字母榜精选

wrote a column ·

Bezos Uses AI to Search for the Next Silicon

Jeff Bezos, founder of Amazon, has invested in another AI company.

This time, his bet is on whether AI can help humanity discover the next 'silicon.'

On July 20, UK-based AI materials company CuspAI announced it had closed a $450 million Series B funding round, valuing the company at $2.6 billion.

The $2.6 billion valuation isn’t unprecedented among AI-native companies, but it is exceptionally high for a firm focused on a specialized scientific domain—especially considering CuspAI was founded less than two years ago.

While mainstream AI competition revolves around large models and agents, CuspAI represents an alternative path:

Can AI evolve from processing existing information to helping humanity discover the unknown?

Humans have always been searching for the next breakthrough material.

If we were to condense the development of industrial civilization into a single sentence, it would resemble a history of continuously seeking new materials.

Bronze ushered in the Bronze Age, steel powered the Industrial Revolution, silicon enabled computing, and lithium has driven the rise of electric vehicles.

Behind nearly every major industrial transformation lies a critical material as its foundation.

Today, humanity is still searching for the next 'silicon.'

For decades, the semiconductor industry has relied on silicon to continually shrink transistor sizes, enabling sustained improvements in computing performance. However, as process nodes approach physical limits, the industry is increasingly turning its attention to new semiconductor materials, interconnect materials, and advanced packaging technologies in search of the next breakthrough.

The new energy sector faces similar challenges. Solid-state batteries are widely seen as promising alternatives to today’s liquid lithium-ion batteries, potentially offering simultaneous improvements in safety, energy density, and range. Yet despite efforts by companies like Toyota, QuantumScape, and CATL, large-scale commercialization has remained elusive.

This year alone, academia has published multiple review papers focused 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.'

Catalysis, aerospace, advanced manufacturing, and environmental protection sectors face similar challenges.

Researchers continue to search for new adsorbent materials that are more efficient and less costly to remove PFAS (per- and polyfluoroalkyl substances), a class of 'forever chemicals' that are nearly impossible to degrade naturally; meanwhile, the ever-growing computing power demands of data centers have made thermal management materials a critical factor influencing energy consumption.

An increasing number of industries are realizing thatengineering capabilities are advancing rapidly, but breakthroughs in materials are becoming slower and slower.

The reason is not hard to understand: discovering a new material is not as simple as searching for an answer in a database. Scientists usually need to formulate a hypothesis first, then design the material’s structure, run computational simulations, conduct experimental validation, and iteratively refine their approach based on results. A single failure often means starting over from scratch.

Moreover, the space of possible materials is virtually infinite.

A material’s properties depend not only on its elemental composition but also on the combined effects of atomic arrangement, crystal structure, defect states, and manufacturing processes. Even minor changes can lead to dramatically different outcomes.

Faced with such an enormous search space, it is becoming increasingly difficult for humans to rely solely on experience and experimentation to incrementally identify truly valuable new materials.

Thus, a new idea is emerging:

What makes materials research so time-consuming and labor-intensive isn’t just experimentation itself—it’s identifying, among countless possibilities, the few candidate solutions worth testing.

And this is precisely the kind of problem AI excels at solving.

Against this backdrop, a new wave of companies focused on AI-driven materials discovery has started attracting investor attention.

CuspAI's latest funding round is the most recent example of this trend. Investors are now betting on a new possibility: the next 'silicon' may emerge from an AI-powered materials research and development system.

Why is Bezos backing this company?

CuspAI is a UK-based AI materials company founded in 2024 and headquartered in Cambridge.

The company was co-founded by two complementary founders. CEO Chad Edwards previously co-founded quantum computing firm Cambridge Quantum Computing (CQC) and led its merger with Honeywell’s quantum business to form today’s Quantinuum. Co-founder Max Welling is a prominent scholar in machine learning, formerly a Distinguished Scientist at Microsoft Research and one of the authors of the seminal paper 'Auto-Encoding Variational Bayes,' wielding significant influence in the field of generative models.

Traditionally, materials research starts with an existing material. Scientists first design a new structure and then evaluate its performance through computational simulations and experimental testing; if results fall short, they tweak the structure and start over.

CuspAI’s MIRA platform aims to transform this process.

It adopts an 'inverse design' approach: rather than starting by asking 'what material should we design?', researchers first define their target properties—such as higher conductivity, greater heat resistance, or lower manufacturing cost—and then let AI generate candidate material structures that could meet those requirements, predicting their properties to help scientists prioritize the most promising options for experimental validation.

In other words, it doesn’t replace scientists in the lab but aims to let them spend more time onvalidating the most promising materialsrather than searching for a needle in a haystack among countless possibilities.

However, as of now, CuspAI has not delivered results impressive enough to shake the industry.

It has neither announced representative new materials like Google DeepMind’s GNoME nor declared that any AI-designed material has already reached commercialization.

Currently, one of CuspAI’s closest cases to commercial validation involves assisting Finnish chemical company Kemira in discovering new materials to remove PFAS contaminants. The company stated that MIRA narrowed down candidate solutions from approximately 300 trillion potential material structures to 20 for further development, but these candidates remain in subsequent validation stages, with no commercial outcomes disclosed yet.

So the question arises:Why would an AI company that hasn’t yet delivered definitive results command a $2.6 billion valuation?

One detail may be worth noting.

On July 20, CuspAI announced not only its $450 million Series B funding round but also the launch of the AI Materials Foundry.

This round was co-led by Kleiner Perkins and NEA, with participation from Bezos Expeditions (Jeff Bezos’s investment vehicle), AMD Ventures, Samsung Ventures, and others. Meanwhile, the AI Materials Foundry has already assembled over 45 partners, including NVIDIA, Meta, Applied Materials, and Hyundai Motor Group.

This list spans the AI, semiconductor, advanced manufacturing, automotive, and materials industries.

Some provide computing power and chips, some possess manufacturing capabilities, some have real industrial needs, and others can undertake subsequent experimentation and industrial validation.

The significance of the list lies not only in how many partners CuspAI has brought on board, but more importantly, in how it has shown capital a market far larger than that of individual material development.

The value of a traditional materials company typically depends on whether it can produce a certain material and secure orders; CuspAI, however, aims to cover the entire R&D chain—from the initial articulation of material requirements, through candidate generation and performance prediction, to validation and eventual industrial adoption.

If this model proves viable, every industry reliant on materials innovation—including semiconductors, automotive, batteries, chemicals, and aerospace—could become its potential client.

This is precisely the vision underpinning its $2.6 billion valuation:CuspAI has the opportunity to evolve from a company searching for new materials into a shared R&D platform serving multiple industries.

Its AI Materials Foundry initiative, along with participation from NVIDIA, Meta, Applied Materials, Hyundai Motor, and others, provides tangible grounding for this vision.It demonstrates that CuspAI is assembling models, computing power, industrial demand, and experimental capabilities into a coordinated network, giving investors confidence that they can take part in building this foundational infrastructure.

If materials R&D gradually evolves into a new collaborative model—in which AI generates candidate materials, industries articulate requirements, computing power runs simulations, and laboratories validate results—then competitiveness will no longer hinge solely on any single company’s R&D prowess, but rather onwho can enter this ecosystem and collaborate with more participants to complete materials development.。

The $2.6 billion valuation reflects the market's assessment of CuspAI’s potential to become a materials R&D platform.

Of course, this valuation is still based on an unrealized future. But in the AI-driven materials space, what capital is willing to pay for upfront is precisely the vast market that could be unlocked once this pipeline proves viable.

Has AI already discovered new materials?

CuspAI is not the first company attempting to reinvent materials research with AI. Before it, Google DeepMind had already demonstrated that AI can scale materials discovery to previously unimaginable levels; MatNex began designing materials around specific industrial needs; and Orbital has further pushed materials modeling into real-world industrial applications.

These players have chosen different paths, yet together they address a key question:How far has AI-driven materials discovery actually come?

Google DeepMind is the most closely watched representative among them.

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

In 2023, Google DeepMind launched the GNoME project.The system uses graph neural networks to predict the stability of crystal structures and identified over 2.2 million potential crystal structures, approximately 380,000 of which were predicted to be stable—nearly ten times the number of stable materials previously known to humanity.

This includes 528 potential lithium-ion conductors, roughly 25 times the number found in prior similar research efforts, offering many more candidates for next-generation batteries.

Traditional materials databases primarily stem from accumulated computational and experimental data from the past. Researchers typically search for patterns within the scope of what has already been documented by humans, whereas GNoME demonstrates an alternative possibility: AI can proactively explore crystal structures that have not yet entered existing databases based on learned material rules.

However, a computer’s prediction that a material is stable does not necessarily mean it can be synthesized in reality.

To verify that these predictions were not confined to simulations, DeepMind subsequently collaborated with Lawrence Berkeley National Laboratory. The latter’s A-Lab system uses algorithms to generate experimental protocols and control robots to carry out tasks such as material mixing, heating, and testing,ultimately succeeding in synthesizing more than 40 new materials.

Nevertheless, GNoME primarily addresses the question of 'which crystal structures might be stable.' Stability does not equate to practicality, and successful synthesis does not guarantee superior electrical conductivity, magnetism, thermal resistance, or cost-effectiveness. Significant hurdles—including performance validation, process development, and scale-up manufacturing—still stand between discovery and actual industrial production.

If DeepMind answers the question 'Can AI discover new materials?', then MatNex focuses on another: 'Can AI design materials according to industrial requirements?'

Originally named Materials Nexus, MatNex is a deep-tech company spun out of the University of Cambridge and founded in 2020.

Similar to CuspAI, Materials Nexus also employs an 'inverse design' approach: using AI to identify novel materials that meet specified performance targets, rather than relying on traditional trial-and-error experimentation.

One of the company’s earliest focus areas was rare-earth-free permanent magnets.

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

The material was designed, synthesized, and tested in just three months—a significant reduction compared to the traditional industrial materials development cycle. The company estimates that MagNex could cost approximately 20% of conventional rare-earth magnets while reducing material-related carbon emissions by 70%.

This case goes beyond merely predicting a single material—it completed an end-to-end process centered on a clear industrial need, from goal setting and algorithmic screening to experimental synthesis and performance testing.

Thus, AI’s value is no longer just about 'discovering more possibilities'—it can also help R&D teams quickly identify a set of viable candidates worth manufacturing.

However, the successful synthesis of MagNex does not automatically equate to commercialization. While it demonstrates that AI can shorten the early-stage R&D cycle, whether this speed can be extended to large-scale production remains to be seen.

Another prominent player in the AI-driven materials space also hails from the UK.

Orbital Materials was founded in 2022 and headquartered in London, later rebranding as Orbital Industries.

Orbital initially pursued an 'AI for Science' approach similar to DeepMind’s: training foundational models capable of understanding atomic structures, material properties, and physical laws, then leveraging these models to predict material performance and screen candidate structures.

In 2024, the company released its materials simulation model, Orb, and made related models publicly available, aiming to lower the barrier for researchers conducting atomistic materials simulations.

But as its business progressed, the company gradually shifted its focus from materials discovery toward industrial applications, designing specific solutions for energy, carbon capture, and AI infrastructure.

One such initiative involves using AI to design porous CO₂ adsorbent materials for direct air capture and attempting to deploy them in data centers. The company stated,Since establishing its lab in 2024, the performance of this material has improved by approximately tenfold.

Orbital is now focused not only on whether it can find better-performing materials, but also on how these materials can be integrated into specific devices and industrial applications. Material models have thus become the starting point for solving real-world problems, rather than end products themselves.

This shift is increasingly making Orbital resemble an AI-driven industrial company: while material discovery remains its core capability, what it ultimately sells may no longer be just models or materials, but comprehensive solutions tailored to specific industrial challenges.

From Google DeepMind to MatNex, Orbital, and CuspAI, the AI-for-materials field has already branched into several distinct development paths.

DeepMind aims to expand the space of materials humans can explore; MatNex focuses on clear requirements and pushes AI-designed materials all the way through synthesis and testing; Orbital embeds material models into concrete applications like carbon capture and data centers; and CuspAI seeks to integrate models, computing power, experimentation, and industry partners into a unified R&D platform usable across multiple sectors.

These advances demonstrate that AI-driven material discovery is not purely conceptual.

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

But regardless of which path they choose, all must confront the same fundamental challenge:AI identifying a potentially viable material does not mean humans already possess a usable one.

From computational prediction to experimental synthesis, from lab-scale samples to stable manufacturing, and finally to large-scale deployment—each step can eliminate numerous candidate materials. The natural world does not relax its validation rules simply because model capabilities improve.

Therefore, AI’s clearest value at present remains its ability to shorten the search process, allowing limited experimental resources to be focused on more promising directions.

It can make 'finding a needle in a haystack' faster, but it still cannot guarantee that the needle retrieved will ultimately support an entirely new industry.

AI might be the first to discover the next 'silicon,' but it is still the real world that ultimately defines what that 'next silicon' truly is. $Amazon (AMZN.US)$$Star Tech Companies (LIST2518.US)$$Warren Buffett Portfolio (LIST2999.US)$$Artificial Intelligence (LIST2136.US)$$Artificial Intelligence (LIST23586.HK)$$Technology (LIST20763.US)$$Technology (LIST20840.HK)$

6359 Views