Not just another AI biotech
Most AI drug discovery companies promise speed. Chai Discovery is claiming something bolder: that the earliest stage of antibody discovery can become programmable. Programmable the way software is. You say what you want a molecule to do, and the system hands you candidates.
It's a big pitch. Traditional antibody discovery means screening huge libraries, running directed evolution experiments, or leaning on what's already known about antibody structure. Every one of those is a search. You comb through possibilities and hope something works.
Chai wants to design instead. Give the model a protein target and an epitope (the patch on the target you want to hit) and it returns antibody sequences that should bind there.
Does this work at the scale pharma cares about? That's the real question. The early data is striking, the partner list keeps growing, and the company reached a $1.3 billion valuation less than two years after founding. That's fast for a company working on one of the hardest problems in biology.
Why antibodies matter
Antibodies are Y-shaped proteins made by the immune system. They latch onto specific targets with high precision, which is exactly what you want in a drug: find the disease protein, stick to it, trigger a response. Compared to small molecules, antibodies have one big advantage. Their binding surface is much larger. More surface to engineer means more specificity and fewer off-target effects. That matters, since non-specific binding causes about 28% of small molecule drug failures.
Then there's pharmacokinetics. Antibodies stay in the blood for 10 to 21 days. Small molecules usually clear in hours. So patients dose less often, and in many disease areas they do better. It's part of why monoclonal antibodies now make up about 25% of new drug approvals. The global market was $288 billion in 2024 and is on track for $628 billion by 2035.
So Chai isn't chasing a niche. Antibodies sit at the center of modern pharma, and the pipeline keeps widening as nanobodies, bispecifics, and antibody-drug conjugates mature. Get meaningfully better at designing them and you earn a seat inside one of the most valuable processes in the life sciences.
The bottleneck: discovery is still slow
Chai isn't trying to fix all of drug development. It's going after the front end of biologics discovery: how fast a team can produce plausible, testable molecules against a target.
Right now, that front end is slow. Rational design starts from what scientists already know about antibody structure and reasons toward something that might bind. Directed evolution takes an existing antibody, mutates it at random, and screens the variants. High-throughput screening means building or buying big libraries and running assays until binders turn up.
All three produce candidates. None of them is quick. Going from target to viable hits usually takes 12 to 24 months. A single screening campaign can cost $500K to $2 million and tie up five to ten people for months per target.
Chai says it can get that down to four to eight weeks. That's a 10 to 20x cut, if it holds across different targets and therapeutic areas.
One caveat matters a lot here. A hit is not a drug. An antibody that binds in a model or an early assay still has to be validated, tested for manufacturability and immunogenicity, run through preclinical work, and eventually put into people. Chai makes the first step much faster. The rest still takes years.
Why now: three curves crossed
Chai exists because three curves crossed at once. None of them alone would have been enough.
First, AI training got cheap. As of early 2026, the cost per floating point operation is down roughly 74% since 2019, falling about 10x a year. Training large generative models on protein and antibody sequences is now practical in a way it wasn't five years ago.
Second, DNA synthesis got cheap too. It cost $25 per base in 1999 and around $0.3 in 2025. That sounds abstract. In practice it means you can synthesize and test computer-designed sequences in bulk, instead of one painful trial at a time. Generate a hundred candidates on a computer, then test them for real.
The third curve, and probably the biggest, is AlphaFold. AlphaFold2 came out in July 2021 and showed that atomic-level protein structure prediction could actually be solved computationally. That raised a follow-up question. If structure is predictable, is it also designable? The research since then says yes, in a meaningful range of cases. Chai-1 and Chai-2 sit on that foundation, plus the generative modeling techniques that grew up alongside language models.
"AlphaFold made structure prediction feel possible. Chai wants to make molecular design feel interactive."
Founding story
Chai was founded in March 2024 by Joshua Meier, Jack Dent, Matt McPartlon, and Jacques Boutreau. It's an unusual mix of machine learning pedigree and structural biology depth.
Meier's path stands out. He grew up in a science family, was a Google Science Fair finalist at 16, and started a company in high school that engineered mosquitoes to deliver West Nile vaccines. He studied computer science at Harvard and trained in Feng Zhang's CRISPR lab. OpenAI recruited him straight out of school, which gave him an early look at what large-scale language model training could do.
Next came Meta's generative biology group, where he helped build ESM1, the first transformer protein-language model. Then three years at Absci as Chief AI Officer, working on de novo antibody design. By the time he co-founded Chai, he'd picked up pretty much every skill the company's technical agenda called for.
Dent studied computer science with Meier at Harvard, then went to Stripe and helped build Stripe Link and Stripe Capital. He brings product and operations sense to an otherwise research-heavy group. McPartlon worked with Meier on Absci's AI research team, specifically on de novo antibody design models. Boutreau came from RNA bioinformatics at McGill and small molecule design at Aqemia. Four people, covering generative modeling, structural biology, product, and chemistry.
The origin story is fun. Sam Altman had tried to recruit Meier for an OpenAI proteomics spinout. Meier said no. He didn't think the science was ready. When he and Dent went back to Altman in 2024 to revisit the idea, the conversation turned into Chai Discovery, and OpenAI became one of the first seed investors. Mikael Dolsten, Pfizer's former Chief Scientific Officer, later joined the board as part of the Series A.
What really sets Chai apart from earlier AI biotechs is that it chose not to build a wet lab. Meier thought the last generation of AI biotech was tied too tightly to lab infrastructure. Chai wanted to be a portable AI platform first: a design tool that plugs into whatever wet lab a pharma company already runs. That choice has real consequences for the business model. More on that later.
Product philosophy: AutoCAD for molecules
Chai calls its product a "computer-aided design suite" for molecules. The reference is AutoCAD, which changed mechanical engineering by letting people specify, simulate, and iterate on designs before building anything. Before AutoCAD, engineers worked from hand-drawn schematics and physical prototypes. After it, design became iterative and computational. Chai thinks molecular biology can make the same jump.
The analogy has holes. Molecular biology has far more variables and a vastly bigger search space. And you can't skip the wet lab the way a mechanical engineer can sometimes skip a physical prototype. Still, the direction of the bet is right. If you can generate and score candidates on a computer before synthesizing anything, you can run far more experiments per dollar and per day than any traditional lab.
The product lineup follows this logic. Chai-1 predicts molecular structure. Chai-2 generates new molecules. Chai-3 aims for the reliability pharma needs in production. Prediction, then design, then deployment inside real drug programs.
Chai-1: the credibility layer
Chai-1 came out in September 2024 as an open-source model under Apache-2.0. It's a multimodal foundation model for structure prediction that handles proteins, small molecules, DNA, RNA, and covalent modifications. That breadth matters. Most structure tools back then were protein-centric, while drug discovery increasingly needs to model interactions across all of these at once.
On 152 antibody-antigen complexes held out of training, Chai-1 reached DockQ scores above 0.8 on 17% of structures. (A score above 0.8 means the predicted binding interface closely matches the real experimental structure.) That's roughly double AlphaFold2 on the same benchmark. Not bad for a brand-new company.
By 2025, 304 published papers had cited Chai-1. The David Baker lab at the University of Washington, arguably the top computational protein design group in the world, said Chai-1 "proved remarkably effective at identifying the most active designs" for enzyme engineering. Pharma pays attention to outside validation like that.
The open-source move is on purpose. Chai doesn't make money on Chai-1. Chai-1 is how it earns the right to charge for Chai-2 and Chai-3. An open model makes it easy for biopharma buyers to run their own evaluations, and strong public benchmark numbers make those evaluations more likely to go well. It's a smart wedge, taken straight from the developer-tools playbook.
Chai-2: the commercial breakthrough
With Chai-2, the story shifts from "we can predict molecular structures" to "we can propose new therapeutic candidates." It launched in August 2025 alongside the Series A. It was the first real test of whether Chai's generative approach works on problems pharma actually cares about.
There are two parts. Chai-2d is the design submodel. You give it a protein target, mark the epitope you want the antibody to bind, and it writes antibody sequences from scratch. No existing library. No known antibody to optimize. Truly de novo. You can pick the format too: full-length IgG, single-domain VHH, or bispecific.
Chai-2f is the folding submodel. It takes Chai-2d's sequences and predicts the structure of the antibody-antigen complex, so you can check on a computer whether a design is likely to work before you synthesize it. On the same benchmark used for Chai-1, Chai-2f scored DockQ above 0.8 on 34% of structures. Twice as good.
The headline number is a roughly 20% experimental hit rate across 52 different targets. Earlier computational antibody design methods sat around 0.1%. That's a 200x jump.
It's worth repeating, though: these are early binders, not development candidates. A molecule still has to clear developability filters, pass preclinical safety, and work in humans. That's a long road. Chai shortens the start of it, which is genuinely valuable. The road is still there.
Chai-3 and Pfizer: the 2026 update
Chai's 2025 story was about model performance. Its 2026 story is about getting deployed inside big pharma.
In June 2026, Chai announced a Pfizer partnership covering platform access, early access to Chai-3, and a custom model trained on Pfizer's proprietary data and built around Pfizer's discovery workflows. That last piece is the big one. Pfizer isn't just licensing a tool. It's wiring Chai's models into its own discovery process and feeding them proprietary data.
According to the company, Chai-3 doubles Chai-2's success rate. It does better on therapeutic binding, multi-specific antibodies, hard-to-drug targets, and generalizing across target classes. That's a real jump on a hit rate that was already strong. Pfizer moving this early suggests its internal tests looked good.
Step back and there's a clear pattern. Open-source credibility with Chai-1. Early commercial use with Chai-2. Deep pharma workflow integration with Chai-3. Each step pulls Chai further into how pharma companies actually run discovery.
Customers and traction
Traction comes in four buckets, and they feed each other.
Academic adoption came first. Researchers started plugging Chai-1 into their workflows almost as soon as it came out in September 2024. By April 2026 it had hundreds of academic citations, and top structural biology labs, the Baker Lab among them, were getting useful results with it on real problems. You can't fake that kind of adoption. It makes pharma procurement conversations a lot easier.
Pharma deals followed. Eli Lilly announced its collaboration in January 2026, paying a mid-eight-figure annual fee to use Chai's antibody design tools across its portfolio. Lilly validated the models internally before signing. At a company that size, that's no rubber stamp. Pfizer came in June 2026 with Chai-3 access and a custom model. Discovery teams at other drugmakers have confirmed they use Chai's tools too, including a J&J scientist who runs Chai alongside AlphaFold for kinase inhibitor screening.
Government work adds institutional weight. The UK picked Chai as a strategic industry collaborator for its OpenBind initiative, alongside Isomorphic Labs and Genentech. The goal is to build the world's largest open dataset of drug-protein interactions for training generative models.
Then there's demand. After Chai-2's announcement, hundreds of pharma and biotech companies asked for early access. That's not revenue. It is a big pool of potential customers.
Business model
Chai runs a hybrid model. Chai-1 is open under Apache-2.0 and free for anyone. That's not charity. It's distribution. Open releases build the scientific credibility that gets pharma buyers to look at the paid products.
Chai-2 and Chai-3 are sold through annual platform access fees. The Lilly deal is the clearest public data point: a mid-eight-figure annual fee (somewhere between $10 million and $99 million a year) for platform access and a custom model trained on Lilly's internal data. AI drug discovery partnerships also usually carry royalties, typically 0.5% to 5% of net sales on molecules found with the platform. Say a Chai-assisted program produces a blockbuster antibody with $1 billion in annual sales. The royalty could be worth $5 million to $50 million a year.
There's a real tension here. Broad platform licensing scales like software. Build the model once, sell it many times, watch margins improve. But the Lilly and Pfizer deals both meant training exclusive models on proprietary data and fitting tools to specific workflows. That's services work, and services don't scale the same way. If every big pharma deal needs its own engineering team, Chai starts to look less like a software platform and more like a high-end contract research organization.
"The bull case is software margins. The bear case is a very expensive AI-enabled contract research organization."
Market size
Chai isn't selling antibodies. It's trying to own the design layer underneath the antibody market.
Monoclonal antibody therapeutics were a roughly $288 billion global market in 2024, headed toward $628 billion by 2035. That's the ceiling if Chai ever develops its own drugs. For now, though, it sells tools that help pharma find antibodies faster. A better comparison is the high-throughput screening industry, worth over $20 billion, which Chai could partly replace at the hit-generation stage.
Add AI platform licensing budgets, custom model work, and future royalties, and the opportunity is large. A lot of it depends on one thing: Chai becoming the default tool inside discovery teams.
The market can also grow sideways. Chai-2 hit 68% success on mini-protein designs (50 to 100 amino acids). Meier has said there's no fundamental reason the same approach can't reach high rates on peptides, RNA, and other modalities. If the platform generalizes, antibodies are just the start.
The competition
The competitive map has three layers, and Chai relates to each one differently.
| Company | Category | Funding / Status | How they compete with Chai |
|---|---|---|---|
| Isomorphic Labs | Generative / structure prediction | $600M raised; ~$3B val (secondary) | DeepMind pedigree, AlphaFold base; lead optimization focus; J&J, Lilly partnerships |
| Xaira Therapeutics | Generative design (RFdiffusion) | $1B raised; ~$2.7B val | David Baker co-founder; exclusive license on RFantibody from UW |
| Boltz | Open-source structure prediction | $28M seed; PBC structure | Fully open-source; Pfizer partnership; Boltz-2 approaches physics-based FEP accuracy |
| Nabla Bio | Antibody design (JAM platform) | $37M raised | 39% VHH hit rates; membrane protein specialization; Takeda partnership |
| Absci | Antibody design + wet lab | Public; $560M market cap | Zero-prior epitope strategy; internal wet lab for proprietary data generation; AMD investment |
| Big Hat Biosciences | Antibody design + wet lab | $99M raised | Milliner platform integrates ML with high-speed wet lab; Eli Lilly partnership |
| EvolutionaryScale | Protein language models | Acquired by Chan Zuckerberg Initiative | ESM3 and novel protein engineering (fluorescent protein); research focus |
Here's the key point: almost everyone starts from the same data. The core public training sets for antibody structure prediction are the Protein Data Bank (about 200,000 protein structures) and the Structural Antibody Database (about 10,000 antibody structures). Every competitor uses them. A former director at Absci put it bluntly: "There's no moat you get from the training set. The ability to generalize from the training set or generate new sequences only now comes from your ML strategy and only comes from the compute you use."
That's a big admission. If the starting data is the same, the edge has to come from somewhere else: model architecture, compute, proprietary pharma data, and how fast you can loop between computational predictions and lab results. The race isn't only about who has the best model today. It's about who builds the best feedback loop across model predictions, wet-lab validation, pharma data, and real programs.
Chai's plan is to win the pharma data flywheel. The Lilly deal trains on Lilly's proprietary data. So does the Pfizer deal. If those exclusive models clearly beat what's possible on public data, Chai has a real advantage that's hard to copy. If they don't, the moat is thinner than it looks.
Valuation and fundraising
Chai raised a $70 million Series A in August 2025 at a $550 million valuation. Four months later came a $130 million Series B at $1.3 billion. OpenAI was an early seed investor. Menlo Ventures led the A, and Oak HC/FT and General Catalyst co-led the B. Individual backers include Greg Brockman, Sarah Guo, and Fred Ehrsam, and Pfizer's former CSO Mikael Dolsten sits on the board.
Investors aren't pricing Chai like a normal early-stage biotech with one or two assets. They're pricing it like a platform that could become infrastructure for biologics discovery. That's a revenue-multiple story, not a pipeline story. To justify it, Chai needs recurring platform revenue that grows with each pharma partnership, plus milestones and royalties from drugs found on the platform.
Here's the catch. Chai hasn't disclosed revenue. The Lilly fee is a real number, but it can't carry a $1.3 billion valuation on its own. The bull case needs Pfizer. Then the next ten partners after Pfizer. Then a royalty stream that only kicks in as Chai-assisted drugs work through clinical development, years from now. That's a long timeline to bet on.
| Revenue mechanism | What Chai gets | What pharma gets | Scalability |
|---|---|---|---|
| Platform access fee | Annual license (mid-8-fig range per partner) | Access to Chai-2/3; tooling | High: same model, many partners |
| Custom model deal | Higher fee + proprietary data access | Exclusive model trained on their data | Medium: requires engineering per partner |
| Discovery royalty | 0.5–5% of net sales from Chai-assisted drugs | Keeps most of drug economics | High once drugs reach market (10–15 yr lag) |
| Co-owned asset | Shared economics on drug | Partner contributes clinical expertise | Low: requires Chai to take on risk + capital |
The bull case
The bull case rests on five legs.
First, the market is huge and it matters. Antibodies aren't a niche. They're central to pharma R&D and headed toward $628 billion globally by 2035. Make antibody discovery meaningfully faster and you're sitting in a very good spot.
Second, the products keep getting better. Chai-1 built credibility. Chai-2 showed generative antibody design at a hit rate no earlier computational method came close to. Pfizer is already adopting Chai-3. Three product generations in two years.
Third, pharma partnerships create data flywheels. Lilly and Pfizer both involve custom models trained on proprietary data. If those models beat public-data baselines, Chai gets harder to replace, and every new data partnership widens the gap. Pfizer's former CSO on the board isn't a coincidence either. It puts Chai right at the top of pharma's networks.
Fourth, Chai can grow past antibodies. Meier describes the underlying tech as modality-agnostic, and Chai-2 already posts high success rates on mini-proteins. RNA therapeutics, peptides, and other biologic drug classes are in scope. Each one is a big market.
Fifth, the platform could change how pharma manages its portfolio. If a discovery team can cheaply test 100 therapeutic ideas instead of five, it can chase riskier, more differentiated targets. Meier put it this way: "Sometimes people are like, how did you decide to try this idea versus this other idea? And we're often just trying tons of them in the lab because it's just easy enough to do it right now." That shifts the logic of R&D across a whole portfolio, not just inside single programs.
The bear case
The core risk: Chai might be very good at making binders, while the market pays for drugs.
First, competitors are catching up. The field moves fast. Isomorphic Labs released IsoDDE in February 2026, and it beat Chai-1 on protein-ligand structure prediction by 3x. Boltz-2 matched the accuracy of physics-based free-energy perturbation methods. The Baker Lab's RFantibody is open-source and trained on the same data as everyone else. New model claims show up every six months. Technical leads are real, but they don't last. Here's that Absci director again: "Every six months there will be a new model claiming it has superiority in silico than the existing models. Any technical edge a company in this space has gets rapidly competed away."
Second, there's a proprietary data gap. An antibody that binds in a model isn't the same as a drug that's safe, manufacturable, and effective. Data on immunogenicity, manufacturability, and formulation mostly lives inside pharma companies and barely shows up in public repositories. There are fewer than 1,000 public data points on antibody immunogenicity. A solid model needs more than 10,000.
Without that data, generative models can produce binders that fail on developability. Chai's answer is to get proprietary data through pharma deals. That's the right idea, but it only works if partners actually hand over data that's useful for training. Nothing guarantees that.
Third, platforms struggle to capture value. This is baked into the business model, not unique to Chai. When a computational platform finds a drug candidate, the pharma company that develops it keeps most of the money. A 0.5% to 5% royalty on a blockbuster sounds good. Next to what the company running the trials and the launch takes home, it's a thin slice.
Look at Schrödinger. It's a computational drug discovery platform with a strong track record, and its market cap is around $870 million as of 2026, roughly a quarter of the value created by assets discovered on its platform. Chai is worth $1.3 billion with far less revenue history. Something has to give.
Fourth, the services trap. Lilly and Pfizer both involve custom models built for specific workflows. Twice is manageable. Twenty times means separate engineering teams maintaining separate codebases while the core platform also has to keep improving. That complexity is exactly what separates a scalable software business from a contract research organization that happens to use AI.
Fifth, nothing has been proven in the clinic. No Chai-designed antibody has made it through clinical development. A computational binder that validates in early assays still has to clear developability screens, preclinical safety, Phase I dose escalation, Phase II efficacy, and Phase III. Hit to approval takes 10 to 15 years. Right now the market is pricing in the platform's potential, not clinical results. If early Chai-designed programs run into toxicity or efficacy problems, the story changes fast.
| Bull case | Bear case |
|---|---|
| 20% hit rate on 52 targets is 200x better than prior computational methods | Hit rates measure binding, not developability, safety, or clinical efficacy |
| Lilly and Pfizer partnerships validate pharma willingness to pay at scale | Custom model deals per partner are services, not software; hard to scale |
| Pharma data flywheels: more partnerships = better models = harder to displace | All competitors start from the same public databases; model claims change every 6 months |
| Platform is modality-agnostic: RNA, peptides, mini-proteins all in scope | No Chai-designed antibody has cleared clinical development yet |
| $1.3B valuation buys time to prove the model before needing to pivot to assets | Historical AI drug discovery platforms captured only a small fraction of drug value (see Schrödinger) |
My take
Chai is interesting because it isn't trying to be a classic biotech, at least not yet. It wants to be the design layer for biologics. That puts it closer to infrastructure than to a therapeutics company. Very different to run. Very different to value.
The most exciting version of Chai goes beyond faster antibody discovery. Picture a closed loop where generative models, structure prediction, pharma data, and wet-lab validation keep improving each other. Each Lilly or Pfizer deal brings in revenue, and also training data for the next generation of models. That makes the next deal easier to close and the models harder to copy with public data. If that flywheel actually spins, it justifies a billion-dollar valuation on a two-year-old company.
The clinical gap is the part I have trouble with. Biology has a habit of punishing clean software analogies. A 20% computational hit rate is impressive given where the field started. The hard questions come later. Will Chai-designed antibodies pass immunogenicity filters? Can they be made at scale? Will any of them work in people? We won't know for years, and the market is already pricing in a yes.
I'd also keep a close eye on the services trap. The Lilly and Pfizer deals validate Chai, but they could box it in too. "Mid-eight-figure annual fee plus custom model" is a great first deal. As a template for the fifth one, it worries me. Every custom engagement that needs Chai engineers to maintain adds headcount cost that never shows up in the licensing number. Eventually the company has to pick. Is it building a platform, or running a bespoke AI research shop for big pharma?
What makes me more optimistic about Chai than most of its AI drug discovery peers is how clear the founders are about what they aren't. They skipped the wet lab on purpose, and that was a real bet. Most earlier AI biotechs built labs because they thought they needed control over their data. Chai decided to get that control through partnerships.
Whether that works depends on pharma sharing the right training data, not just using the platform. That's the variable I'd watch in every new deal announcement.
Chai Discovery is one of the clearest tests of whether AI drug discovery can be a product business before it becomes an asset business. If Chai becomes the default tool scientists reach for when designing antibodies, it could be one of the defining AI biotech companies of the decade. If the value keeps flowing mostly to pharma's own assets, or competitors catch up on model performance, Chai may have to move deeper into developing its own drugs, like plenty of platform biotechs before it.
The next two to three years will start to show which way it goes. Watch the clinical data from Chai-assisted programs. Watch the next handful of pharma deals.