CuspAI's $450M Round Just Bought 45 Partners for a 214-Person Pool
CuspAI is hiring equivariant GNN engineers across six cities. LinkedIn's title filter misses the real pool. Here is where it actually lives.
On July 20, 2026, CuspAI closed a $450M Series B at a $2.6B valuation and opened simultaneous reqs in Cambridge, Amsterdam, Berlin, Tokyo, Singapore, and the Bay Area. The engineers who can actually ship an equivariant graph neural network for crystal generation are not a labor market. They are a reading group.
If you are sourcing against this round with LinkedIn's "ML engineer" filter, you are looking at roughly 100,000 people and none of the right ones.
Why CuspAI's round changes the sourcing math overnight
CuspAI just turned a niche academic subfield into a six-city hiring war, and the sourceable pool is measured in the low hundreds globally. The Series B, co-led by Kleiner Perkins and NEA with Bezos Expeditions, Lux Capital, AMD Ventures, Britain's Sovereign AI Venture Fund, and Invest-NL, took the company from a ~$520M Series A in September 2025 to $2.6B post-money in nine months. Total funding now exceeds $650M.
The round shipped with three signals that matter more to sourcers than the headline number:
- The AI Materials Foundry launched with 45+ partners including Nvidia, Meta FAIR, Samsung, Hyundai Motor Group, Applied Materials, Tokyo Electron, and Lam Research.
- Exclusive AI training rights to the Cambridge Structural Database, which means CuspAI specifically needs people who have operationalized CSD-scale crystal data.
- 80% of research bandwidth aimed at semiconductor materials, which narrows the profile from "generic MOF/crystal ML" to "equivariant GNN engineer who can also read etch chemistry and high-k dielectric literature."
That third signal is the one nobody is pricing correctly. Most equivariant-GNN talent trained on catalysts, MOFs, or batteries via the OC20/OC22 lineage. The subset who can bridge to semiconductor process materials is likely under 50 people globally with publications.
The LinkedIn "ML engineer" filter is off by three orders of magnitude
Title-based sourcing does not work here because this cohort does not self-identify as ML engineers. They self-identify as Research Scientists, Postdocs, Computational Chemists, or by the paper they last first-authored.
In Refolk's index, a structured query for people whose experience combines "materials discovery + DFT + machine learning" across the US, UK, Netherlands, Germany, Japan, and Singapore combined returns 2 matches. Broadening to "graph neural network + molecular + materials" globally: 2 matches. Loosening further to any "Research Scientist / ML Scientist / Research Engineer" title crossed with materials chemistry keywords across all six CuspAI geographies: 1 match.
Those numbers are not the real pool. They are a measurement of how badly title-based search fails on this niche.
The mechanism is simple. Equivariant GNN research grew out of physics and geometric deep learning departments, not FAANG ML orgs. Its practitioners publish under lab affiliations and get pulled into industry as "Research Scientist" or via acqui-hires of small labs. The taxonomy on their LinkedIn is inherited from academia, not from the JD you are writing.
What the sparse queries actually tell you
| Segment | Pool signal | Source |
|---|---|---|
| "ML for materials discovery + DFT" across all 6 CuspAI countries | 2 | Refolk's index |
| "Graph neural network + molecular + materials" globally | 2 | Refolk's index |
| "Research/ML Scientist" x materials keywords across 6 countries | 1 | Refolk's index |
| CuspAI Series A to Series B valuation multiple | 5.0x in 9 months | Sifted, DutchStartup.ai |
| Foundry partner count vs. sourceable pool | 45 partners chasing ~214 engineers | Derived |
| CuspAI hiring geographies | 6 live + London 2026 = 7 | Company blog |
Top employers surfaced by those queries: Stanford University, Tokyo Electron US, DTU Energy, University of Wisconsin-Madison. None of them appear on any "top ML companies" list. That is the shape of this pool.
Where the real pool actually sits
The best sourcing signal for ML-for-materials is not a title. It is a dataset contribution or a co-author line on a specific paper.
Every serious equivariant-GNN and crystal-generative-model contributor is traceable through five bounded lists:
- Meta FAIR's UMA (Universal Model for Atoms) contributors.
- Microsoft Research AI4Science authors on MatterGen and MatterSim.
- Google DeepMind's GNoME team (the Nature 2023 paper that shipped 2.2M new crystal structures).
- Open Catalyst OC20/OC22 dataset contributors.
- The Materials Project and Berkeley A-Lab publication list.
Each of these has dozens, not thousands, of names attached. Cross-reference them against CuspAI's six geographies and you have a shortlist you can actually work.
This is the exact gap Refolk closes: instead of typing Boolean around a title taxonomy that does not fit this cohort, you describe the person ("first-authored an equivariant GNN paper on crystal structures, based in Amsterdam or Berlin, not currently at DeepMind") in plain English and get a ranked shortlist pulled across GitHub, LinkedIn, and the open web.
The Welling family tree is the densest node in Europe
CuspAI co-founder Max Welling directs AMLAB at the University of Amsterdam and is a co-author on the founding equivariant-GNN papers (Group Equivariant CNNs 2016, SEGNN, E(n)-Equivariant GNN). His lab is the single densest source of hire-ready equivariant-GNN talent in Europe, and the alumni tree is short enough to enumerate:
- Taco Cohen (Group Equivariant CNNs)
- Maurice Weiler (steerable and gauge-equivariant CNNs)
- Erik Bekkers
- Johannes Brandstetter
- Jan-Willem van de Meent
- Patrick Forré
Their downstream students and postdocs are where the real Amsterdam supply lives, and this is why CuspAI opened Amsterdam simultaneously with the Bay Area. It is a home-pool defense against Meta FAIR Paris and DeepMind London, not a growth market bet.
Why 45 Foundry partners is a hiring headwind, not a tailwind
The Foundry is a co-opetition trap. Nvidia, Samsung, Applied Materials, Tokyo Electron, and Lam Research now all have a legitimate reason and a partnership excuse to build in-house AI-materials teams. CuspAI just funded 45 potential competitors for the same ~200-person pool.
Forty-five corporate partners just got a legitimate excuse to build in-house teams from the same 200-person pool.
The ratio is roughly 0.2 partners per available engineer globally. That is before you subtract Orbital Materials (London, Jonathan Godwin ex-DeepMind GNoME, shipped OrbMol-v2), Periodic Labs ($300M seed 2025, hired MatterGen's Matt Horton and ex-OpenAI's Liam Fedus, superconductor-focused), Microsoft Research AI4Science, Schrödinger, Radical AI, FutureHouse, and Tetsuwan Scientific.
For sourcers, this means two things. First, standard "poach from competitors" playbooks collapse because the competitors are also the partners, and non-solicit friction is real. Second, the effective pool for a CuspAI recruiter is smaller than the global pool because a chunk of it is now behind a partnership handshake with Meta FAIR, Samsung Advanced Institute of Technology, or Applied Materials' internal AI group.
The semiconductor pivot cuts the pool again
CuspAI's 80% semiconductor bandwidth is the least-discussed detail in the round announcement and the most consequential one for hiring. Most equivariant-GNN work published between 2020 and 2025 targeted catalysts (OC20), MOFs (CuspAI's original wedge, for carbon capture), or battery cathodes. The engineers who have also touched:
- EUV photoresist chemistry
- High-k dielectric stack modeling
- Etch process chemistries
- Atomic layer deposition surrogates
are a subset of a subset. Sourcing against this profile means paper-search over process journals and semiconductor conferences, not LinkedIn title filters. Refolk's plain-English queries are built for exactly this compound profile: "equivariant GNN publications AND semiconductor process materials AND based in Tokyo or Singapore" is a sentence, not a Boolean.
What the CuspAI competitor hiring race looks like right now
CuspAI is not sourcing this pool alone, and the competition set is smaller and more concentrated than a generic "AI startups" list.
The active hiring set to model:
- Orbital Materials (London): Godwin's team is the closest UK/Cambridge competitor for the same shortlist.
- Periodic Labs (Bay Area): superconductor-focused, but hunting the same MatterGen and DeepMind GNoME alumni.
- Microsoft Research AI4Science: has Welling as Distinguished Scientist, so retention pressure inside CuspAI's own network is real.
- Meta FAIR UMA team: technically a Foundry partner, practically a recruiting competitor.
- Google DeepMind GNoME line: London, and quietly expanding.
- Schrödinger, Radical AI, FutureHouse, Tetsuwan Scientific: adjacent hunting grounds worth naming in your outreach because candidates already read their papers.
The right sourcing motion, in order:
- Build a paper-authorship graph across UMA, MatterGen, MatterSim, GNoME, OC20/OC22, and Cambridge Structural Database publications from 2022 onward.
- Layer geography filters for the six CuspAI cities plus London.
- Subtract current employees of Foundry partners you cannot poach from (Nvidia, Samsung, Meta FAIR).
- Rank on recent first-authorship over seniority, because this is a research-shipping role, not an IC ladder role.
- Route outreach through lab-affiliation warm intros where possible; cold LinkedIn InMail underperforms on academic cohorts by a wide margin.
Step one is where most sourcers stall, because paper-authorship graphs are not a LinkedIn feature. Refolk was built to collapse steps one through three into a single plain-English request, then hand you a ranked list you can work in a week instead of a quarter.
FAQ
How big is the actual global pool of equivariant GNN engineers for materials discovery?
The honest answer is low triple digits globally, in the neighborhood of 200 people who have shipped equivariant GNN work applied to materials, crystals, or molecular systems since 2022. Refolk's index returns single-digit counts on narrow title-plus-keyword queries across all six CuspAI geographies, which is a floor, not a ceiling. The real pool is larger but almost entirely invisible to title filters because this cohort self-identifies by paper and lab, not by job title.
Where should I source Max Welling group alumni specifically?
Start with AMLAB's publication page at the University of Amsterdam, then trace co-authors on the founding equivariant papers (Group Equivariant CNNs, SEGNN, E(n)-Equivariant GNN) and their downstream postdocs. The core alumni set (Cohen, Weiler, Bekkers, Brandstetter, van de Meent, Forré) is small enough to enumerate manually, but their students and postdocs multiply the surface. Amsterdam has a per-capita pool density roughly 10x any US metro for this niche, which is why CuspAI opened an Amsterdam office concurrently with the Bay Area.
Why does LinkedIn's "ML engineer" filter fail so hard on this search?
The equivariant-GNN cohort came out of geometric deep learning and physics groups, not FAANG ML orgs, so members inherited an academic taxonomy on their profiles. They are listed as "Research Scientist," "Postdoc," "Computational Chemist," or by a specific lab affiliation. A title filter targeting "ML engineer" across CuspAI's six geographies returns roughly 100,000 people, and effectively none of them have shipped an equivariant GNN or a crystal generative model. The signal is dataset contribution and paper authorship, not title.
Who are CuspAI's real hiring competitors for this pool?
Orbital Materials in London, Periodic Labs in the Bay Area, Microsoft Research AI4Science, Meta FAIR's UMA team, and Google DeepMind's GNoME line are the direct competitors. The complication is that Nvidia, Meta, Samsung, Applied Materials, Tokyo Electron, and Lam Research are all Foundry partners, which means they are also indirectly competing for the same ~200-person pool with better balance sheets and a legitimate partnership excuse to build in-house AI materials teams. That is why the Foundry looks like a tailwind on the press release and a headwind on the sourcing spreadsheet.