COMPANY PROFILE / MATERIALS INFORMATICS
CuspAI
A UK start-up that calls itself a "search engine for materials", and fills the gaps in experiments and data through its "Foundry"
CuspAI is a materials-design AI company founded in Cambridge, UK, in 2024. Its idea is that you enter the performance you need and it designs candidate materials with generative AI and molecular simulation. In July 2026 it announced the "AI Materials Foundry", a framework that pulls together data, labs, compute and expertise, and at the same time announced a US$450 million Series B (company announcements). This profile separates what official sources confirm from what is still a company claim.
- CuspAI in 30 seconds, and where it stands in materials informatics
- The logo and how the name is written
- Where the company sits in materials informatics
- Scale
- Resources committed to materials informatics
- What it has done so far
- What comes next
- What the company is aiming for
- Map of partnerships
- Technology, products and services
- The risks this company carries
- Glossary / References / Claim-to-source audit
1. CuspAI in 30 seconds, and where it stands in materials informatics
Max Welling (CTO)As stated on the official website
Series A over US$100M
Series B US$450MJune 2024 / September 2025 / July 2026
What the company presents as its strength is "designing candidate materials" Sourced. The work of actually synthesising and measuring a designed material, and the data used for training and evaluation, rely on partner labs and data providers, and the Foundry is what ties these together. Most of the results made public so far are numbers of designed candidates. For example, the PFAS-removal case with Kemira is at the stage where experimental performance checks are still to come (Chapter 10).
2. The logo and how the name is written
A note on names: the company is Cusp AI Limited; the brand and website use CuspAI (cusp.ai). AI Materials Foundry is the name of a framework the company announced, not the name of the company. The logo is taken from the SVG embedded in the official website, with its colour fixed to black.
Official site: https://cusp.ai/3. Where the company sits in materials informatics
Materials informatics (MI) runs data, then AI prediction and design, then narrowing by calculation, then confirmation by experiment. CuspAI covers the first half (design and calculation) with its own platform and leaves experiments and part of the data to partners.
Drawing the flow the company shows on its website, split into the part done by calculation and the part that needs experiments, gives the following.
4. Scale
CuspAI is a private company, and its revenue and headcount could not be confirmed from official sources. The table shows what the register and company announcements let us confirm Sourced.
| Item | Detail | Source |
|---|---|---|
| Legal entity | Cusp AI Limited (UK private limited company, number 15549537). Status: active | Companies House (UK) |
| Incorporated | 9 March 2024 | Companies House (UK) |
| Registered office | 20 Station Road, Cambridge, UK | Companies House / the official privacy notice |
| Sites (company statement) | Cambridge, Amsterdam, Berlin and Tokyo. It said a London (King's Cross) office would open in January 2026 | Official blog (December 2025) |
| Employees | Not disclosed. The company says team size and contract value grew twentyfold in 21 months (base value not disclosed) | Official blog (company claim) |
| Revenue and profit | Not disclosed | - |
| Listing | Unlisted | - |
| Funding (company-stated) | Seed US$30 million (June 2024), Series A over US$100 million (September 2025), Series B US$450 million (July 2026) | Official blog |
Funding amounts are company announcements and do not show revenue or how mature the business is. This article does not touch the valuations or investment views that appear in press coverage.
5. Resources committed to materials informatics
| Angle | What we could confirm | Type |
|---|---|---|
| People | The founders are Prof. Max Welling (CTO; formerly of Microsoft Research and Qualcomm, and professor at the University of Amsterdam) and Dr Chad Edwards (CEO). Prof. Aron Walsh (Imperial College London) joined as Chief Scientific Officer in 2025. Headcount is not disclosed | Sourced |
| Money (R&D) | R&D spend and capital expenditure for MI are not disclosed | Not disclosed |
| Money (funding) | Three rounds as in the table above. Series B was led by Kleiner Perkins and NEA with Bezos Expeditions among participants (company announcement). Series A was co-led by NEA and Temasek, with NVentures (NVIDIA), Samsung Ventures and Hyundai Motor Group taking part | Sourced |
| Investments and acquisitions | No investment in or acquisition of another company by CuspAI could be confirmed | Not confirmed |
| Universities, institutes, public funding | Joint work with Meta and the Georgia Institute of Technology (a dataset of MOFs for direct air capture; Chapter 10). Catalyst research with Topsoe and the Technical University of Denmark (said to have been announced at the 2025 AI in Science Summit). It says it was named in the UK government's "AI for Science" strategy and was among the first to get significant access to the Isambard-AI supercomputer (company claims). No US public grant such as DOE or NSF could be confirmed | Sourced |
Status of the main models, products and facilities
| Name | Kind | Main specifications (company-stated) | Form of release and licence | Stage and date |
|---|---|---|---|---|
| CuspAI platform | A "search engine" for materials (generation + simulation + choosing experiments) | Generates synthesisable candidates from a target performance. Said to be up to 10 times faster than traditional discovery (Series A announcement; basis of comparison not disclosed) | Not public. Offered through joint projects with partners | Deployed with partners (2025 onward) |
| kUPS | Molecular simulation engine (JAX-based) | Lets MD, Monte Carlo and optimisation be combined in one common form. 49 times the throughput of the existing tool RASPA on one adsorption calculation (conditions in Chapter 10) | Open source (Apache-2.0) | Released April 2026 |
| Tojax | Library that converts PyTorch functions to JAX | Brings pre-trained models such as MACE into kUPS | Open source (Apache-2.0) | Released April 2026 |
| ODAC25 | Dataset of MOFs for direct air capture (DAC) | 15,000 MOFs and nearly 60 million DFT single-point calculations. The paper also says trained machine-learning potentials are released | Said in the paper to be released. Licence not checked in this article | August 2025 (with Meta and Georgia Tech) |
| AI Materials Foundry | A network of data, labs, compute and expertise | More than 45 founding partners (company statement). It also says it is setting up US "foundry" operations | Participation is an arrangement between companies | Announced 20 July 2026. US operations: Not yet confirmed |
| Own laboratory facilities | The company's own lab | The size and location of its own lab could not be confirmed from official sources | - | Not confirmed |
In the PFAS-removal case with Kemira, the companies say they searched about 300 trillion candidate structures (metal-organic frameworks, MOFs), produced more than 5,000 designs and narrowed them to about 20 Sourced. What a materials engineer should notice is that this narrowing rests on predicted adsorption performance from calculation. In real water, competing ions and organics, stability in water and ease of regeneration all matter. The company lists "water-stable, synthesisable and low-cost" as design conditions, but measured data for the candidates have not been published.
A figure such as "chosen from trillions" shows how wide the search was; it does not show whether the material works in practice. Until results from the next stage (synthesis and testing) appear, it is safest to read it as a company claim.
6. What it has done so far
| When | What happened | Meaning |
|---|---|---|
| March 2024 | Cusp AI Limited incorporated in the UK (register) | The company's starting point |
| June 2024 | Raised a US$30 million seed round (led by Hoxton Ventures). Said it would work with Meta on open science and the OpenDAC effort | Public debut as a "search engine for materials" |
| March 2025 | Announced an advisory board (including Geoffrey Hinton) | Governance announced |
| July 2025 | Strategic partnership with Kemira. First theme: PFAS removal | The first published industrial partner |
| August 2025 | ODAC25 dataset (with Meta and Georgia Tech) posted on arXiv | Data for DAC MOFs made public |
| September 2025 | Series A (over US$100 million). Announced partnerships with Hyundai Motor Group, Kemira and Meta | Partnerships across sectors |
| 2025 | Catalyst research with Topsoe and the Technical University of Denmark; Prof. Aron Walsh became Chief Scientific Officer | Catalysis and research team |
| April 2026 | kUPS and Tojax released as open source (Apache-2.0) | Tools released for the research community |
| May 2026 | With Kemira, announced the design results for PFAS-removal materials (300 trillion searched, over 5,000 designs, about 20 priority candidates, six months) | A design-stage result; experiments come next |
| July 2026 | Announced the Series B (US$450 million) and the AI Materials Foundry | Scale-up |
Note: for some posts the date shown on the official blog (Medium) and the date in the text disagree. The seed-round post is listed with a January 2025 date, but its text is dated 18 June 2024. This article follows the dates in the text.
7. What comes next
All of these are the company's plans and intentions. Their execution and results could not be confirmed Sourced Not yet confirmed.
None of the following is stated in official sources Not yet confirmed.
The facilities and size of its own lab; the name, scale and training data of the models that run the platform; measured data for the designed materials; the amounts and number of contracts; and what Foundry participants will actually supply (their comments are statements of intent).
8. What the company is aiming for in materials informatics
This part includes things the company has not said directly. It sets out inferences we draw from primary sources, with the evidence.
| What we read from it | Primary-source basis | Type |
|---|---|---|
| It aims to deliver results by taking on design for each customer's problem | Project-by-project partnerships with Kemira (PFAS), Hyundai, Topsoe and others. The founders say commercial partnerships are how they create impact | Inference |
| Its weight is shifting to semiconductors and advanced compute | Foundry participants include Applied Materials, Lam Research, Tokyo Electron, imec and KIOXIA. The 2025 review says it "significantly expanded deployments in semiconductors" | Inference |
| It draws in the research community with open tools and data | It released kUPS and Tojax, published ODAC25 with Meta, and keeps working with Meta | Inference |
| The Foundry is a way of filling the gaps in experiments and data from outside | The Foundry names data providers (CCDC, ICSD, Wiley) and labs (imec, A*STAR and others) and includes "lab validation" in the flow | Inference |
| It claims to run a closed loop that includes experiments | It says results are fed back to influence later decisions. SkyVault carbon-capture materials went from design to experimental validation in six months (company claim; details not disclosed) | Sourced (claim) |
Among the comments by Foundry participants, makers of semiconductor equipment, memory and materials stand out Sourced. For instance Tokyo Electron says the concept "could disrupt the current business model in the industry", and KIOXIA speaks of "new possibilities in semiconductor memory materials" (both are participants' comments, not contract terms).
At the time of the Series A, the main partners were Kemira (water treatment), Hyundai (mobility) and Meta (carbon capture). Set against that, we read the weight as moving toward semiconductors Inference. But participation is a "statement of intent", and which company signed what scale of contract has not been disclosed.
9. Map of partnerships in materials informatics
| Counterparty | Relationship | Type |
|---|---|---|
| Meta (FAIR) | Alliance on open science in June 2024. Works with CuspAI on carbon capture and published ODAC25 (a MOF dataset for DAC) together | Sourced |
| NVIDIA | Collaboration on ALCHEMI and help with GPU tuning of kUPS. NVIDIA's NVentures took part in the Series A | Sourced |
| Kemira | Design of PFAS-removal materials (a commercial partnership). Design results announced jointly in May 2026 | Sourced |
| Hyundai Motor Group | Joint research on materials for next-generation mobility. Also took part in the Series A | Sourced |
| Topsoe and the Technical University of Denmark | Joint catalyst research (said to have been announced in 2025) | Sourced |
| Foundry participants | Applied Materials, Lam Research, Tokyo Electron, Mitsui Chemicals, Fujifilm, Resonac, KIOXIA, Henkel, 3M, imec and others appear with comments or logos | Sourced (intent) |
10. Technology, products and services
CuspAI aims to take a requirement for a material (for example "adsorb only one particular molecule") and return designed candidates that meet it. The platform itself is not public and is provided through joint projects. The table lists the published results and their conditions.
| Item | What the company shows | Conditions and cautions |
|---|---|---|
| Kemira: PFAS-removal materials | A search space of about 300 trillion MOF structures. For three molecules (GenX, PFBS, PFOS), over 5,000 new designs with property data, narrowed to about 20 priority candidates, reached in six months | Results of synthesising and testing the candidates have not been published (they are "advancing to further development"). "The first commercial partnership" is a company claim |
| SkyVault: carbon capture | Carbon-capture materials went from generative design through synthesis to experimental validation in six months | A statement in the company's year review. Measured values and conditions could not be confirmed |
| Speed claim | Produces synthesisable candidates up to 10 times faster than earlier discovery methods (Series A announcement) | Basis of comparison not disclosed |
| kUPS performance | On one NVIDIA L4 GPU, MD of a 114-atom MOF ran 1.8 times faster than MACE through the standard ASE interface. CO2 adsorption (GCMC) ran 49 times faster than RASPA on one CPU core | The company's measurements on two workloads. Performance on other systems is not checked |
| ODAC25 | 15,000 MOFs and nearly 60 million DFT single-point calculations. Adsorption of CO2, H2O, N2 and O2 | Stated in the paper (arXiv). Data for sorbent discovery in DAC |
Note: every figure is a statement by the company or its co-announcers. We could not find reproduction by an independent third party or experimental confirmation.
1. How your own data would be used. The platform is not public, so using it means setting up a joint project. The official flow (steps 2, 3 and 4 going back and forth) reflects the idea that handing over your own experimental data and evaluation criteria brings the design closer to the target.
2. The public tools you can try yourself. kUPS is open source that combines MD and Monte Carlo in one framework and can be used for adsorption calculations in MOFs (Apache-2.0). It also shows how to bring in models such as MACE and UMA that you may already run.
3. The gap between "designed" and "made". For structures proposed by generative AI, whether they can be synthesised is the biggest hurdle. The company says it returns "synthesisable candidates", but its criteria for that have not been published.
11. The risks this company carries
| Risk | Detail | Type |
|---|---|---|
| Published results are at the design stage | The Kemira case stops at choosing priority candidates, with experimental performance not yet public. The six months for SkyVault is also without details | Sourced |
| The platform cannot be checked from outside | The names, scale and training data of the models and the evaluation methods are not public. The basis of "up to 10 times" is not disclosed | Sourced |
| The Foundry is at the stage of intent | Participants' comments are expressions of expectation. Contract terms and the scope of data and labs actually provided are not disclosed. US operations are still being set up | Not yet confirmed |
| Reliance on outside labs and data | Because labs and data depend on partners, partners' circumstances can affect progress | Inference |
| Disclosure to match large funding | Contract amounts, number and revenue are not disclosed. The base of the "twentyfold contract value" is unknown | Sourced |
| Competition in the field | Other firms in this series (Meta, Orbital Materials and others) also release models and data, so differences are hard to see | Inference |
12. Glossary
- Generative AI for materials design
- A method in which AI is asked to propose new material structures from the properties you want. Whether a proposal can really be made is the challenge.
- MOF (metal-organic framework)
- A crystal full of tiny voids built from metals and organic molecules. Used to adsorb gases or substances in water.
- PFAS
- A group of persistent man-made organofluorine compounds. Technology is needed to remove them from water at very low concentrations.
- Direct air capture (DAC)
- Technology that takes CO2 directly from the air. The performance of the sorbent is the key.
- Molecular simulation
- Reproducing how molecules move and adsorb on a computer: molecular dynamics (MD), Monte Carlo methods and so on.
- DFT
- Density functional theory. The standard way to calculate the properties of matter from quantum mechanics: precise but slow.
- Machine-learning potential
- An AI model that predicts the forces on atoms and the energy quickly. Used instead of DFT to gain speed.
- Series A / B
- Names for stages of start-up fundraising. Moving from A to B refers to a later-stage raise.
13. References
- CuspAI Official website. AI Materials Foundry, founders and advisors, the design-to-experiment flow, and partner comments and logos. https://cusp.ai/
- CuspAI Official blog (Medium), "Launching our AI Materials Foundry and $450 million Series B" (20 July 2026). https://medium.com/@CuspAI/launching-our-ai-materials-foundry-and-450-million-series-b-f5603d8259dd
- CuspAI Official blog, "Securing our $100M+ Series A" (10 September 2025). Investors and partners (Hyundai, Kemira, Meta). https://medium.com/@CuspAI/securing-our-100m-series-a-to-revolutionise-materials-discovery-with-ai-e6705e8c8fd3
- CuspAI Official blog, "CuspAI secures $30M" (the text is dated 18 June 2024). Founders, the Meta link and the business description. https://medium.com/@CuspAI/cuspai-secures-30m-to-combat-climate-change-with-ai-designed-materials-e334b4caf560
- CuspAI Official blog, "Our partnership with Kemira" (July 2025). https://medium.com/@CuspAI/pioneering-the-future-of-chemicals-using-ai-our-partnership-with-kemira-f5db8c450504
- Kemira Oyj / CuspAI Kemira stock exchange release (Inderes-hosted page, 21 May 2026). Results of the PFAS-removal materials design (search space, designs, candidates). https://www.inderes.fi/en/releases/kemira-oyj-new-ai-designed-materials-show-promising-potential-to-remove-forever-chemicals-from-drinking-water-in-industry-first-breakthrough
- CuspAI Official blog, "AI-designed materials for PFAS removal" (21 May 2026). https://medium.com/@CuspAI/new-ai-designed-materials-show-promising-potential-to-remove-forever-chemicals-from-drinking-42c8ec5afb11
- CuspAI Official technical blog, "kUPS: a molecular simulation engine for the AI era" (23 April 2026). Benchmarks with conditions. https://medium.com/@CuspAI/kups-a-molecular-simulation-engine-for-the-ai-era-b213963a2359
- GitHub CuspAI's open-source organisation (cusp-ai-oss). kUPS, tojax, widom and others, all Apache-2.0. https://github.com/cusp-ai-oss
- CuspAI Official blog, "A Year at the Cusp: 2025" (18 December 2025). Summary of 2025 partnerships, sites, funding and claims. https://medium.com/@CuspAI/a-year-at-the-cusp-2025-eab5480ea580
- CuspAI Official blog, "Announcing the CuspAI Advisory Board" (5 March 2025). https://medium.com/@CuspAI/announcing-the-cuspai-advisory-board-2d04a7e2f0df
- arXiv Sriram et al., "The Open DAC 2025 Dataset for Sorbent Discovery in Direct Air Capture", arXiv:2508.03162 (August 2025; v2 September 2025). https://arxiv.org/abs/2508.03162
- Companies House (UK) Public record for Cusp AI Limited (company number 15549537). Incorporation date, type and registered office. https://find-and-update.company-information.service.gov.uk/company/15549537
14. Claim-to-source audit
| Claim in the text | Type | Source |
|---|---|---|
| Legal name, incorporation date (9 March 2024), type and registered office | Sourced | Reference 13 https://find-and-update.company-information.service.gov.uk/company/15549537 |
| Founders, leaders and advisors; Foundry participants and their comments | Sourced | Company statement. Reference 1, 10, 11 https://cusp.ai/ https://medium.com/@CuspAI/a-year-at-the-cusp-2025-eab5480ea580 https://medium.com/@CuspAI/announcing-the-cuspai-advisory-board-2d04a7e2f0df |
| Seed round of US$30 million (18 June 2024) | Sourced | Company announcement. Reference 4 https://medium.com/@CuspAI/cuspai-secures-30m-to-combat-climate-change-with-ai-designed-materials-e334b4caf560 |
| Series A (over US$100 million, September 2025), investors and partners | Sourced | Company announcement. Reference 3 https://medium.com/@CuspAI/securing-our-100m-series-a-to-revolutionise-materials-discovery-with-ai-e6705e8c8fd3 |
| Series B of US$450 million, the Foundry announcement (20 July 2026), more than 45 partners | Sourced | Company announcement. Reference 2 https://medium.com/@CuspAI/launching-our-ai-materials-foundry-and-450-million-series-b-f5603d8259dd |
| The partnership with Kemira (July 2025) | Sourced | Reference 5 https://medium.com/@CuspAI/pioneering-the-future-of-chemicals-using-ai-our-partnership-with-kemira-f5db8c450504 |
| PFAS-removal design results (about 300 trillion searched, over 5,000 designs, about 20 priority, six months) | Sourced | Company and partner announcements. No experimental results stated. Reference 6, 7 https://www.inderes.fi/en/releases/kemira-oyj-new-ai-designed-materials-show-promising-potential-to-remove-forever-chemicals-from-drinking-water-in-industry-first-breakthrough https://medium.com/@CuspAI/new-ai-designed-materials-show-promising-potential-to-remove-forever-chemicals-from-drinking-42c8ec5afb11 |
| 2025 review: sites, SkyVault, work with Hyundai, Topsoe and Meta, twentyfold team and contracts | Sourced | Company claim. Base value not disclosed. Reference 10 https://medium.com/@CuspAI/a-year-at-the-cusp-2025-eab5480ea580 |
| Release of kUPS and Tojax, and performance (one L4, 1.8 times ASE, 49 times RASPA) | Sourced | Company measurement conditions. Reference 8, 9 https://medium.com/@CuspAI/kups-a-molecular-simulation-engine-for-the-ai-era-b213963a2359 https://github.com/cusp-ai-oss |
| The open-source organisation and licences (Apache-2.0) | Sourced | Reference 9 https://github.com/cusp-ai-oss |
| The ODAC25 dataset (15,000 MOFs, nearly 60 million DFT calculations) | Sourced | Reference 12 https://arxiv.org/abs/2508.03162 |
| The claim of "up to 10 times faster than before" | Sourced | Company claim. Basis not disclosed. Reference 3 https://medium.com/@CuspAI/securing-our-100m-series-a-to-revolutionise-materials-discovery-with-ai-e6705e8c8fd3 |
| Revenue, headcount, own lab facilities, model details and contract amounts | Not yet confirmed | Could not be confirmed from official sources. Not estimated in this article Reference 1, 2, 10 https://cusp.ai/ https://medium.com/@CuspAI/launching-our-ai-materials-foundry-and-450-million-series-b-f5603d8259dd https://medium.com/@CuspAI/a-year-at-the-cusp-2025-eab5480ea580 |
| US foundry operations, site expansion and sharing of semiconductor results | Not yet confirmed | Company plan. Reference 2, 10 https://medium.com/@CuspAI/launching-our-ai-materials-foundry-and-450-million-series-b-f5603d8259dd https://medium.com/@CuspAI/a-year-at-the-cusp-2025-eab5480ea580 |
| The four types of MI company, where CuspAI sits, and the colour-coded five steps of design | Our calculation | Our own organisation |
| "Design per customer problem", "moving toward semiconductors", "drawing in the community with open tools", "the Foundry fills gaps" | Inference | Our own reading. Evidence is in Chapter 8 |
| The risks of design-stage results, a non-public platform and a Foundry at the stage of intent | Inference | Our own reading. Evidence is in Chapter 11 |
Last updated: 22 October 2026 / Troy Technical
The facts in this article rest on CuspAI's official website and blog, Kemira Oyj's release, the UK Companies House record, a published paper and GitHub. We have not used market estimates from research firms or figures that appear only in press reports. We do not touch valuations or investment judgements. Performance and result figures are statements by the company or its co-announcers, not independent third-party verification. Passages marked "Inference" are our own reading of primary sources. The official blog (Medium) returns HTTP 403 to automated requests, so we read the posts through the company's RSS feed.