CuspAI Just Hit $2.6B on 68 People. The Equivariant-GNN Pool Lives in 3 Labs.
CuspAI's $450M Series B opened a seven-city hiring war for equivariant-GNN engineers. Here is where the tiny pool actually sits and how to source it.
On July 20, 2026, Cambridge-based CuspAI closed a $450M Series B at a $2.6B valuation, a 5x jump in nine months, and launched the AI Materials Foundry with 45+ partners including Nvidia, Meta FAIR, Samsung, ASML and Hyundai. The company employs roughly 68 people and is now hiring simultaneously in Cambridge, London, Amsterdam, Berlin, Tokyo, Singapore and the Bay Area. The problem: the pool of engineers who can actually ship an E(3)-equivariant graph neural network on crystal structures lives inside three or four labs.
Why CuspAI's raise is really a talent story
The headline is $2.6B. The real number is $38M raised per employee, which means the constraint is not capital, it is the count of humans who understand equivariant message passing on atomic graphs.
CuspAI's own math tells the story: ~$650M total raised, ~68 heads, and seven active hiring cities. That per-employee density puts it in Anthropic/Mistral territory, not "materials startup" territory. When a company is that capital-dense and that small, every senior hire moves the org chart by a percentage point and moves the market comp band by a step.
The second signal is co-founder Max Welling, who co-invented the variational autoencoder and ran the Amsterdam Machine Learning Lab (AMLab) at UvA. Welling is not on the cap table for optics. He is the reason Amsterdam is one of the seven hiring cities, because the densest per-capita source of equivariant-GNN engineers outside Boston and London is his own former lab.
The pool is three institutions, not a market
The playbook for equivariant graph neural networks on materials was written by three teams, and nearly every qualified senior IC on the planet has trained inside one of them.
The three institutions:
- DeepMind GNoME (London). The Nequip-style equivariant GNN implemented in e3nn-jax that discovered 2.2 million candidate crystals and 381,000 predicted-stable materials. Named ICs include Amil Merchant and Ekin Dogus Cubuk.
- Meta FAIR Chemistry (Pittsburgh/NYC). The UMA (Universal Models for Atoms) family, built on the eSEN equivariant GNN and trained on ~500M atomic systems across molecules, materials and catalysts. Brandon Wood is the lead author.
- The NequIP/MACE academic cluster. Simon Batzner (NequIP, now at Radical AI/DeepMind), Ilyes Batatia and Gábor Csányi (MACE, Cambridge Cavendish), plus Welling's AMLab alumni behind E(n)-equivariant GNNs.
Everyone else, and this includes the ML teams inside Schrödinger, XtalPi ($2.5B), Orbital Materials, Microsoft's MatterGen group, Entos, Toyota Research Institute, and Isomorphic Labs on the adjacent side, is essentially a fourth-generation copy of one of those three lineages. That is not a knock. It is a sourcing map. If you cannot name which of the three someone trained under, you probably have not found a real one.
Why the moat is people, not GPUs
NequIP, the foundational architecture, outperformed prior models with up to three orders of magnitude fewer training examples. That single result reshaped the field's economics. When your model needs 1,000x less data to hit accuracy, the bottleneck stops being compute and becomes the small number of people who understand the group-theoretic tricks (SO(3)-equivariant tensor products, spherical harmonics, message-passing on E(3)) that make the sample efficiency work.
That is why CuspAI can raise $450M and still be short-staffed. There is no market you can buy your way into. There is a group chat.
The seven-city hiring map
CuspAI is hiring in Cambridge, London, Amsterdam, Berlin, Tokyo, Singapore and the US, and each city maps to a specific talent source you should be sourcing against right now.
| City | Real reason it's on the map | Who to source first |
|---|---|---|
| Cambridge (UK) | Csányi/Batatia MACE group at the Cavendish | MACE contributors on GitHub, Cavendish PhD alumni |
| London | DeepMind GNoME alumni pipeline | Ex-DeepMind Chemistry/Materials ICs, Isomorphic adjacencies |
| Amsterdam | Welling's AMLab, E(n)-equivariance origin | UvA AMLab PhD alumni, Qualcomm AI Research NL |
| Berlin | TU Berlin ML + Fritz Haber Institute crossover | DFT-to-ML converts, BASLEARN, MPI-CBG |
| Tokyo | Tokyo Electron / Hitachi / RIKEN AIP for materials | RIKEN AIP, Preferred Networks matlantis team |
| Singapore | A*STAR IHPC, Temasek portfolio proximity | NTU/NUS computational chemistry PhDs |
| US (Bay/Boston) | Meta FAIR Chemistry, Materials Project (Berkeley), Radical AI | Ex-FAIR UMA authors, LBNL Persson group, MIT DMSE |
The mistake most recruiters make is treating this as one search. It is seven searches, each with its own reference set of 20 to 50 people, and the reference sets barely overlap. Sourcing "ML + materials" on LinkedIn from a US desk will surface postdocs who published one paper on a graph convnet and call it a day. That is the exact gap Refolk closes: you describe the person in plain English (say, "ex-DeepMind or Meta FAIR chemistry ICs who have shipped an equivariant GNN, based in EU") and get a ranked shortlist across GitHub, LinkedIn and the open web, rather than a keyword soup.
What "qualified" actually means on the JD
A qualified equivariant-GNN engineer for a CuspAI-tier role has shipped a model in one of three architecture families (NequIP/Allegro, MACE, or eSEN/UMA-style), knows E(3) group theory well enough to debug an irreps mismatch, and has touched real DFT or experimental crystal data.
Use this as a filter, not a wishlist:
- Architecture fluency. Has authored or contributed to a repo in the e3nn, MACE, NequIP, SchNetPack, or FAIR Chemistry (fairchem) ecosystem. GitHub commit history beats resume claims here.
- Physics grounding. Can explain why permutation-equivariance is not enough and rotation-equivariance matters for energy prediction. If they cannot, they are not senior for this work.
- Data pipeline experience. Has worked with Materials Project, OQMD, Alexandria, or Open Catalyst datasets. Bonus if they have run VASP or Quantum ESPRESSO to generate training labels themselves.
- Machine learning interatomic potential (MLIP) production. Has deployed an MLIP for MD or structure relaxation at scale, not just benchmarked one on MD17.
Anyone who checks three of four is in the pool. Anyone who checks one is a training project, which at $38M per head is not what CuspAI's investors bought.
The Foundry is a poaching signal, not a partnership
The AI Materials Foundry's 45+ founding partners, including Nvidia, Meta, Samsung, Hyundai, Applied Materials, Tokyo Electron, Lam Research, Henkel and Merck, effectively hand CuspAI a warm-intro channel into the internal ML-materials teams of half the semiconductor supply chain.
Read that partner list as a recruiter, not as a corp dev person. Every senior IC inside Applied Materials' ML group, Lam's computational team, or Tokyo Electron's process modeling group now has a plausible path to a CuspAI conversation via a Slack channel they share with the Foundry team. If you are hiring against CuspAI at Orbital Materials, XtalPi, Radical AI, Schrödinger or a stealth competitor, assume your target's inbox already has a message from someone the target respects.
When 45 semiconductor giants share a Slack with your top competitor, your retention plan is the offer letter you send this quarter.
The counter-move is not a bigger comp package alone. It is speed. The half-life of an equivariant-GNN IC being "available" once CuspAI's Foundry channel warms up is measured in weeks, not quarters. Recruiters who run a monthly rescan of the same 200 target profiles will lose to recruiters who get a fresh signal the day a target updates a GitHub bio, changes their LinkedIn headline, or posts a preprint. Refolk's re-rank on new signals is built for exactly that cadence.
The five moves that actually work right now
The winning sourcing plays for AI materials science hiring in Q3 2026 are narrow, fast, and skip the generic keyword pool entirely.
- Mine UvA AMLab and Cavendish MACE alumni first. Both groups have public alumni pages and thesis repositories. Reference-check via co-author graphs, not LinkedIn endorsements.
- Rank by GitHub contribution graph on e3nn / MACE / NequIP / fairchem. Commit history in the last 12 months is a stronger signal than any title. Sourcing computational chemistry engineers through GitHub outperforms Boolean-on-LinkedIn by a wide margin here.
- Attend ML4Materials at NeurIPS and the AI4Science workshop. Every serious contributor either presents, reviews, or shows up in the hallway. The room is small enough that an evening of coffees maps 30% of the global pool.
- Poach adjacent, not identical. MLIP engineers at Toyota Research and Argonne can retrain on inorganic crystals in a quarter. Waiting for a pure "crystal ML" person adds six months.
- Skip the recruiter template. These candidates get one boilerplate InMail a week from CuspAI's in-house team. A three-sentence note that names their NequIP PR or their MACE benchmark result gets replies. A "hi {first_name}" does not.
For ML materials discovery talent specifically, the two failure modes are (a) confusing generative-molecular-design people (drug discovery, protein) with inorganic crystal-structure people (they are not interchangeable) and (b) hiring on paper count rather than shipped-model count. The Foundry's rare-metals framing (iridium, ruthenium substitutes for EUV lithography) means CuspAI needs the second cohort, not the first.
What this predicts for the next 90 days
Expect three things: CuspAI will hire 30 to 40 more people by year end, Orbital Materials will raise a defensive round, and comp for equivariant-GNN seniors in London and Amsterdam will re-price up by roughly one Anthropic band.
The 30 to 40 number is straightforward math. CuspAI has ~$650M in the bank, seven open cities, and a stated Foundry roadmap that requires domain leads per partner vertical. Even a conservative headcount plan doubles the company inside 18 months. That doubling has to come from the same three-institution pool that Meta FAIR, DeepMind, Microsoft MatterGen, and Radical AI are also fishing in.
The right response for a founder or eng leader competing in this space is not "post the job and wait." It is a named list of 60 to 120 people, refreshed weekly, with a specific opening line per person. That is the workflow Refolk is built for: describe the target ("MACE or NequIP contributors, PhD from Cambridge, Amsterdam or Berkeley, currently at DeepMind, Meta, Microsoft, Schrödinger or a national lab"), get a ranked shortlist, then work it as a pipeline instead of an inbox.
CuspAI's round did not just move a valuation. It compressed the timeline for everyone else hiring against the same 100-person pool. The recruiters who treat this like a normal search will lose the next four hires. The ones who treat it like a targeted operation, with named references, GitHub evidence, and a weekly re-rank, will win them.
FAQ
How large is the equivariant-GNN materials talent pool globally?
There is no clean public number, and any recruiter who quotes one is guessing. What is verifiable: the field's foundational architectures (NequIP, MACE, eSEN/UMA, E(n)-equivariant GNNs) trace to fewer than a dozen senior authors, and the qualified senior IC pool that can ship these systems in production almost certainly lives inside DeepMind GNoME, Meta FAIR Chemistry, the NequIP/MACE academic cluster, and a handful of national labs. Practically, plan your outreach against a target list in the low hundreds, not thousands.
Why is Amsterdam one of CuspAI's seven hiring cities?
Because Max Welling, CuspAI's co-founder, ran the Amsterdam Machine Learning Lab at UvA, which produced foundational equivariance work including the E(n)-equivariant GNN and the variational autoencoder. AMLab alumni are one of the densest per-capita sources of the exact skill set CuspAI is buying. Amsterdam also gives CuspAI proximity to Qualcomm AI Research NL and Invest-NL, one of the Series B backers.
Who are CuspAI's most direct competitors for these hires?
Orbital Materials (a DeepMind spinout), XtalPi (valued at $2.5B), Radical AI, Microsoft's MatterGen team, Google DeepMind's GNoME group and Meta's FAIR Chemistry team are the direct competitors. Schrödinger, Entos, PhaseCraft and PASQAL are adjacent. Isomorphic Labs competes for the same equivariant-GNN talent even though its target domain is biological, because the underlying skill set transfers.
What is the fastest way to source computational chemistry engineers with real equivariant-GNN experience?
Skip Boolean searches on job titles. Rank by GitHub contribution history to e3nn, MACE, NequIP, SchNetPack and fairchem repos over the last 12 months, cross-reference with recent ML4Materials and AI4Science workshop author lists, and enrich against LinkedIn only to confirm current employer and location. This is exactly the shape of query Refolk is built for: plain-English description in, ranked shortlist with cross-source evidence out.