Chai Discovery

Phase A — Understand the business

Lens 1 · Company Overview

Chai Discovery is an AI-native molecular-design company — building foundation models that turn drug discovery from wet-lab trial-and-error into a software-design problem. Founded March 2024, HQ San Francisco, ~31 employees as of Apr 2026. The framing the founders use: a "computer-aided design suite for molecules" — the AutoCAD-for-biology analogy.

What they actually sell — two products, two motions:

• Chai-1 (Sep 2024) — a multimodal structure-and-interaction prediction model (proteins, small molecules, DNA, RNA, covalent mods). Open-source, Apache-2.0, free for academic and commercial use. This is the credibility/adoption flywheel, not the revenue engine. An "atomic-level microscope".
• Chai-2 (Jun 2025) → Chai-3 (Jun 2026) — the de novo antibody design engine. This is the product they charge for. Zero-shot: a user specifies an epitope + modality (full-length IgG, VHH/nanobody, bispecific) and the model generates novel binder sequences from scratch — "Photoshop for antibody design," generating rather than optimizing.

Business model — two-pronged:

1. Platform licensing — annual access fees to pharma/biotech for Chai-2/3. The Eli Lilly deal is a "mid-eight-figure annual access fee".
2. Research partnerships — custom bespoke models trained on a partner's proprietary data, typically with a 0.5%–5% royalty on product net sales for the discovery contribution.

Customers / partners (the proof points):

• Eli Lilly (Jan 2026) — deploy Chai's platform at Lilly + train an exclusive bespoke model on Lilly's proprietary data for multi-target biologics. Lilly internally validated the models before signing.
• Pfizer (Jun 2026) — license agreement: early access to Chai-3 plus a custom Pfizer-data model.
• UK OpenBind (Jun 2025) — alongside Isomorphic Labs + Genentech, generating the largest open drug–protein interaction dataset.
• Confirmed bottoms-up adoption: J&J scientist running Chai alongside AlphaFold for kinase work; David Baker lab (UW) found Chai-1 "remarkably effective" for enzyme design.

Why this is interesting (the MenFem teaching frame): the conventional view is "AI drug discovery has overpromised for a decade — Recursion, Exscientia, BenevolentAI all underdelivered." The contrarian read is that those were target-ID and small-molecule screening plays; Chai is doing something narrower and more falsifiable — designing a protein binder that you can synthesize in a 24-well plate and measure binding on within two weeks. The feedback loop is short, the result is physical, and the hit rate is going up release-over-release. That is a different category of claim.

Lens 2 · Supply Chain

For an AI-bio platform the "supply chain" is compute → data → model → wet-lab validation → pharma deployment. Named stakeholders along the chain:

• Upstream — compute: GPU capacity. Not separately disclosed, but the OpenAI relationship (seed investor; offices) and the AI-frontier pedigree imply NVIDIA-class training infrastructure, likely via a hyperscaler. n/a — specific cloud/GPU vendor not disclosed.
• Upstream — training data (the real input): public structural biology. ~200K Protein Data Bank (PDB) structures + ~10K SAbDab antibody structures are the shared substrate every competitor trains on. This is the chain's most important fact: the raw material is a commons. Chai's edge cannot come from data it owns — it must come from ML architecture + compute + the design loop.
• Midstream — Chai (the model layer): Chai-1 (open) and Chai-2/3 (commercial).
• The chokepoint — wet-lab validation: Chai deliberately runs lab-light — it forgoes significant internal wet-lab capacity and relies on partners (and CRO-style validation) to confirm designs. The industry-wide constraint has flipped: "experimental throughput is now the constraint, not design capacity". Chai produces millions of in-silico candidates; somebody still has to express, purify and assay them (SPR, ELISA, cell assays).
• Downstream — proprietary data Chai does NOT have: immunogenicity, manufacturability, formulation data. Public immunogenicity data is <1K points; proper training needs >10K. This data lives inside Lilly/Pfizer — which is exactly why the partnership model exists: Chai trades platform access for access to the proprietary developability data it can't otherwise get.
• Downstream — end customer: pharma R&D pipelines (Lilly, Pfizer) → eventually patients, 10–15 years out.

Single-source dependency: the PDB/SAbDab commons (everyone shares it) and — critically — pharma partners as the sole source of the developability data that turns a binder into a drug. That dependency is the supply-chain risk and the moat question in one.

Lens 3 · Competitive Advantages (moats)

Be honest: in a field where everyone trains on the same public structures and models leapfrog every ~6 months, the data moat is weak by construction. So where could a durable edge actually live?

1. The design-loop + cadence (the real candidate moat). Chai-1 → Chai-2 → Chai-3 in under two years, each roughly doubling the prior model's success rate (Chai-2 DockQ>0.8 on 34% of the 152-complex benchmark vs Chai-1's 17%; Chai-3 doubles Chai-2 again). A moat made of velocity — staying 6–12 months ahead — is real but rentable, not owned.
2. Talent density. Joshua Meier co-led ESM1 (the first transformer protein language model) at Meta and ran AI at Absci; Matt McPartlon did de novo antibody design at Absci/VantAI; Jack Dent built ML infra at Stripe. For a velocity moat, the team is the moat.
3. Proprietary developability data via partnerships (the moat they're trying to build). Each Lilly/Pfizer bespoke-model deal feeds Chai private immunogenicity/manufacturability data the open competitors can't see. If this compounds, it converts a weak data position into a defensible one. This is the single most important strategic bet in the company.
4. Open-source distribution as a top-of-funnel moat. 1.9K+ GitHub stars, 304 papers cited Chai-1 in 2025, 400+ total citations by Apr 2026. Chai-1 being the default academic structure tool seeds the commercial funnel — but Boltz (also open) contests this directly.

Bargaining power: today Chai needs pharma more than pharma needs Chai — Lilly/Pfizer have alternatives (Isomorphic, internal teams, AlphaFold3) and own the proprietary data. The leverage only flips if Chai-designed molecules start reaching the clinic and working, which is a decade away. Net: the moat is "stay ahead + lock in proprietary data via partners." Both are contestable. This is a great team renting an edge, working to convert it into an owned one.

Lens 4 · Segments

n/a — private; no segment-level financial disclosure. No segments.csv data (header only). Qualitatively, two revenue lines exist:

• Platform access fees (recurring, annual) — Lilly mid-eight-figures/yr is the only sized data point; Pfizer undisclosed.
• Partnership / royalty (custom models + 0.5%–5% net-sales royalties) — no products on market, so royalty revenue is $0 today and is a 10-yr option, not income.

Geographic split: US-centric (SF), with a UK research footprint via OpenBind. n/a — not disclosed.

---

Phase B — Measure performance

Lens 5 → Funding & valuation trajectory (+private swap)

The "earnings" of a private is its round history. Chai's trajectory is steep and tier-1-validated:

• Valuation step-ups: $150M → $550M (~3.7x in ~12mo) → $1.3B (~2.4x in ~4mo). The Series B closed the same quarter the Lilly deal was being finalized — revenue proof drove the markup, not just hype.
• Syndicate quality (an IPO-proximity tell): Thrive, Menlo, General Catalyst, Oak HC/FT (a healthcare-crossover specialist), plus Glade Brook and Emerson Collective in the B. GC + Oak HC/FT co-leading is exactly the late-stage healthcare-crossover signature that precedes an S-1 by ~18–36 months.
• Angel/strategic roster reads like an AI-establishment cap table: Greg Brockman, Sarah Guo (Conviction), Fred Ehrsam, Lachy Groom, Will Gaybrick, the Hartzes, Martin Chavez.
• Burn signal: 31 people on $231M raised = the cash is going into compute + talent, not headcount. Lab-light keeps burn lower than wet-lab-heavy peers (Xaira, Generate). Runway: not disclosed; estimate multi-year given a fresh $130M B and a thin team.

Lens 5b → Pipeline / model cadence (+clinical swap — the asset IS the model)

For Chai the "pipeline" is the model line and its wet-lab-validated capability, not clinical assets (it has none in the clinic). The capability table:

The de-risking question (the one that matters): is the hit rate improving and generalizing? Evidence says yes — 17% → 34% structural accuracy, ~0.1% → 16% → (Chai-3) ~30%+ design hit rate, and developability now >85%. That trajectory, not any single number, is the bull thesis.

Lens 6 → Founder / public signal (+private swap for "earnings calls")

No earnings calls. The sentiment proxy is founder communication + scientific publishing cadence:

• Tone: confident, benchmark-forward, deliberately publishing wet-lab failure modes alongside wins (the Nov 2025 paper discloses expression-failure rates) — a credibility-building move rare in hype-cycle AI-bio.
• Strategic framing shift: from "open structure model" (Chai-1, science credibility) → "commercial antibody designer with drug-like outputs" (Chai-2/3, revenue). Meier's Series B line — "molecules that have properties we'd want from actual drugs, and tackle challenging targets that have been out of reach" — signals the pivot from can it bind to is it a drug.
• Recurring phrases: "zero-shot," "de novo," "drug-like properties," "hard-to-drug targets," "engineering discipline." Stopped saying: nothing structural yet — the company is <2.5 yrs old.

Lens 7 → Cap table, comps & secondary marks (+private + mechanism-comps swap)

Comps are by category, not P/E (nobody here has earnings). USD valuations are secondary/round marks, ``:

Reading the table: Chai is mid-pack on valuation ($1.3B) but arguably top-tier on the specific axis it competes on (wet-lab-validated zero-shot antibody hit rate with signed pharma revenue). Isomorphic is bigger, better-capitalized, and claims a 3x structural-prediction edge over Chai-1 — but Isomorphic's public proof is mostly structure/affinity prediction, whereas Chai's is generative design validated in a plate. Different proof, overlapping ambition. Nabla's 39% VHH hit rate is a genuine flag that Chai's lead is contestable.

Crossover-fund marks: GC + Oak HC/FT (B) and Menlo (A) are the IPO-proximity tells. No public mutual-fund markup yet. Secondary marks: n/a — not disclosed.

Lens 8 → Funding/product catalysts (the events that re-rated the company)

Mapping the "stock-moving" events for a private (each step-changed perceived value):

• Sep 2024 — Chai-1 open release → instant scientific credibility (the "we're real" moment).
• Aug 2025 — $70M Series A / $550M, Menlo lead, Mikael Dolsten (ex-Pfizer CSO) to board → institutional + pharma validation.
• Jun/Jul 2025 — Chai-2 + the "24-well plate" biorxiv paper → the 16% hit rate that made the field take notice; "hundreds of early-access requests."
• Nov 2025 — Chai-2 full-length IgG + GPCR + >85% developability data → de-risked "binder ≠ drug" critique.
• Dec 2025 — $130M Series B / $1.3B → 2.4x markup.
• Jan 2026 — Eli Lilly mid-eight-figure deal → first marquee revenue proof.
• Jun 2026 — Chai-3 + Pfizer license → second pharma + a model that doubles the prior.

Pattern: the market re-rates Chai on (1) a published wet-lab hit-rate step-change and (2) a named pharma signing. Those are the two catalyst types to watch. The next re-rate is either Chai-3 wet-lab numbers in print, a third major pharma, or the first Chai-designed molecule entering IND-enabling studies.

---

Phase C — Judge people & books

Lens 9 · Management

Joshua Meier — CEO/co-founder. The signal hire of the field. Google Science Fair finalist at 16; founded a biotech (Provita) in high school; Harvard + Feng Zhang's CRISPR lab; OpenAI (2018, research/eng); co-led ESM1 at Meta (first transformer protein LM); Chief AI Officer at Absci where he worked on de novo antibody design tied to clinical molecules. Founder-archetype, technical, with a prior in exactly this problem — not a generalist parachuting into bio. Skin in the game: founder equity, undisclosed %.

Jack Dent — President/co-founder. Harvard CS classmate of Meier; cold-emailed into Stripe as a teenager, built Stripe Link + Stripe Capital + large-scale ML infra. The systems/commercial half — explains the unusually clean enterprise motion (Lilly/Pfizer signed fast for a 2-yr-old).

Matt McPartlon — CTO. De novo antibody design at Absci; protein-protein interaction models at VantAI/Proxima. Jacques Boutreau — generative molecular affinity (McGill/Aqemia).

Board: Mikael Dolsten, Pfizer's former Chief Scientific Officer, joined at Series A — and Pfizer is now a customer. That is either the best possible validation or a related-party flag worth noting (an ex-Pfizer-CSO board member while Pfizer becomes a paying licensee). I read it as validation, but flag it.

Capital allocation: disciplined — 31 people on $231M, lab-light, spending on compute + talent not real estate or a wet lab. No buybacks/M&A (too early). Red flags: none material; the only watch-items are (a) the Dolsten/Pfizer adjacency and (b) the founders' youth + first-time-CEO status against a giant TAM and brutal competition. Archetype: mission-driven technical founders with a genuine domain prior and a commercial co-founder. For this stage, close to ideal.

Lens 10 · Forensic Red Flags & Regulatory

Accounting: n/a — private, unaudited, no financial statements. No revenue-recognition, lease, SBC, or goodwill analysis is possible without filings; none invented.

Regulatory findings (from regulatory/regulatory-findings.md, generated 2026-06-24):

• SEC (EDGAR EFTS — LR + AAER): 0 findings. Chai has no CIK — private, not required to file. No SEC enforcement is searchable.
• Non-SEC web search — "Chai Discovery" (FTC OR DOJ OR FDA OR CFPB OR consent decree OR settlement OR fine OR penalty) enforcement: no material enforcement, litigation, fine, or consent-decree hits surfaced across this and the Phase A–B searches.
• Item 3 (Legal Proceedings): n/a — no 10-K exists.

The real "forensic" risks for a company like this are scientific, not accounting — and the skeptic literature flags them precisely (carried into Lens 13):

• Survivor bias in the developability stats — 8% of IgGs and 37% of VHHs failed expression and were excluded before the >85%-developability figure was computed. The headline number is conditioned on the molecules that survived.
• Framework selection bias — designs deliberately used VH3-23/VH3-66 frameworks chosen for thermostability, flattering developability metrics.
• Hit-rate variance — 4% (CXCR6) to 100% (CCR5); some pMHC targets at 0%. The "16%" is a mean over a wildly bimodal distribution.

Verdict on the books: clean on the (nonexistent) accounting; the diligence burden is entirely on whether the science generalizes beyond favorable targets and survives the developability funnel. "No material regulatory or legal findings — verified via SEC EDGAR EFTS (LR, AAER), web search, and (n/a — no 10-K) as of 2026-06-24."

---

Phase D — Project & stress-test

Lens 11 → IPO-readiness & path-to-tradeable (+private swap)

Where it is: post-Series B, ~$1.3B, two flagship pharma deals (Lilly + Pfizer) producing real ARR, a model cadence that's still accelerating. Stage: late-private, commercially-validated, pre-IPO.

IPO-readiness: MEDIUM, and rising — but an S-1 is ~2–3 years out. What gates it:

1. ARR scale + durability — one mid-eight-figure deal (Lilly) + Pfizer is a start, not an IPO-grade book. Public markets will want a recurring, diversified, growing licensing line (call it $100M+ ARR across 5+ pharma) and ideally early royalty optionality.
2. A clinical validation event — the field's overhang is that no AI-designed molecule has completed clinical development. Chai doesn't need a Phase 3, but a Chai-designed antibody entering IND-enabling / Phase 1 at a partner would be the single biggest de-risking catalyst, and that's plausibly 2–4 years out.
3. Surviving the next competitive cycle — staying ahead of Isomorphic/Xaira/Nabla through Chai-3 → Chai-4.

Estimated tradeable window: 2028–2029. Be-early action: this is exactly the name to track now — there is no public way to own it, and the watch is when (pre-IPO secondary, or the eventual IPO), not if it's interesting.

Write-back note: research/private-watch.json does not exist / has no chai-discovery entry. Per wave boundaries I do not edit watchlist/private-watch state from this run — flag for the curator to add chai-discovery with stage: late-private, ipo_readiness: medium, catalyst: "ARR scale + first Chai-designed molecule into IND-enabling studies", dossier: companies/chai-discovery/deep-dive-2026-06-24.md.

Forecast (Brier): skipping forecast.ts create (no EPS for a private; per --watchlist/private rules, log forecasts only on a binary). Candidate binary for a future log: "Chai Discovery signs a 3rd top-15-pharma licensing deal by 2026-12-31" — p≈0.55; not logged this run.

Lens 12 · Bull vs Bear

Bull case. Chai is the first AI-antibody platform to clear the bar that killed a decade of AI-bio hype: a wet-lab-validated, generalizing, release-over-release-improving design hit rate (~0.1% → 16% → ~30%+), and a real enterprise business (Lilly mid-eight-figures + Pfizer) at <2.5 years old. The team is the best-pedigreed in the field (ESM1 + Stripe + Absci). The TAM is enormous — mAbs were $288B in 2024 → ~$628B by 2035, 23% of 2025 drug approvals were antibodies. If the partnership model converts proprietary developability data into a compounding edge, and if even one Chai-designed molecule reaches the clinic, this re-rates from "$1.3B AI tool" to "the antibody-design layer of the pharma industry." The open Chai-1 funnel + the licensing engine is a clean software business with biotech-scale upside.

Bear case (3 things that permanently impair it):

1. The platform-vs-asset value trap. History is brutal to platforms: Schrödinger captured 3.7% of the Nimbus TYK2 value ($147M of a $4B exit); Recursion and Generate both pivoted to owning assets because platforms capture crumbs. At 0.5–5% royalties, Chai is the pick-and-shovel that watches pharma keep the gold. A $1.3B valuation on a royalty/licensing model needs enormous volume to justify.
2. Commoditization + a better-funded leader. Everyone trains on the same PDB/SAbDab commons; models leapfrog every ~6 months. Isomorphic ($3B, IsoDDE claims 3x Chai-1) and Xaira ($2.7B, Nobel-laureate + own wet lab) are deeper-pocketed. A velocity moat is rented; the moment Chai stops shipping the best model, the thesis erodes.
3. The services trap. Bespoke Lilly/Pfizer models divert engineering from the core platform and risk turning Chai into a glorified CRO with multiple codebases to maintain — non-scalable, low-multiple.

Pre-mortem (it's Dec 2027 and the thesis broke): Chai-4 lands ~level with Chai-3 (the hit-rate curve flattened — favorable targets were the easy 80%); Isomorphic's IsoDDE-derived design tool overtakes it; Lilly renews at a lower fee because internal teams + AlphaFold closed the gap; the Pfizer deal lapses; and zero Chai molecules reached the clinic, so the platform never earned a royalty. The $1.3B looks like a 2025 AI-bubble mark.

Is the multiple too high? On traction (one sized deal + one undisclosed), $1.3B is rich — it's priced on the trajectory and the team, not the book. Defensible only if you believe the hit-rate curve keeps bending and the licensing line compounds.

Contrarian view (what the market is refusing to see): the bears anchor on "platforms capture no value" — but that precedent is from the small-molecule + target-ID era, where the platform's contribution was fuzzy and hard to attribute. Zero-shot de novo design is different: the binder didn't exist until the model drew it. Attribution is clean, which makes a defensible, higher royalty (and renewable annual licenses) more plausible than the Schrödinger analogy implies. If Chai is the company that proves "AI designed this drug from scratch" with a clinical asset, the value-capture math the bears use breaks in Chai's favor.

Lens 13 · Devil's Advocate (short-seller)

Short thesis: Chai is a brilliant research group with a fragile business priced like a winner.

• Where revenue concentrates → the kill shot. Effectively two customers (Lilly, Pfizer), one of them sized. Lose or fail-to-renew Lilly and the revenue story — and the $1.3B mark — collapses. Worse: a Pfizer ex-CSO sits on Chai's board while Pfizer becomes a licensee — great optics until you ask how arm's-length that deal is, and whether it's repeatable with pharmas that don't have a friendly board member.
• The hit rate is a cherry-picked mean. Independent teardown of the Chai-2 paper: GPCR hits ranged 4% → 100%; pMHC targets KRAS-G12V 4%, p53-R175H 0%, gp100 0%. The GPCR agonist wins required near-perfect active-state structures with a bound agonist — only a tiny fraction of the ~800 PDB GPCR structures qualify, so the capability doesn't scale to the targets pharma actually wants. And the agonists were 7x–200x weaker than natural ligands (CCR8 272nM vs 40nM; CXCR4 164nM vs 0.83nM). "Designs a functional GPCR agonist" is doing heavy lifting.
• Survivorship inflates developability. 37% of VHHs failed expression and were dropped before the ">85% developable" stat. The funnel is leakier than the headline.
• The moat is rented. Same public data as everyone; Isomorphic claims 3x Chai-1 on structure, Nabla shows a 39% VHH hit rate — Chai's lead on its own axis is already contested. Compute + talent are buyable; EvolutionaryScale's team just got absorbed by CZI, proving these groups are fungible.
• Binder ≠ drug, and the clock is 10–15 years. No AI-designed molecule has ever completed clinical development. Immunogenicity/manufacturability data — the stuff that actually kills antibodies — isn't in the public data Chai trains on (<1K immunogenicity points vs >10K needed). Chai is designing for the easy half of the problem and renting the hard half from partners.
• Valuation breaks on a 20–30% growth miss. A $1.3B mark on a thin, concentrated, royalty-capped book is a 2025-vintage AI-bio bubble price. If the next model is merely good (not 2x), if a third pharma doesn't sign in 2026, or if Lilly renews flat/down, the down-round math is ugly.

The one scenario that permanently impairs it: the hit-rate curve flattens at the favorable-target ceiling and Isomorphic/Xaira ship a generative design tool at parity — Chai loses its only real moat (being ahead) with no asset ownership to fall back on. Plausibility: moderate. It's the bet that has to be wrong for the bull case to win.

Lens 14 · Management Questions (15, ordered by information value)

1. Of the 16% mean Chai-2 hit rate, what is the median and distribution by target class for the targets your paying partners actually prioritize (vs the published 52) — and how does Chai-3 change that distribution?
2. What is total ARR today and the renewal/expansion rate on existing licenses — and how concentrated is it across Lilly + Pfizer?
3. For GPCRs and other hard targets, how do you design without a near-perfect active-state structure — does the capability scale beyond the tiny fraction of PDB structures that qualify?
4. How do you convert partner-derived proprietary developability data into a defensible, owned moat rather than data the partner restricts to their own programs?
5. Where do you draw the line on the platform-vs-asset question — will Chai ever own a wholly-owned therapeutic asset, and if not, how do you escape the Schrödinger 3.7% value-capture trap?
6. What royalty rate are you actually getting (0.5–5% is wide), and why is it defensible given de novo attribution is cleaner than legacy platforms?
7. How do you avoid the services trap as bespoke pharma models multiply — how many partner codebases can you maintain before platform velocity suffers?
8. Isomorphic claims IsoDDE is 3x Chai-1 on protein-ligand; Nabla shows 39% VHH hit rates — where, specifically, is Chai's durable edge and how long is your lead in months?
9. What is your honest expression/developability funnel end-to-end (design → express → purify → developable → manufacturable), un-survivor-biased?
10. When does the first Chai-designed molecule enter IND-enabling / clinical studies, at which partner, and what's the gate?
11. What's the burn and runway on the $130M B, and what milestones must hit before the next raise / an S-1?
12. How exposed are you to immunogenicity surprises in vivo given the public data gap (<1K points), and what are you doing to close it?
13. How real is the GPCR-agonist function story given 7–200x weaker potency than natural ligands — is this a marketing capability or a drug capability?
14. What governance exists around the Dolsten/Pfizer relationship to keep partner deals arm's-length and repeatable with non-affiliated pharma?
15. What does the org look like at 100 people — does headcount discipline (31 today) hold, and where do the next hires go (research vs commercial vs wet lab)?

---