Company teardown · Periodic Labs

How to invent a better superconductor with no physics degree?

A frontier-AI research crew, a robotic chemistry lab and $605M — aimed at one material that doesn't exist yet. We read 71 of the ~72 identifiable staff behind it, and the roster tells a different story than the launch announcement did.

Get the full visual brief as a PDF.
Founded
2025 · Menlo Park
People
71 of ~72 staff
Raised
$605M · $300M seed
North star
A high-Tc superconductor
Sells
Intelligence for R&D

0.1 — Periodic Labs in one page

A frontier-AI research crew, a robotic chemistry lab and $605M — aimed at one material that doesn't exist yet.

71
People in 17 months
Hired in the order the bets required — the AI crew first, from the founders' own networks; the laboratory now, from fabs and national labs.
$605M
Raised, starting at $300M
A $300M seed out of stealth with zero products, customers or public code — and a ~$7B valuation in reported talks a year later.
41: 18
Bits to atoms, today
Launch week framed it as half AI, half physical scientists. A year on, our read of the roster is 41 to 18.
1
The idea
Nature as the grader
Today's AI learned math and code because there was always a grader. Science has no grader — so Periodic is building a laboratory where nature grades the model.
2
The business
An intelligence layer
Not a materials company: “a software business” sold to R&D teams in semiconductors, space and defence. The lab is how the model earns the right to be sold.
3
The founders
Bits and atoms
Liam Fedus helped build ChatGPT at OpenAI. Ekin Dogus Cubuk led GNoME at DeepMind, which predicted 2.2 million new crystals. One from bits, one from atoms.
4
The team
Seventy-one people
Hired in the order the bets required: the AI crew first, from the founders' own networks; the laboratory now, from fabs and national labs.
5
The lessons
What transfers
Work the founders' ledger, don't filter on domain, hire recruiters before HR — and know that this order costs a $300M seed.

Everything here is public — LinkedIn, funding disclosures, the job board, and three founder appearances with full transcripts. The reading is ours.

§ 01 — Part one of five · The idea

What if a model could run its own experiments?

Periodic's founders think AI progress in science is stuck on data — and that the fix is a laboratory, not a bigger dataset.

1.1 — The problem they're solving

Why a model that has read every paper still can't do chemistry.

Reason oneThe data is noisyAsk the literature for a material's property and the answers span orders of magnitude — different samples, different methods. A model trained on that inherits the noise.
Reason twoThe failures are missingExperiments that don't work never get published. The model learns only from the wins — the opposite of how a scientist learns.
Reason threeReading isn't actingScience advances by proposing, testing and being wrong. A model that can only read the past cannot be corrected by the present.

— The founders' framing

“Nature is our RL environment.”

— Liam Fedus, a16z Podcast, September 2025.

— RL environment, translated

Reinforcement learning: a model tries something, gets a score, improves. The recent leaps in math and code came from graders that score answers automatically. Periodic's grader is the experiment — the lab exists to run it.

1.2 — The idea

An AI scientist with hands.

Models propose materials and recipes; simulation filters them; a robotic lab makes and measures them and hands the result back. Every cycle yields experimental data that exists nowhere else.

01 · Propose

Frontier models An LLM that reads the literature and drives specialist materials networks as tools — the “AI scientist.”

02 · Filter

Physics simulation First checks on stability — on legacy “complicated Fortran code” the team is still mastering, says Cubuk.

03 · Test

Robotic laboratory Robots mix and heat precursor powders, then measure what formed. Humans still work the bench beside them.

↺ measurements return as training signal — nature grades the model

What the robots makePowder synthesisThe workhorse of solid-state chemistry: grind precursor powders together, heat them in a furnace, see what crystal forms. Repetitive steps — which is why a robot can do it.
What the robots measureCharacterisationMeasuring what you made. X-ray diffraction reveals the crystal structure; resistance against temperature reveals whether it superconducts, and below what temperature.

1.3 — The north star

One target for the whole company: a better superconductor.

— What's a superconductor?

A material that carries electricity with zero resistance — no heat, no loss — but only below a critical temperature. The best known work at about −138 °C (135 K) at ambient pressure, so they live in liquid-nitrogen baths. One that worked warmer would change power grids, MRI machines, chips and fusion magnets.

— Why this target

It's robust
A phase transition is “robust to some of these details that we cannot simulate yet” — a target a simulation-heavy team can aim at.
It matters before it ships
A 200 K result “says so much about the universe” — worth having even if no product follows.
It unites the team
“Quite rare to find a topic that unites the whole team”: physicists, chemists and ML researchers all want it. Chosen partly as a hiring tool.

“Real scientific progress has never come from a static dataset.”

— Ekin Dogus Cubuk, RAISE Summit, Paris, 2026.

The academic bench behind the target: Z.-X. Shen and Steven Kivelson (Stanford), Mercouri Kanatzidis and Chris Wolverton (Northwestern) sit on the advisory board.

§ 02 — Part two of five · The business

A software business that owns a lab.

Periodic isn't planning to sell materials. It plans to sell the intelligence that finds them — to industries that already run R&D labs.

2.1 — The product, as described

An intelligence layer for R&D — sold one problem at a time.

01
Who buys
R&D teams in semis, space and defence
Industries that already run costly physical experiments and hold decades of proprietary results in silos.
02
What they get
A co-pilot for the lab
Three asks Fedus hears: automate the simulations, put the data in one place, and go beyond search — a model trained on the customer's own results.
03
How it's sold
Land and expand
One well-scoped problem with clear evals, then the next. Delivered by forward-deployed engineers — the first joined in July 2026; two more reqs are open.
04
What it isn't
A discovery-royalty model
“A software business,” says Fedus — not a biotech-style bet on owning a molecule. The superconductor is the proof, not the product.
05
Today
Partnerships and grants
No product yet. A university group's syntheses partly failed; Periodic's AI explained why. A grant programme and the advisory board carry the go-to-market for now.

Forward-deployed engineer: an engineer embedded with the customer, building on their data — the Palantir motion, adopted by AI companies.

2.2 — The money

A $300M seed, then $605M in a year.

$300M
Seed round — out of stealth, September 2025.
$605M
Raised in total by mid-2026.
~$7B
Valuation in reported talks — Bloomberg.
0
Products, customers or public code when it was priced.
I1
a16z · Founders Fund · Nvidia
The institutions
Lead capital. a16z put both founders on its podcast in launch week — the richest public record of how the company was designed.
I2
Radical Ventures · Conviction
Innovation Endeavors · Paradigm
Deep tech and AI-native. Toronto's Radical Ventures — not the rival Radical AI — beside Eric Schmidt's and Sarah Guo's funds.
I3
Bezos · Schmidt · Dean · Gil · Deng
Personal cheques
The angels. Google's chief scientist backing two ex-Googlers; Elad Gil then hosted Fedus on No Priors. The investors are also the distribution.

Where it goes, per Fedus: “primarily a compute cost.” The GPUs cost more than the laboratory — but the lab “has very large lead times,” so it was started early. As automation improves, the bottleneck moves back to intelligence.

2.3 — Who else is trying

Five neighbours, one talent market.

01
Lila Sciences
Same bet, older
Flagship Pioneering's “scientific superintelligence” — also pairs models with autonomous labs. Periodic hired its lab-automation principal, Sam Cross, onto the founding team.
02
Radical AI
Same bet, models first
AI for materials. The category's loudest public voice. Two of its alumni — Riebesell and Gangan — now wear Periodic's badge.
03
CuspAI · Discovered Materials
Same cohort
The set, widening. CuspAI on generative materials design; Discovered Materials surfaced in August 2026 with AI-found semiconductors and an open benchmark.
04
SparksMatter
The opposite bet
MIT's Buehler lab runs the whole discovery cycle purely in software. If that works without a lab, Periodic's wager is wrong — this is the control group.
05
The A-Lab
Ceder Group · Berkeley Lab
The origin. The canonical academic autonomous lab, and GNoME's experimental partner. Periodic hired out of it — Bernardus Rendy, nearly five years there.

For a recruiter these are one market — and Periodic is winning alumni from two named rivals and the field's founding lab.

Talent brief

Want the full visual breakdown?

Download the PDF version of this teardown — the three-pillar org chart and its six sub-teams, ten names to know, the build order month by month, and the five moves worth stealing.

§ 03 — Part three of five · The founders

One helped ship ChatGPT. The other found 2.2 million crystals.

Two Google alumni, one from each side of the bits-and-atoms line, sharing the CEO title.

3.1 — Who they are

Two Google alumni, one from each side of the line.

— Liam Fedus · Co-founder & co-CEO · the bits half

Physics
A physicist first: dark-matter research before machine learning.
Mila
PhD at Mila, Université de Montréal — root of the company's Canadian ring.
Google Brain
First author of the Switch Transformer, the trillion-parameter sparse model.
OpenAI
VP of Research, Post-Training; one of the small team that shipped ChatGPT.

Brings the post-training playbook, the OpenAI bench and a conviction that physicists are AI's next talent pool.

— Ekin Dogus Cubuk · Co-founder & co-CEO · the atoms half

Harvard
Physics PhD — machine learning on glassy solids before the field had a name.
Google Brain
AutoAugment and RandAugment — augmentation methods vision models still use.
DeepMind
Led GNoME — 2.2M predicted crystals, ~380k stable: nine times the known catalogue.
Eight years
At Google and DeepMind in total. The atoms half of the founding pair.

Brings the materials-ML lineage, the DeepMind bench — and the lab's bar: powder synthesis at “coffee-robot” grade.

3.2 — What they believe that most people don't

Six contrarian takes, in the founders' words.

01 · The rewardNature as the graderMath and code leapt because graders could score answers automatically. “Nature is our RL environment”: the experiment is chemistry's grader — the only honest one.
02 · Scaling“What is this y-axis?”Models keep improving on what they were trained on. Off-distribution the slope “may not be good enough”; a coding model “is not going to then cure cancer.” So: train on lab data.
03 · Build vs buyZero effort on coding modelsCodex and Claude Code are “a huge accelerator,” not rivals. Periodic builds models only where the frontier “is not sufficiently good” — the LLM drives specialist nets as tools.
04 · The robotsThe robot is the easy partCubuk's bar for lab robotics is an airport coffee robot. Fedus: robotics is “very commoditized” — “we employ people as well.” Deciding what to make is the hard part.
05 · The billPrimarily a compute costThe GPUs cost more than the laboratory. But the lab “has very large lead times,” so it was started early; as automation improves, the bottleneck moves back to intelligence.
06 · The businessA software businessNot a royalty on discoveries. “An intelligence layer” sold to R&D teams, land-and-expand. The superconductor proves the model; the model is the product.

Sourced from three founder appearances with full transcripts — the a16z Podcast, No Priors, and a GPU Mode keynote.

3.3 — How they hire

“No background in physics” — and they mean it.

Told to candidates explicitly: even the best physicist on staff has more to learn about this problem than they already know, so a newcomer's gap is “not that different.” The screen is not the domain.

— The screen, in the founders' words

Deeply curious
Wants to understand the ML and the science, not just ship.
Pragmatic
Careful process, solution-oriented, “gets to goals quickly.”
World-class on one axis
Not across all of them.
A sense of urgency
Cubuk's addition to the list.
Cares about the mission
“Probably the biggest determinant” — plus an anti-pitch: want to improve a megacorp's product? Go there.

— How the culture absorbs newcomers

No stupid questions
“The dumbest physics question, the dumbest ML question” are welcome.
Weekly teaching
ML researchers, physicists and chemists teach each other; faculty on staff run the sessions.
Bridge people
Hired deliberately: people inside the ML–simulation–experiment triangle who translate — “CS people think in APIs.”
No degree required
No physics, no PhD; otherwise “very high overlap” with the frontier-lab bar.

The roster agrees: ads-ranking (Snap, TikTok), trading, Stripe, Waymo and Anthropic engineers sit on the AI side; six postdocs came straight from university benches.

§ 04 — Part four of five · The team

Seventy-one people, hired in the order the bets required.

The founders' beliefs, turned into an org chart — and where the org chart and the roster still disagree.

4.1 — The org chart, from the bets

Three pillars, “like a fractal” — each one a bet, staffed.

L1
LLM · post-training & RL
Fedus · Bahdanau · Piché · Rishabh A. · Huang · Nakano
Serves the bet nature as the grader. RL scientists from Brain and DeepMind, the ServiceNow pair, OpenAI alumni. Huang: “physical rewards.”
L2
LLM · infra & inference
Hsu · Bayya · Singhal · van der Staay · Sigaev
Serves the bet primarily compute. xAI's inference lead, a supercomputing engineer, Meta GPU infra; in August, a security architect.
S1
Simulation · theory
Wakefield · Manki Kim · Kaba · Hoyer · Cheon
Serves the bet where physics is robust. Theorists from MIT, Princeton and Mila; differentiable physics from Google (Hoyer).
S2
Simulation · ML for materials
Cubuk · Aykol · Riebesell · Gangan · Sheriff · Ren
Serves the bet build only what's missing. GNoME's lead, a DeepMind staff scientist, two ex-Radical AI. All did AI-for-materials elsewhere first.
E1
Experiment · chemistry & physics
Chica · Rom · Koay · Minyong Han · Safari · Pan
Serves the bet the superconductor. Solid-state chemists and thin-film physicists; Argonne superconductors; a visiting professor.
E2
Experiment · automation & facilities
Cross · Rendy · Gueble · Feng · Tatum · Ogden
Serves the bet robots are the easy part. Lila and A-Lab automation, Apple/QuantumScape controls, a fab director: the robots' keepers.

4.2 — Ten names to know

The hires that set the bar.

01
Liam Fedus
Co-Founder & Co-CEO
OpenAI VP of Research, Post-Training; ChatGPT co-creator. Google Brain. Mila PhD — root of the Canada ring.
02
Ekin Dogus Cubuk
Co-Founder & Co-CEO
Eight years at Google and DeepMind; led the materials work behind GNoME. Harvard PhD. The atoms half of the founding pair.
03
Dzmitry Bahdanau
MTS · Founding wave
Bahdanau attention — the mechanism behind modern translation; ServiceNow Research lead, remote from Montreal. His line: “Move bits to construct new things from atoms.”
04
Stephan Hoyer
Member of Technical Staff
Ten years at Google on AI weather models; creator of Xarray. A differentiable-physics req followed him.
05
Byron Hsu
MTS · RL systems & inference
xAI's Inference Lead, reporting to Musk; founding inference-team member. Liger-Kernel lead. Now: “RL system for atoms.”
06
Rishabh A.
Reinforcement Learner
Brain → DeepMind Staff RS → Meta Superintelligence Labs for four months → Periodic. McGill adjunct, in Canada.
07
Bernardus Rendy
Member of Technical Staff
Nearly five years in the Ceder Group's A-Lab at Berkeley Lab — the canonical autonomous lab. Straight from the PhD.
08
Sam Cross
MTS · Founding team
Lab-automation principal at Lila Sciences — a named rival — then founding-team MTS here. Samsung Research before.
09
Jun Feng
AI for Semiconductor Process
Ex-Director of Process Engineering, Applied Materials; Harvard PhD. A fab director running a lab's process bench.
10
Isabel Gueble
Controls Engineering
Ex-Director of Controls Engineering; Apple iPhone Display; QuantumScape. The person who makes the robots move.

MTS = Member of Technical Staff — the flat title convention imported from OpenAI, and the one most of this roster carries.

4.3 — The shape of the team

A mid-career research crew, titled flat.

8.6
Years average experience — not a company built on new graduates, nor on twenty-year veterans.
23%
Hold PhDs — 16 of 71.
86%
Senior level or above — by title.
32%
Big-tech background — Google, OpenAI, DeepMind, Meta, Apple, xAI.
80%
In the Bay Area — 57 of 71. The exceptions are Montreal, Amsterdam and London.

Most of the roster carries one title — Member of Technical Staff — the flat convention imported from OpenAI. Underneath it: 39 joined at senior level, 23 mid-career, six veterans, three early-career.

Schools underneath: UC Berkeley ×6, Stanford ×6, Columbia ×4, Harvard ×3, McGill ×3 — no dominant feeder school. The company's LinkedIn page adds MIT ×11 across all 90 associated members.

4.4 — The build order

Bits in 2025. Atoms in 2026.

Bits = models and infrastructure. Atoms = the laboratory. Read the start dates in order and the sequencing is unmistakable.

Apr – Jul 2025
Before the founders' own start dates: the ML bench. Kurtulus (Stanford SAIL), Aykol (DeepMind), Nakano and Pandey (OpenAI), Fu (Meta FAIR), Cheon (Google). One atoms hire: Menon, process engineering.
Sep 2025
The founding wave and the $300M seed. Fedus and Cubuk formalise; Bahdanau (remote, Montreal), Rishabh A., Riebesell (Radical AI), Sam Cross (Lila, lab automation), Lauterbach as Head of Operations.
Oct – Dec 2025
The first wet-lab scientists — Chica, Rom, Khorshidi-Zadeh — then Safari (Argonne, superconducting materials), facilities manager Ogden (CFS · Tesla · PsiQuantum); Sheriff converts from first intern.
Jan – Mar 2026
Infra and the second site. van der Staay (Meta), Singhal (Crusoe), Hoyer (Google), Hwang (Anthropic); Piché on-site in Montreal, Shliazhko in Amsterdam, Gangan in London. Jun Feng from Applied Materials.
Apr – Jun 2026
The xAI lift and the recruiters. Hsu and Bayya; Chau and Vu. Controls and automation (Gueble, Max P., Fei), Rendy from the A-Lab, theorists Wakefield and Manki Kim, four strategy and ops hires in June.
Jul – Aug 2026
The process-development bench. Tatum (AMAT · TSMC), Wu, Lin, Koay to MTS-Atoms, a nanofab intern — beside Sigaev (xAI), Haas (recruiter), the first Forward Deployed Engineer and a commercialization lead.

Through June, software out-hired the bench two to one; July and August are the first months to split evenly.

4.5 — Where they came from

Two networks, one PhD apart.

The first thirty hires came from where the founders had worked — OpenAI 6, Google DeepMind 5, Google 8 — recruiters included. The second ring is Montreal: Fedus's PhD city explains a founding-wave hire, an on-site Montreal researcher, a Mila resident scientist and a McGill adjunct. Five people, one city.

Google · 8OpenAI · 6Google DeepMind · 5Stanford · 5xAI · 3ServiceNow · 3McGill · 3Radical AI · 2
The frontier-lab ledger
Ekin Dogus CubukReiichiro NakanoRohan PandeyMuratahan AykolKate LauterbachJason AiDustin ChauPhuong Vu
OpenAI ×6 · Google DeepMind ×5 — recruiters included
one bridge
a Mila PhD
The Montreal ring
Dzmitry BahdanauAlexandre PichéSékou-Oumar KabaRishabh A.
Mila · ServiceNow Research · McGill

Liam Fedus is the bridge — a Mila PhD who then ran post-training at OpenAI, so both rings are his. Beyond them: Riebesell, Gangan and Cross came from Radical AI and Lila Sciences; Hsu, Bayya and Sigaev were lifted from xAI between April and July 2026. School and employer overlaps are otherwise thin. The structure is the founders' résumés, not a school.

4.6 — Who builds the lab

The wet lab is real, small — and built by fab veterans.

About 18 of 71 sit on the atoms side. The academics are here — but the laboratory itself is being built by people from semiconductor fabs, national labs and hard-tech facilities.

A1
Rendy · Cross · Fei
A-Lab · Lila · Berkeley SDL
The autonomous-lab lineage. Berkeley Lab's A-Lab, Lila's lab automation, Berkeley self-driving-labs research — the three who have built this kind of lab before.
A2
Feng · Tatum · Menon · Wu · Lin
Applied Materials · TSMC
The fab veterans. A process director, an AMAT/TSMC device technologist, two process developers — materials made repeatably, at chip-plant discipline.
A3
Gueble · Max P. · Ogden
Apple · QuantumScape · CFS
Controls and facilities. Controls from Apple and QuantumScape; automation from Chevron and Plenty; facilities from Commonwealth Fusion, Tesla and PsiQuantum.
A4
Chica · Rom · Koay · Han · Safari
+ Alizadeh · chemistry, films
The scientists. Synthesis, solid-state chemistry, thin films, superconductors (Argonne, ESRF), polymers. Safari's specialty is the north star itself.
A5
Jordan S. · Kim · Johnson · Wang
OpenAI robotics · LLNL · fab
The technicians. An OpenAI robotics technician, a Hacker Fab alum, an LLNL mechanical engineer, a nanofab intern. “Right now we employ people as well” — Fedus.

Against it, the job board still lists 19 Science roles of 28. Built, but far from finished.

4.7 — The gap between plan and team

A quarter of the roster. Two-thirds of the job board.

1in 4
Of the roster is atoms
~18 of 71 sit on the laboratory side of the house.
68%
Of open roles are Science
19 of 28 — thin films, theory, powder, nanofab, lab ops, EHS.
2: 1
Atoms hires, 2026 vs 2025
The bench is being bought now, not at founding.

— The roster · 71 people, by our read · Sep 2026

— The job board · 28 open roles

Launch-week framing — half AI, half physical scientists — went uncorrected. A year on, our read is 41 to 18. The AI scientist is largely hired. The laboratory is what's being hired now — a full lab org chart, posted.

§ 05 — Part five of five · What transfers

What the rest of us can steal.

You can't copy the résumés or the round. The order, the screens and the sourcing channels transfer.

5.1 — Sourcing patterns

Six patterns — almost none of them ran through a job board.

01 · Competitor raidsThree hires from named rivalsRiebesell and Gangan from Radical AI, Sam Cross from Lila Sciences — both rivals were on the map before the roster existed. Periodic is winning that market.
02 · Unit liftsRecruit the pair, not the personHsu and Bayya — xAI inference and supercomputing — landed within two months of each other. Bahdanau and Piché came from one ServiceNow lab in Montreal.
03 · The national-lab channelA feeder no software startup sharesLawrence Livermore ×3, Berkeley Lab, Argonne, Brookhaven. For a company building a lab, the Department of Energy's laboratories are a defensible channel.
04 · Try before you buyFive conversions to staffRen, Sheriff and Gangan began as interns; Koay and Minyong Han as in-house postdocs. All five converted. Two interns and a resident scientist are in the pipeline now.
05 · The academic comp leverSix postdocs, straight from the benchFedus calls academics “very under-compensated” and hires accordingly: six came straight from postdocs at Princeton, Stanford and Penn State.
06 · Recruiters before HRThree recruiters, no HR leaderChau (DeepMind, May), Vu (OpenAI, June), Haas (D. E. Shaw, July) — one a month — while the Founding HR Leader req stays open. Acquisition speed over people operations.

5.2 — The lessons

Five moves worth stealing.

01 · SourcingWork the founders' ledgerOpenAI ×6, DeepMind ×5, Montreal ×5: the co-CEOs' own alumni reproduced first, recruiters included. Map both founders' careers before the job board opens — the first thirty hires are in there.
02 · ScreeningDon't filter on the domainNo physics required, the founders tell candidates — and they hired ads-ranking, trading and fintech engineers as MTS. Screen for curiosity, pragmatism, urgency, one world-class axis — and, above all, the mission.
03 · SequencingBits first, atoms second — on purposeThe ML bench was mostly hired before the first wet-lab scientist. If the software half is the founders' home turf, hire it while the network is warm, then buy the bench.
04 · The poolThe lab lives in fabs and national labsApplied Materials, TSMC, Livermore, Argonne, Berkeley Lab, CFS, PsiQuantum. Autonomous-lab talent isn't on an AI recruiter's list — it's in semiconductor and hard-tech.
05 · Units of hireLift pairs, convert internsHsu + Bayya from xAI; Bahdanau + Piché from one ServiceNow lab; Riebesell + Gangan from Radical AI. Five intern and postdoc conversions. Recruit relationships; test-drive the rest.
The catchThis order costs $605M — and it isn't finished.Bits-first works when a seed round can carry a year of researchers before the lab exists. Without that cheque, invert it: hire the bench that generates the data first.

The same trust-graph sourcing we read at Eigen AI and Elorian AI — and the mirror of Unconventional AI, which bought its physical bench first.

Base to Base · Recruiting

A hiring order is a balance sheet of which risk you think is cheapest to hold — and Periodic priced the laboratory as the one worth waiting on.

Read the build order before the org chart, and you can tell which half of a company its founders actually know how to hire.

— The takeaway

They hired the AI scientist first.
The laboratory is being built
underneath it — right now.

Brief
Periodic Labs · Talent Brief
Prepared by
Base to Base · Recruiting
Written for
Founders building AI-for-science teams

Methodology & limitations

Wide coverage — honestly counted.

— Sources

  • The LinkedIn company page (90 associated members), with full profile histories for 71 of ~72 identifiable staff.
  • A 30-day public-signal scan — X, Hacker News, GitHub, YouTube, Aug 2 – Sep 1.
  • Three founder appearances with full transcripts: the a16z Podcast, No Priors, and a GPU Mode keynote.
  • Funding disclosures, press and the company's own job board.

— Limitations

  • Nine investor-advisors and one portfolio-fund profile are excluded from team statistics; one visiting scientist is retained and flagged.
  • The 41 : 18 : 12 bits/atoms/ops split is our classification of public titles, not the company's own — people who straddle the line were assigned to one side.
  • Seniority is read from public work histories, which are incomplete for part of the roster, so the buckets are noisy and the shape is not.
  • Collected September 2026. Teardowns like this are how our searches begin — this one's on us.