AI-Bio
PrivateNot investable — absorbed into CZ Biohub (non-profit) Nov-2025 as an acqui-hire; the science is frontier-grade but the equity is gone. The IPO path is dead; track the people (Rives) and the ESM model line, not the cap table.
Research
The verdict
"Not investable — absorbed into CZ Biohub (non-profit) Nov-2025 as an acqui-hire; the science is frontier-grade but the equity is gone. The IPO path is dead; track the people (Rives) and the ESM model line, not the cap table."
EvolutionaryScale was a frontier AI research lab for biology — a "foundation-model lab for proteins," explicitly modelling itself on the OpenAI/Anthropic playbook but pointed at the language of life rather than human language. Founded 2023 (incorporated July 2023), dual-HQ New York City + San Francisco, structured as a public-benefit corporation (PBC).
What it actually made (the product): the ESM (Evolutionary Scale Modeling) family of protein language models — most notably ESM3, a 98-billion-parameter multimodal generative model that reasons jointly over a protein's sequence, structure, and function, and can generate novel proteins to a prompt (not just predict structure). Later shipped a lighter ESM C ("ESM Cambrian") representation-model line.
Business model (intended): a model-as-a-service / API platform — the Forge API gated the large 7B/98B ESM3 variants (with safety filtering), while small open variants (1.4B / ESM C 300M–600M) were released under non-commercial / open licenses to seed academic adoption. Distribution was outsourced to the hyperscalers: AWS (Bedrock, SageMaker, HealthOmics) and NVIDIA BioNeMo NIM microservices. The pitch: sell to pharma R&D; EvolutionaryScale claimed reach to "nine of the top ten global pharmaceutical companies" via the AWS partnership — note this is platform availability, not signed enterprise revenue.
Customers / suppliers / competitors (one line each):
n/a — private, not disclosed.The structural problem (visible already in Lens 1): the business model — charge for inference on a protein model whose open-source rivals were closing the gap monthly — never demonstrably cleared escape velocity. That tension is the whole bear case, and it is why the company landed inside a non-profit 17 months after launch rather than raising a Series A.
For a model lab the "supply chain" is compute → training data → model → distribution rail → end user. Named nodes:
Chokepoint / single-source dependency: the company was structurally dependent on two strategic investors (NVIDIA + Amazon) for both compute and go-to-market. That is efficient for a lab but means EvolutionaryScale never owned its distribution — and a foundation-model lab that owns neither its data (open) nor its distribution (hyperscaler marketplaces) has a thin defensible core. Names present → lens satisfied.
Where the moat was real:
Where the moat was thin (and this is the crux):
Bargaining power: weak over suppliers (NVIDIA/AWS held the compute and the rails); weak over customers (pharma could wait for the free open model or use AlphaFold3). The only durable bargaining chip was the people — and that is exactly the chip CZI bought. Net: the moat was the team, not the product.
n/a — private, pre-revenue, no segment disclosure. There is no segments.csv data and no public revenue breakdown by product or geography. The closest to a "segment" view is the model line (qualitative, ``):
No revenue, EBITDA, or earnings exists to break out. Do not fabricate a split.
With no P&L and no clinical pipeline, the analogue to an earnings print is the model release cadence and capability milestones. As a "pipeline" table (all ``):
| Release | What it is | "Phase" / status | Significance |
|---|---|---|---|
| ESM1 (2019, at Meta) | First transformer protein LM | Foundational (legacy) | Created the category |
| ESM2 / ESMFold (2022, at Meta) | Structure prediction at scale | Mature, widely adopted | Atlas of ~600M predicted structures |
| ESM3 (Jun 2024) | 98B multimodal generative (seq+struct+function) | Shipped; public beta API | The flagship; Science 2025; esmGFP |
| ESM C / Cambrian (2024–25) | Efficient representation models | Shipped (open + gated) | Adoption funnel + embeddings |
The "print" that matters: the esmGFP result — a de novo bright fluorescent protein at 58% identity to known FPs, peer-reviewed in Science (Jan 2025). This is the single hardest evidence that the platform worked as generative biology, not just prediction. Balance-sheet/cash flags: none disclosed; on $142M seed with a ~50-person team running 98B-param training runs on H100s, burn was high and the absence of any Series A in 17 months is the tell — the runway-to-next-raise math did not close (see Lens 11). "Market reaction": the ultimate market verdict was the Nov-2025 acqui-hire at undisclosed (presumably modest) terms — i.e. the private market declined to fund an independent Series A at a step-up, and a non-profit absorbed the team instead.
No earnings calls exist. Substituting the founder/narrative track (``):
Sentiment trajectory: hype (2024) → legitimacy (early 2025) → soft-landing rationalisation (late 2025). The thing they stopped saying was anything about API revenue, enterprise pharma contracts, or a Series A. That silence is the signal.
Comps are by capability (protein/biomolecular foundation models), not by multiple — there is no public multiple for any of these to anchor on. Every "valuation" cell is n/a to avoid fabrication.
| Company | Model / asset | Funding (last known) | Status | Note |
|---|---|---|---|---|
| EvolutionaryScale | ESM3 / ESM C | $142M seed (Jun 2024) | Acquired by CZ Biohub Nov-2025 | Subject |
| Isomorphic Labs (Alphabet/DeepMind) | AlphaFold3 | $2.1B raised (2025) | Private, drug-developer | The 800-lb gorilla; develops drugs itself |
| Chai Discovery | Chai-1 / Chai-2 | ~$30M+ seed (OpenAI-backed) | Private | Multimodal structure prediction |
| Profluent | ProGen / OpenCRISPR | undisclosed | Private | Claims protein-design "scaling laws"; open-sources |
| Generate Biomedicines | generative protein platform | ~$1B+ cumulative | Private, drug pipeline | Owns therapeutic assets |
| Cradle Bio | protein-design SaaS | ~$73M+ | Private | Direct API/SaaS competitor |
| Baker Lab / IPD (academic) | RoseTTAFold All-Atom | grant-funded | Academic/open | Free; Nobel-laureate-led (Baker) |
| Boltz-2 (MIT + Recursion) | open structure+affinity | open-source | Free | Closing the gap, open weights |
Valuation, all rows: n/a (private; no comparable multiple exists). The comp that matters: Isomorphic (AlphaFold3) raised $2.1B and owns its drug pipeline — it monetises the model by becoming a pharma, not by selling API calls. EvolutionaryScale chose the harder, thinner business (sell the tool) and could not raise against it. That strategic contrast is the core lesson of the whole comp set.
No stock; the analogue is value-inflection events (``):
Pattern: the value-creating events were all scientific/distribution milestones; no commercial-traction milestone ever landed (no announced enterprise pharma revenue, no Series A). For a tool-company that is the missing catalyst — and its absence is exactly what the Nov-2025 outcome priced.
The team was the asset. This is a pure scientific-founder company, and an unusually credentialed one:
Capital allocation: raised $142M, spent it on frontier compute + an elite ~50-person team, shipped a Science-grade model in ~12 months. That is technically excellent allocation. Commercially, the allocation produced no demonstrated revenue engine — and when the next-round math didn't close, the team chose a non-profit soft landing over a down-round or shutdown. For the scientists that is arguably the best available outcome (permanent funding, no commercial treadmill); for equity holders it is, at best, a return-of-capital-ish event (see Lens 11).
Red flags (governance): none of the fraud/related-party variety surfaced. The relevant "flag" is structural, not ethical: a world-class research team paired with an under-proven commercial model, exiting fast into philanthropy. Founder archetype implication: these are researchers who want to do research — the CZI/Biohub landing fits their revealed preference perfectly, which is why it happened and why betting on them as commercial operators was always the weaker read.
Forensic accounting: n/a — private, no audited financials, no filings. There is no income statement, balance sheet, or cash-flow statement to forensically examine. No 10-K, no segment reporting, no SBC disclosure. Nothing here is auditable; treat all financial framing as unaudited per public sources.
Regulatory findings (per regulatory/regulatory-findings.md, generated 2026-06-29):
n/a — no filings exist.Summary: No corporate/accounting red flags found (and none auditable). The material risk is regulatory/biosecurity around the technology itself — and notably, the CZI/Biohub non-profit home reduces the commercial-misuse-incentive surface while keeping the capability concentrated. Verified via SEC EDGAR EFTS (LR, AAER — 0 hits), web enforcement search (0 hits), and the absence of any filings, as of 2026-06-29.
This is the be-early payoff lens, and the answer is unusually clean: the path-to-tradeable is CLOSED.
Biohub/esm migration) the model line moved into Chan Zuckerberg Biohub, a non-profit philanthropic research entity. There is no S-1 path, no secondary market, no tradeable security../:n/a — not disclosed. The strategic investors (NVIDIA, Amazon) plausibly valued the ecosystem/relationship over the markup.Write-back: I will add an evolutionaryscale entry to research/private-watch.json marking stage: acquired, ipo_readiness: 0, catalyst = "absorbed into CZ Biohub Nov-2025; no tradeable path", with this dossier path — so privates.ts shows the name correctly de-risked off the IPO board.
No Brier forecast logged (per --watchlist rules, and there is no scoreable EPS/readout — the binary "will it IPO" already resolved NO).
Bull case (for the technology/people, since the equity is gone): EvolutionaryScale built the best-pedigreed protein-LM team on earth and a genuine Science-grade generative model (esmGFP). Inside CZI/Biohub — with Mark Zuckerberg & Priscilla Chan's multi-billion-dollar philanthropic commitment and freedom from commercial pressure — the team can pursue cell-modelling and "cure/manage/prevent all disease" ambitions with more compute and a longer horizon than any VC would fund. If you could somehow own exposure, the option value on the science is real.
Bear case (why this was never a great investment): (1) No moat — open-source models (Boltz-2, Chai, RoseTTAFold) commoditised the core capability within ~18 months; (2) No revenue engine — selling API inference to pharma is a thin, contested business vs. Isomorphic's "become-a-pharma" model; (3) No pricing power — distribution owned by AWS/NVIDIA, data freely available. Pre-mortem (it's 18 months out and the thesis broke): it already broke — the company couldn't raise a Series A at a step-up and was absorbed into a non-profit, exactly the failure mode a skeptic would have flagged at the seed: brilliant science, unproven business, commoditising fast.
Multiples too high? n/a — no multiple; but the implied seed valuation (raising $142M in 2024) was priced on AI-bio euphoria that the subsequent commercial reality did not support.
Contrarian view (what the market refused to see): The market in 2024 treated "frontier protein model" as a durable moat like a frontier LLM. It wasn't — protein models are smaller, train on open data, and get open-sourced fast, so the lab has far less defensibility than an OpenAI. The contrarian-but-correct read was that the value would accrue to whoever owns the downstream drug asset (Isomorphic, Generate) or the open ecosystem — not to the pure tool-maker. EvolutionaryScale's trajectory confirms it.
If this were a tradeable security, here is the short.
(Now addressed to Rives/leadership inside CZI/Biohub, since that is where they sit.)
facebookresearch/esm) going forward?Not a tools company anymore — a sub-NAV cash shell mid-conversion into Treeline's oncology pipeline; the only edge is the deal-spread between ~$325M market cap and the ~$460M net cash being delivered, and that spread is a bet on Bilenker's KRAS/BCL6 readouts, not on CyTOF.
A re-rated single-asset AI-antibody story — ABS-101's half-life miss quietly killed the old lead, so the entire ~$1.15B cap now rides on one binary (ABS-201 alopecia interim PoC, H2 2026) against a 26%-of-float short and an 18-month runway. Own the readout, not the platform.
A fortress-margin vertical-SaaS monopoly trading at a growth-stock funeral price (~20x forward EPS, near 52-wk lows) because the market is pricing a Salesforce-Agentforce CRM war that threatens the contested ~40% (Commercial) while ignoring the defensible, faster-growing ~60% (R&D/Quality); BULLISH at $153 on a 1–3Y view, but the CRM-migration-to-2030 is a real, watchable execution overhang — not a phantom.