAI seeks to span the entire research chain—from identifying material needs, generating candidates, predicting properties, completing verification, through to industrial deployment.

Should this model prove viable, every industry reliant on materials progress—semiconductors, automotive, batteries, chemicals, aerospace—could become its customer base.

This is the imagination behind the $2.6 billion valuation: CuspAI has a chance to evolve from a company seeking new materials into a research platform used across multiple industries.

The AI Materials Foundry and participation from NVIDIA, Meta, Applied Materials, Hyundai, and others provide real footholds for this vision. It shows CuspAI is organizing models, compute, industrial demand, and experimental capability into a collaborative network, and it convinces investors they have a chance to participate in building this infrastructure.

If material development eventually evolves into a new collaboration model—AI generates candidate materials, industry specifies needs, compute performs simulations, labs verify results—then competitive advantage will no longer rest solely on any single company's research capabilities but on who can enter this system and work with more participants to complete material research.

This valuation certainly rests on a future yet to be realized. But in AI materials, capital is precisely willing to prepay for the enormous market that could emerge once this chain fully operates.

CuspAI is not the first to attempt using AI to reshape materials development. Before it, Google DeepMind already proved that AI can extend material search to scales once unimaginable. MatNex began designing materials around concrete industrial needs. Orbital pushed materials models further into real industrial scenarios.

These players chose different paths yet collectively answered one question: how far has AI materials discovery actually come?

If AlphaFold brought AI into life sciences, then GNoME (Graph Networks for Materials Exploration) represents DeepMind's significant foray into materials science.

In 2023, Google DeepMind released the GNoME project. The system used graph neural networks to predict crystal structure stability and discovered over 2.2 million potential crystal structures, of which approximately 380,000 were predicted to be stable—roughly ten times the number of stable materials known to humanity. This includes 528 potential lithium-ion conductors, about 25 times the number from prior work, offering far more battery candidates.

Traditional materials databases stem from past computational and experimental accumulation, with researchers typically finding patterns within ranges already documented by humans. GNoME demonstrated another possibility: AI can actively explore crystal structures not yet in any database by applying learned material principles.

To verify these predictions were not merely computational artifacts, DeepMind partnered with Lawrence Berkeley National Laboratory. The institution's A-Lab system can use algorithms to generate experimental plans, control robots for material mixing, heating, and testing, and ultimately successfully synthesized over 40 new materials.

However, GNoME primarily addresses "which crystal structures might be stable." Stability is not the same as utility; successful synthesis does not guarantee better conductivity, magnetism, heat resistance, or lower manufacturing costs. A gap remains before genuine industrial production: performance validation, process development, and scaling.

If DeepMind answered "Can AI find new materials?", MatNex asks a different question: Can AI design materials according to industrial requirements?

MatNex, originally Materials Nexus, was founded in 2020 and is a deep-tech company spun out from Cambridge University.

Like CuspAI, Materials Nexus uses an "inverse design" approach: based on target performance requirements, it uses AI to find possible new materials that could meet needs rather than relying on traditional trial-and-error experiments.

Rare-earth permanent magnets are widely used in electric vehicles, motors, and wind turbines but depend heavily on specific rare-earth resources and supply chains. In 2024, MatNex announced that its AI platform screened and designed rare-earth-free permanent magnet material MagNex from over 100 million candidate combinations.

The material went from design, synthesis, to testing in just three months—dramatically shorter than traditional industrial development cycles. The company projects MagNex costs could reach approximately 20% of traditional rare-earth magnets while reducing material carbon emissions by 70%.

This case goes further than simply predicting one material—it fulfills a specific industrial need through a complete process of setting targets, algorithm screening, experimental synthesis, and performance testing.

AI's value thus transcends simply "discovering more possibilities" to actively helping development teams find answers worth manufacturing faster.

But MagNex's successful synthesis does not directly equate to completed commercialization. It proves AI can shorten early development cycles, but whether it can extend this speed into scaled production remains to be seen.

Orbital Materials was founded in 2022 and is headquartered in London before later becoming Orbital Industries.

Orbital initially chose a path similar to DeepMind's "AI for Science" approach: training foundational models that understand atomic structure, 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 opened-sourced related models, seeking to lower barriers for researchers conducting atomic-scale materials simulation.

As its work progressed, however, it gradually shifted focus from materials discovery to industrial application, beginning to design concrete solutions for energy, carbon capture, and AI infrastructure.

One project uses AI to design porous CO₂ adsorbent materials for direct air capture and attempted deployment in data centers. The company reports that since establishing its laboratory in 2024, performance of these materials has improved approximately tenfold.

Orbital's concern extends beyond finding better-performing materials to how these materials enter specific equipment and industrial contexts. The materials model thereby becomes a starting point for solving practical problems rather than the final product.

This makes Orbital increasingly resemble an AI industrial company: materials discovery remains its core capability, but what it ultimately sells might be not just models or materials but solutions designed around specific industrial challenges.

From Google DeepMind to MatNex, Orbital, and CuspAI, several distinct development paths have emerged in AI materials.

DeepMind seeks to expand the material space humans can explore; MatNex advances AI design to synthesis and testing around clear needs; Orbital embeds materials models in concrete scenarios like carbon capture and data centers; CuspAI aims to organize models, compute, experiments, and industrial partners into a research platform usable across multiple industries.

AI can now predict crystal structures previously unrecorded, has already helped researchers synthesize new materials, and is beginning to shorten design and testing cycles around specific industrial demands.

Yet regardless of which path is chosen, all must confront the same challenge: AI finding a material that might exist is not the same as humans possessing a material they can use.

From computational prediction to experimental synthesis, from sample to stable manufacturing, through to scaled deployment, each step can eliminate vast numbers of candidates. The natural world does not change its verification rules because model capabilities improve.

Therefore, AI's clearest present value remains shortening the search process, preserving limited experimental resources for more promising directions.

It can make "finding a needle in a haystack" faster, but it cannot guarantee the needle found will ultimately support a new industry.

The next "silicon" might be found first by AI, but what truly defines "the next silicon" remains the real world.