Company teardown · Chai Discovery
A two-year-old AI neolab's classic play for getting big pharma to pay for its models early.
Josh Meier led the ESM protein language models at Meta; Jack Dent scaled engineering teams at Stripe. Their lab now licenses antibody-design models to Lilly, Pfizer, Novartis, argenx and Bristol Myers Squibb — with a research team of about ten. We read 36 of 37 profiles to see how: a team kept “slightly over capacity” on purpose, a recruiting agency that moved in-house, and hiring waves timed to each model.
0.1 — Chai Discovery in one page
A lab that designs antibodies on a computer and sells the model, not the drug — built small on purpose, staffed in waves, sold by scientists.
Everything here is public — LinkedIn, GitHub, the Ashby job board, bioRxiv author lists, funding and deal announcements, and six founder and staff appearances with transcripts. The reading is ours.
§ 01 — Part one of six · The idea
Stop screening molecules. Start designing them.
Chai's founders treat drug design as a scaling problem — data, models, compute — and sell the model to the companies that make the drugs.
1.1 — The problem they're solving
Why finding a drug still looks like searching a haystack.
— The founder's framing
“There's not many more than 20 people on the planet, I think, capable of building something like this.”
— Joshua Meier, Core Memory, December 2025.
— De novo antibody design, translated
Writing an antibody sequence from scratch to bind a chosen spot on a target protein, instead of finding one in an animal or a library. “Zero-shot”: no known antibody for that target as a starting point. Chai-2 tests 20 or fewer designs per target, in about two weeks.
1.2 — The idea
Predict the structure, generate the binder, then let the lab feed the model.
Meier's frame is the “bitter lesson” for molecules: data, models and compute beat hand-built rules. McPartlon's test for any addition — incremental, or compounding? Chai-1 has 23 submodules and he refuses to add the 24th.
01 · Predict
02 · Generate
03 · Validate
↺ once the model is accurate enough, it generates the training data for the next one
1.3 — The wager
A research lab that bets on staying small.
Isomorphic and the big labs run research groups many times this size. Chai's bet is that a dozen first-principles researchers plus compute beats a large lab — so it holds research at ten, puts growth into platform, science and partnerships, and keeps everyone, in Meier's words, “a little bit slightly over capacity.”
— The wager, in his words
“Keep the leanest team possible, really maintain that kind of focus and figure out just what you need.”
— Joshua Meier, Unsupervised Learning, August 2025.
— And it checks out
The founders' headcount claims match the roster: “around a dozen” in July 2025 — we count 12; “about 20” in December — 20; “only 30” in August 2026 — 34. They slightly understate their own size, which is the rare direction for a startup to err in.
§ 02 — Part two of six · The business
Sell the model, not the medicine.
Chai licenses its design models to the largest pharma companies, trains bespoke versions on their data, and runs no drug pipeline of its own.
2.1 — The product, as described
Licences to big pharma, bespoke models — and open source as the shop window.
Chai runs 100–200 internal projects to stress-test the models, and buys most wet-lab work from partner labs — the lab is a benchmark, not a pipeline.
2.2 — The money
A $30M seed, then three rounds in eleven months — $630M in all.
A lead investor from each funded round holds a board seat — Thrive's Miles Grimshaw, Menlo's Greg Yap, Index's Nina Achadjian. General Catalyst, Redpoint and DCVC hold stakes without one. Investors and the one advisor are excluded from every headcount on this page.
2.3 — Who else is trying
A crowded field — and Chai hires from it.
Talent brief
Want the full visual breakdown?
Download the PDF version of this teardown — the five-contract sales motion, the six-group org chart, ten names to know, the build order month by month, the two pools that barely touch, and six sourcing patterns.
§ 03 — Part three of six · How it sells
Five contracts in eight months — and the model did the selling.
A loud launch, year-long pilots, an annual access fee, Pfizer's former R&D chief on the board — and a commercial team of seven in which nobody is titled sales.
3.1 — The motion, reconstructed from five announcements
Release, then pilot, then licence — the model does the prospecting.
↺ each release restarts the inbound: Chai-3 reached Pfizer before it was announced
Reconstructed from the Lilly, Pfizer, Novartis and TuneLab announcements, Forbes (Jun 2026), Contrary Research (Apr 2026) and Endpoints (Jan 2026). No customer has described its own buying process, and Pfizer declined to comment.
3.2 — Eighteen months, from launch to five logos
Every contract followed a model release — and the money followed the contracts.
3.3 — The five contracts, as disclosed
Five logos, one public price — and a rival inside every account.
Chai's own language: Lilly and Pfizer are “customer agreements”; Novartis, argenx and BMS are “collaborations.” Only Lilly's price has leaked, and Chai has never confirmed it.
3.4 — The contract shape
Access fees, not biobucks — the dogma Chai chose to challenge.
— The dogma
“When we started, people told us the only way to make money is to make your own assets and become a drug company. That's the dogma we had to challenge.”
— Jack Dent, Forbes, June 2026.
— And why a customer believes it
“If you want to be the trusted one that traditional industries feel comfortable teaming up with, you cannot at the same time try to have your own little shop.”
— Mikael Dolsten, board member, Forbes, June 2026.
Contrary's counter-case, the “services trap”: every bespoke model is engineering spent off the core platform. Three partner-trained models are public so far — Lilly, Pfizer and the TuneLab miniprotein suite.
3.5 — Who opened the doors · public record only
A former Pfizer R&D chief, an OpenAI seed — and investors who tell the commercial story.
3.6 — The commercial team, from 36 profiles
Seven people touch the sale — and none of them is titled “sales.”
3.7 — The funnel, in the numbers that are public
From 304 papers to five contracts — the open model is the top of the funnel.
Chai-1 is free and open; Chai-2 and Chai-3 are controlled-access. The free model buys citations — the Baker lab reports it “proved remarkably effective” — and, in Forbes' words, “lets potential pharma company customers test drive some of its tech.” The controlled models turn attention into evaluations, and evaluations, months later, into licences.
— What a launch felt like
“It was like we dropped a bomb on the field. People were messaging on LinkedIn at 2 a.m., saying, ‘I am so excited I can't sleep.’”
— Jack Dent on the Chai-2 launch, Forbes, June 2026.
— What “in talks” counts
Chai's own figure, given to Forbes in June 2026. Three of those conversations — Novartis, argenx and BMS — became announcements within eleven weeks. Twelve or more remain unconverted, or unannounced.
3.8 — What the record does not show
Five things we could not verify — and one number we rejected.
3.9 — What transfers from the sales motion
Five moves from the sales motion worth stealing.
§ 04 — Part four of six · The founders
Three AI scientists and a Stripe engineer.
Three of the four founders led AI at drug-discovery companies before Chai; the fourth scaled engineering teams at Stripe. The company is hired in exactly that shape.
4.1 — The four who built it
Heads of AI from Absci and Aqemia, ESM's lead developer, and a Stripe engineer.
4.2 — The leadership layer
A platform lead, an antibody veteran, a recruiter-founder — and no VP layer.
Thirteen of 36 carry “Member of Technical Staff”, two more “Member of Scientific Staff”. Below four co-founders there is no VP, Head or Director.
4.3 — How they hire · No Priors · Unsupervised Learning · Core Memory · Sachin & Adam · Training Data · Latent Space
Hire for the next model, keep everyone over capacity, and want researchers who engineer.
§ 05 — Part five of six · The team
Thirty-six of 37, hired in three waves.
AI researchers first, then lab scientists and product engineers, then platform, security and partnerships — each wave timed to a model release, and each drawn from a different pool.
5.1 — The org chart, from the bets · inferred from profiles
Six groups — and the second-largest writes software, not papers.
Groups are our reading of 36 profiles, not an org chart Chai publishes. Research is still the bigger group, at eleven counting the two founders who do research — which is the same bench McPartlon calls “around ten.” The ten platform engineers, none with biology on the CV, are the bigger surprise.
5.2 — Ten names to know
The hires that set the bar.
5.3 — The shape of the team · 36 profiles
Senior, small, and in the room.
— Career stage at join · all 36 profiles classified
A Pfizer director, a Stripe staff engineer and a maths graduate from Uppsala can carry the same title: Member of Technical Staff. Above it, four co-founders; between, nothing. The three remote exceptions — Boston, Portland, Tel Aviv — are two senior scientists and a contract recruiter.
Schools, by LinkedIn's census over all 45 members: Berkeley 9, Stanford 7, Harvard 5, Cambridge 4, MIT 3. That census spans the investors and advisor too, so it is quoted as found rather than recomputed over the 36.
5.4 — The build order · first Chai role, 36 profiles
Researchers in 2024. Scientists and recruiters in 2025. Platform in 2026.
Sixteen of 36 joined in 2026 — and only four of them are ML researchers. Six are platform engineers.
5.5 — Where they came from
Two pools — and four founders as the bridge.
Fourteen of 36 came from AI-drug-discovery companies. Eleven came from software companies with no biology at all — Stripe, Nuro, Imbue, Vanta, Watershed, Sieve. Three of the four founders seeded a chain from a former employer, and the chains are still running.
These count prior affiliations — employers, schools and programmes — across the 36 profiles we read. Highlighted chips are the chains — a founder's last employer (Absci, Aqemia, Stripe), a rival's bench (Seismic) or the embedded agency (Plus Plus). The other non-zero chips are repeat sources without a person-to-person chain — a shared school, a scouting programme, or two hires from the same rival months apart. The Isomorphic zero is real — no one has moved between Chai and the company every write-up compares it to, in either direction.
founders bridge them
Everyone else — 11 of 36 — comes from academia, CERN, talent, partnerships and ops. Excluding big tech, 14 of 36 share a prior employer with someone who joined before them ; counting Google and Facebook overlaps, 18.
5.6 — The agency that moved in
The recruiting agency became the talent team.
Plus Plus, a nine-year-old technical-recruiting agency, listed Chai among its clients. In May 2025 two of its talent partners started at Chai on contract; in September its founder joined full-time. Four of 36 people now recruit — and a fifth “Talent” role is open.
5.7 — Physicists first, biologists later
“Only me and Kevin have a bio background” — already out of date.
McPartlon said it of the research team on August 11, 2026. It was true of the founding bench: physicists and mathematicians who learned proteins on the job. But since September 2025 the company has hired five people with biology or chemistry doctorates — the science bench is arriving while the public story is still “first-principles ML people.”
5.8 — The census and the job board
Platform absorbed a year of hiring — and the board wants researchers back.
— The company · our reading of 36 profiles, Sep 2026
— The job board · 17 open roles, all San Francisco · Ashby
LinkedIn's own census over 45 members reads Engineering 20, Business Development 13, Research 3 — self-reported functions that file ML researchers under engineering and investors under BD. The board tells the truer story: after a year of platform hiring, research is the biggest slice again, plus a first finance hire and an “Outliers & Polymaths” pipeline.
§ 06 — Part six of six · What transfers
What the rest of us can steal.
You can't copy ESM's lead developer or a $3.8B valuation. Hiring by model generation, embedding the agency and lifting a rival's ML bench transfer.
6.1 — Sourcing patterns
Six patterns — most of them are chains, not job posts.
6.2 — The lessons
Five moves worth stealing.
— The takeaway
Three AI scientists and a Stripe engineer
kept the lab at thirty-seven —
and hired each wave for the model that came next.
Methodology & limitations
A near-complete read — honestly counted.
— Sources
- The LinkedIn company page — 45 associated members, of whom 7 are investors and 1 an advisor, excluded by scope — with profile histories for 36 of 37 staff, plus its function, school and location census.
- The GitHub org chaidiscovery, the Ashby careers page (17 roles, fetched Sep 27), and bioRxiv author lists for Chai-1, Chai-2 and the November 2025 report.
- Six founder and staff appearances with transcripts — No Priors, Unsupervised Learning, Core Memory, Sachin & Adam, Sequoia Training Data and Latent Space.
- OpenAlex citations for 13 members, a 30-day public-signal scan (Aug 29 – Sep 28), and funding and deal announcements. Part three adds Business Wire, Forbes, TechCrunch, Endpoints, PharmExec and Contrary Research, fetched October 3.
— Limitations
- The citation total is a floor: only primary OpenAlex author IDs were counted, and five researchers with common names could not be told apart from namesakes, so they are blank rather than guessed. Read ~16.5K as “at least.”
- Seventeen of the 36 function labels came from an LLM reading the title, because “Member of Technical Staff” tells a rules-based classifier nothing. Every group on 5.1 is our reading of profiles, not an org chart Chai publishes.
- The school census spans all 45 members, investors and advisor included, so it is quoted as published rather than recomputed over the 36 we read.
- No comparative baseline. Chai is the first Series C company in our AI-biomolecular-design cohort, and a comparison needs five or more.
- Headcount is a moving target: the founders said “only 30 people” in August 2026 and we count 34 by that month. Collected September 28, 2026. Teardowns like this are how our searches begin — this one's on us.