Not just another AI biotech
Most AI drug discovery companies promise to make the process faster. Chai Discovery is making a more aggressive claim: that the earliest stage of antibody discovery can become programmable. Not just computer-assisted, not just faster than a wet lab. Programmable, in the way software is programmable. You specify what you want a molecule to do, and the system generates candidates.
That is the pitch. And it is a big one. Traditional antibody discovery starts with screening large libraries, running directed evolution experiments, or relying on prior knowledge of antibody structure. All of these approaches share the same basic character: you search through a space of possibilities, hoping to find something that works. Chai wants to flip that. Instead of searching, you design. You give the model a protein target and an epitope, the region on the target you want to hit, and it returns antibody sequences that should bind there.
Whether this actually works at the scale pharma cares about is the central question. Chai's early data is striking. Their partnership list is growing. And the company went from founding to a $1.3 billion valuation in less than two years. That is a fast trajectory for a company working on one of the hardest problems in biology.
Why antibodies matter
Antibodies are Y-shaped proteins produced by the immune system. They bind to specific targets with high precision, which is exactly the property you want in a drug: find the disease-associated protein, stick to it, and trigger a therapeutic response. What makes antibodies particularly attractive compared to small molecule drugs is the size of their binding surface. There is more area to engineer, which means more specificity and fewer off-target effects. Non-specific binding is responsible for about 28% of small molecule drug failures. Antibodies, with their larger and more tunable binding interfaces, can sidestep a lot of that.
There is also the pharmacokinetics story. Antibodies have a serum half-life of 10 to 21 days, while small molecules are typically cleared in hours. That translates to less frequent dosing and, in many disease areas, better patient outcomes. It is one of the reasons monoclonal antibodies now account for about 25% of new drug approvals and represent a $288 billion global market as of 2024, growing toward $628 billion by 2035.
Chai is not attacking a niche. Antibody therapeutics are at the center of modern pharma, and the pipeline keeps growing as new modalities like nanobodies, bispecifics, and antibody-drug conjugates mature. Whoever gets meaningfully better at designing antibodies faster gets to insert themselves into one of the most valuable processes in the life sciences.
The bottleneck: discovery is still slow
The problem Chai is solving is not the entire drug development process. It is solving the front end of biologics discovery: how quickly a team can generate plausible, testable molecules against a target.
Traditional antibody discovery relies on a handful of approaches, and none of them are fast. Rational design means starting from what scientists already know about antibody structure and reasoning toward something that might bind. Directed evolution means taking an existing antibody, introducing random mutations, and screening variants for the properties you want. High-throughput screening means building or buying large libraries of antibody variants and running assays to find binders. All of these generate candidates, but the cycle from target identification to viable antibody hits typically runs 12 to 24 months. A single high-throughput screening campaign can cost $500K to $2 million and requires teams of five to ten people working for months per target.
Chai says it can compress that front-end discovery cycle to four to eight weeks. That is a 10 to 20x reduction in time, if the claim holds up across diverse targets and therapeutic contexts. The important caveat is that a hit is not a drug. An antibody that binds in a computational model or initial assay still needs to be experimentally validated, developability-tested for manufacturability and immunogenicity, pushed through preclinical work, and eventually into humans. Chai is making the first step of that process dramatically faster. The rest of the pipeline still takes years.
Why now: three curves crossed
Chai exists because three curves crossed at the same time. On their own, none of them would have been enough.
The first is AI training costs. As of early 2026, the cost per floating point operation has fallen roughly 74% since 2019, declining at approximately 10x per year. That makes it practical to train large generative models on protein and antibody sequence data in ways that were economically impossible five years ago. The second is DNA synthesis costs. Synthesis has fallen from $25 per base in 1999 to around $0.3 per base in 2025. That sounds abstract until you think about what it enables: computationally designed antibody sequences can now be synthesized and experimentally validated at costs that support high-throughput iteration rather than one-at-a-time trial and error. You can generate a hundred candidates computationally and test them physically.
The third, and probably the most important, is AlphaFold. When AlphaFold2 was released in July 2021, it demonstrated for the first time that atomic-level protein structure prediction was a solvable computational problem. That opened something deeper: if structure is predictable, maybe it is also designable. The research that followed showed this was true in a meaningful range of cases. Chai-1 and Chai-2 are built on that foundation, combined with the generative modeling techniques that matured in parallel from language model research.
"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. The founding team is unusual in how it assembles machine learning pedigree with structural biology depth.
Meier's path is the most interesting one. He came from a science family, was a Google Science Fair finalist at 16, and started a company in high school focused on genetically engineered mosquitoes delivering West Nile vaccines. He studied computer science at Harvard, trained in Feng Zhang's CRISPR lab, and was recruited directly to OpenAI after graduating, where he got exposure to what large-scale language model training could do. He then moved to Meta's generative biology group, where he helped develop ESM1, the first transformer protein-language model. From there he went to Absci as Chief AI Officer, where he spent three years working on de novo antibody design. By the time he co-founded Chai, he had basically threaded every needle that the company's technical agenda required.
Dent had studied computer science alongside Meier at Harvard, then went to Stripe where he helped build Stripe Link and Stripe Capital. He brings product and operational sensibility to a team that is otherwise very research-heavy. McPartlon worked with Meier at Absci on the AI research team, specifically on de novo antibody design modeling. Boutreau came from RNA bioinformatics at McGill and small molecule design at Aqemia. The four of them cover generative modeling, structural biology, product, and chemistry in a team of four.
The startup origin is interesting too. Sam Altman had been trying to recruit Meier for an OpenAI proteomics spinout. Meier passed because he felt the underlying science was not ready. When he and Dent reconnected with Altman in 2024 to revisit the idea, the conversation turned into Chai Discovery, with OpenAI becoming one of the first seed investors. Mikael Dolsten, Pfizer's former Chief Scientific Officer, later joined the board as part of the Series A.
The thing that makes Chai look different from earlier AI biotech companies is the deliberate choice not to build a wet lab. Meier's view was that the previous generation of AI biotech was too tightly integrated with lab infrastructure. Chai wanted to be a portable AI platform first, a design tool that could sit on top of whatever wet lab a pharma company already had. This is a bet with real implications for the business model, which I will get to.
Product philosophy: AutoCAD for molecules
Chai describes its core product philosophy as a "computer-aided design suite" for molecules. The reference point is AutoCAD, which transformed mechanical engineering by making it possible to specify, simulate, and iterate on physical designs before building anything. Before tools like AutoCAD, engineers worked from hand-drawn schematics and physical prototypes. After them, design became iterative and computational. Chai thinks the same transition is possible in molecular biology.
The analogy is not perfect. Molecular biology has far more variables, the search space is astronomically larger, and wet-lab validation is unavoidable in a way that physical prototyping is not always unavoidable for mechanical engineering. But the direction of the bet is right: if you can computationally generate and evaluate molecular candidates before synthesizing anything, you can run far more experiments per dollar and per day than a traditional lab allows.
Chai's product reflects this. Chai-1 is about understanding molecular structure. Chai-2 is about generating new molecules. Chai-3 pushes further into production-grade pharma reliability. The trajectory is from prediction to design to deployment inside real drug programs.
Chai-1: the credibility layer
Chai-1 was released in September 2024 as an open-source model under an Apache-2.0 license. It is a multimodal foundation model for molecular structure prediction, handling proteins, small molecules, DNA, RNA, and covalent modifications. The breadth matters: most structure prediction tools at the time were protein-centric, and drug discovery increasingly requires modeling interactions across all of these molecule types simultaneously.
When benchmarked on 152 antibody-antigen complexes held out of training, Chai-1 achieved DockQ scores above 0.8 on 17% of structures. A DockQ score above 0.8 means the predicted binding interface closely matches the actual experimental structure. This was roughly twice AlphaFold2's performance on the same benchmark, which is a meaningful result from a freshly founded company.
By 2025, Chai-1 had been cited in 304 published papers. The David Baker lab at the University of Washington, which is arguably the most prestigious group in computational protein design, reported that Chai-1 "proved remarkably effective at identifying the most active designs" for enzyme engineering. That is the kind of third-party validation that accelerates pharma trust.
The open-source strategy is deliberate. Chai-1 is not where Chai makes money. It is where Chai earns the right to charge for Chai-2 and Chai-3. Scientific credibility through an open model lowers the barrier for biopharma buyers to run internal evaluations, and a good track record on public benchmarks makes those evaluations more likely to go well. This is a smart wedge, borrowed from the playbook of developer-focused software companies.
Chai-2: the commercial breakthrough
Chai-2 is where the company's story changes from "we can predict molecular structures" to "we can propose new therapeutic candidates." It was announced in August 2025 alongside the Series A and represents the first real test of whether Chai's generative approach works on problems pharma actually cares about.
The architecture has two parts. Chai-2d is the design submodel. You give it a protein target and specify an epitope, the region on the target you want the antibody to bind, and it generates antibody sequences from scratch. Not from an existing library. Not by optimizing a known antibody. Truly de novo. You can also specify the format: full-length IgG, single-domain VHH, bispecific. Chai-2f is the folding submodel. It takes the sequences Chai-2d generates and predicts the structure of the resulting antibody-antigen complex, so you can evaluate computationally whether the designs are likely to work before synthesizing anything. On the same benchmark as Chai-1, Chai-2f achieved DockQ scores above 0.8 on 34% of structures, twice the performance of Chai-1.
The headline result is a roughly 20% experimental hit rate across 52 diverse targets. Prior computational antibody design methods had success rates around 0.1%. That is a 200x improvement. The caveat worth repeating is that these are early binders, not development candidates. Getting from a 20% hit rate at the bench to a molecule that clears developability filters, passes preclinical safety, and works in humans is a long road. Chai is shortening the front end of that road, which has real value. But the road still exists.
Chai-3 and Pfizer: the 2026 update
Chai's 2025 story was about model performance. Chai's 2026 story is about deployment inside big pharma.
In June 2026, Chai announced a partnership with Pfizer that includes platform access, early access to Chai-3, and a custom model trained on Pfizer's proprietary data and tailored to Pfizer's discovery workflows. That last part is significant. Pfizer is not just licensing a tool. It is embedding Chai's models into internal drug discovery processes and contributing proprietary data to improve them.
Chai-3, according to the company, doubles the success rate of Chai-2. It shows improved performance on therapeutic binding, multi-specific antibodies, hard-to-drug targets, and generalization across target classes. Doubling an already strong hit rate is a meaningful jump, and the fact that Pfizer is moving early on it suggests the internal validation was compelling.
The pattern here, open-source credibility with Chai-1, early commercial deployment with Chai-2, pharma workflow integration with Chai-3, looks like a deliberate platform expansion. Each step gets Chai deeper into how pharma companies actually run discovery programs.
Customers and traction
Traction falls into four buckets, and they reinforce each other in a way that matters for the business.
Academic adoption came first. Chai-1's open-source release in September 2024 got researchers integrating it into workflows almost immediately. By April 2026, it had accumulated hundreds of academic citations, and leading structural biology labs including the Baker Lab were reporting that Chai-1 was producing useful results on real research problems. That kind of adoption is hard to manufacture. It is organic credibility that makes pharma procurement conversations easier.
Pharma partnerships followed. Eli Lilly announced its collaboration in January 2026, paying a mid-eight-figure annual access fee to deploy Chai's antibody design tools across its therapeutic portfolio. Importantly, Lilly internally validated Chai's models before signing, which is not a formality for a company Lilly's size. Pfizer came in June 2026 with access to Chai-3 and a custom model tailored to Pfizer's workflows. Multiple discovery teams at other pharmaceutical companies have also confirmed active use of Chai's tools, including a J&J scientist who reported using Chai in parallel with AlphaFold for kinase inhibitor screening.
Government collaborations add institutional credibility. Chai was selected as a strategic industry collaborator for the UK's OpenBind initiative, alongside Isomorphic Labs and Genentech, aimed at building the world's largest open dataset of drug-protein interactions for generative model training.
Early access demand signals what is coming. After Chai-2 was announced, the company received hundreds of early access requests from pharmaceutical and biotech companies. That pipeline is not disclosed revenue, but it does represent a large potential commercial surface.
Business model
Chai operates a hybrid model. Chai-1 is open-access under Apache-2.0, free for anyone. That is not charity. It is distribution. Open-source releases build the scientific credibility that makes pharma buyers willing to evaluate the premium products.
Chai-2 and Chai-3 are commercial products sold through annual platform access fees. The Eli Lilly deal is the clearest public data point: a mid-eight-figure annual fee, which means something between $10 million and $99 million a year, for platform access and a custom model trained on Lilly's internal data. Standard partnership arrangements in AI drug discovery also carry royalty clauses, typically 0.5% to 5% of net sales for molecules discovered using the platform. If a blockbuster antibody with $1 billion in annual sales comes out of a Chai-assisted program, that royalty could be worth $5 million to $50 million per year.
The tension in this model is real. Broad platform licensing scales like software: build the model once, sell it many times, margins improve with scale. But the Lilly and Pfizer deals both involved training exclusive models on proprietary data and tailoring tooling to specific workflows. That is services work, and services do not scale the same way. If every major pharma deal requires a dedicated engineering team to customize and maintain, Chai starts looking more like a high-end contract research organization than a software platform.
"The bull case is software margins. The bear case is a very expensive AI-enabled contract research organization."
Market size
Chai is not capturing the antibody market directly. It is trying to capture the design layer underneath the antibody market.
The global market for monoclonal antibody therapeutics was about $288 billion in 2024, growing toward $628 billion by 2035. That is Chai's ultimate addressable universe if the company ever develops its own assets. But in the near term, Chai is selling tools that help pharma companies discover antibodies faster. The more comparable market is the high-throughput screening industry, valued at over $20 billion, which Chai could partially displace at the hit-generation stage. Add in AI platform licensing budgets, custom model development, and potential future royalty streams, and the addressable opportunity is large, but a meaningful fraction of it depends on whether Chai's platforms actually become the default tool inside discovery teams.
The market also has an important expansion dimension. Chai-2 showed 68% success rates on mini-protein designs (50 to 100 amino acids). Meier has stated there is no fundamental reason the same approach cannot achieve high rates on peptide therapeutics, RNA, and other molecular modalities. If the platform generalizes, the addressable market is not just antibodies.
Competitive landscape
The competitive map has three distinct layers, and Chai sits in a different relationship to each.
| 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 |
The most important thing to understand about this competitive landscape is that most players start from the same place. The core public training data for antibody structure prediction is the Protein Data Bank (around 200,000 protein structures) and the Structural Antibody Database (around 10,000 antibody structures). Every competitor uses these. A former director at Absci put it directly: "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 is a significant admission. If the starting data is the same, differentiation has to come from model architecture, compute scale, proprietary data from pharma partnerships, and speed of the iteration loop between computational predictions and experimental validation. The race is not just who has the best model today. It is who gets the best feedback loop between model predictions, wet-lab validation, pharma data, and real programs.
Chai's strategy is to win the pharma data flywheel. The Lilly deal involves training on Lilly's proprietary data. The Pfizer deal does the same. If these partnerships generate exclusive models that meaningfully outperform what is possible on public data alone, Chai builds a genuine advantage that is hard to replicate. If the exclusive models do not perform much better than the public baseline, the moat is weaker than it appears.
Valuation and fundraising
Chai raised a $70 million Series A in August 2025, valuing the business at $550 million. Four months later it raised a $130 million Series B at $1.3 billion. The cap table includes OpenAI as an early seed investor, Menlo Ventures leading the A, Oak HC/FT and General Catalyst co-leading the B, and a collection of individual investors including Greg Brockman, Sarah Guo, Fred Ehrsam, and Pfizer's former CSO Mikael Dolsten on the board.
Investors are not valuing Chai like a normal early-stage biotech with one or two assets. They are valuing it like a platform company that could become infrastructure for biologics discovery. That implies a revenue multiple story, not a pipeline story. The bar for that valuation is recurring platform revenue that grows with the number of pharma partnerships, plus the potential for milestone payments and royalties from drugs discovered using the platform.
The tension is that Chai has not disclosed revenue figures. The Lilly deal is a mid-eight-figure annual fee, which is a meaningful number but not enough to justify a $1.3 billion valuation on its own. The bull case requires Pfizer, and then the next 10 partners after Pfizer, and then a royalty stream that kicks in as Chai-assisted drugs move through clinical development years from now. That is a long timeline to underwrite.
| 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 has five distinct legs.
First, the market is large and structurally important. Antibody therapeutics are not a niche. They are at the center of pharma R&D, growing toward a $628 billion global market by 2035. Whoever makes antibody discovery meaningfully faster is sitting on a valuable position.
Second, the product progression is real. Chai-1 established credibility, Chai-2 demonstrated generative antibody design at a hit rate no prior computational method approached, and Chai-3 is already being adopted by Pfizer. That is a three-generation product track record in two years from founding.
Third, the pharma partnership model creates data flywheels. The Lilly and Pfizer deals both involve training custom models on proprietary data. If those models outperform public-data baselines, the companies become harder to displace. Each new proprietary data partnership widens the moat. And Pfizer's former CSO on the board is not a coincidence: it signals that Chai is being positioned inside pharma networks at the top of the house.
Fourth, Chai can expand beyond antibodies. The technical infrastructure Meier describes as modality-agnostic. Chai-2 already shows high success rates on mini-proteins. RNA therapeutics, peptides, and other biological drug classes are within scope, and each is a sizable market on its own.
Fifth, the platform shifts how pharma manages its portfolio. If discovery teams can cheaply test 100 therapeutic hypotheses instead of five, companies can pursue riskier and more differentiated targets. As Meier put it: "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 changes the logic of drug discovery R&D at a portfolio level, not just at the level of individual programs.
The bear case
The central risk is that Chai may be very good at producing binders, while the market ultimately pays for drugs.
The first risk is competitive convergence. The field is moving fast. Isomorphic Labs released IsoDDE in February 2026, which outperformed Chai-1 on protein-ligand structure prediction by 3x. Boltz-2 reached the accuracy of physics-based free-energy perturbation methods. RFantibody from the Baker Lab is open-source and trained on the same data everyone else uses. New model claims appear every six months. Technical leads are real but fragile. As the same Absci director noted: "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."
The second risk is the proprietary data gap. Antibody binding in a computational model is not the same as a drug that is safe, manufacturable, and efficacious. Data on immunogenicity, manufacturability, and formulation is largely proprietary to pharma companies and scarce in public repositories. There are fewer than 1,000 publicly available data points on antibody immunogenicity, while training a robust model requires over 10,000. Without that data, generative models can produce binders that fail on developability grounds. Chai's strategy for this, getting proprietary data through pharma partnerships, is the right approach, but it requires those partnerships to actually transfer training-useful data, which is not guaranteed.
The third risk is the platform value capture problem. This is structural to the business model, not specific to Chai. When a computational platform discovers a drug candidate, the pharma company developing the drug captures most of the value. A royalty of 0.5% to 5% on net sales from a blockbuster antibody sounds meaningful, but it is a small slice of the economics compared to the pharma company that does the clinical development and commercial launch. Schrödinger, a computational drug discovery platform with a strong track record, has a market cap of around $870 million as of 2026, roughly a quarter of the value created by assets discovered using its platform. Chai is valued at $1.3 billion on far less revenue history. That gap has to close somehow.
The fourth risk is the services trap. Both the Lilly and Pfizer deals involve building custom models tailored to specific workflows. Done twice, this is manageable. Done with 20 pharma partners, it starts to require separate engineering teams maintaining separate codebases while also advancing the core platform. That organizational complexity is the difference between a scalable software business and a contract research organization that happens to use AI.
The fifth risk is the clinical validation gap. No antibody designed by Chai has completed clinical development. An antibody that binds computationally, then validates in initial assays, still has to pass developability screens, preclinical safety studies, Phase I dose escalation, Phase II efficacy studies, and Phase III trials. That pipeline takes 10 to 15 years from hit to approval. The market is currently pricing in the potential of the platform, not validated clinical outputs. If early clinical programs using Chai-designed antibodies run into unexpected toxicity or efficacy problems, the narrative 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 is not trying to be a classic biotech at first. It is trying to become the design layer for biologics. That makes it feel more like infrastructure than a therapeutics company, which is a fundamentally different kind of business to run and to value.
The most exciting version of Chai is not just faster antibody discovery. It is a closed-loop system where generative models, structural prediction, pharma data, and wet-lab validation continuously improve each other. Every Lilly and Pfizer deal is not just revenue. It is training data for the next generation of models, which makes the next deal easier to close and the models harder to replicate on public data alone. That flywheel, if it actually spins, is what justifies a billion-dollar valuation on a two-year-old company.
The part I find harder to underwrite is the clinical gap. Biology has a way of punishing clean software analogies. A 20% computational hit rate is genuinely impressive relative to where the field started, but the harder questions are downstream: will Chai-designed antibodies pass immunogenicity filters? Will they be manufacturable at scale? Will any of them work in humans? We will not know the answers to those questions for years, and the market is already pricing in a strong positive answer.
The services trap is the other thing I would watch closely. The Lilly and Pfizer deals are both validating and potentially constraining. The "mid-eight-figure annual fee plus custom model" structure is a great first deal and a worrying template for the fifth one. Each custom engagement that requires Chai engineering resources to maintain is a headcount cost that does not appear in the licensing revenue number. At some point, the company has to decide whether it is building a platform or running a bespoke AI research service for big pharma.
The thing that makes me more optimistic about Chai than most of its AI drug discovery peers is the founding team's clarity about what they are not. Deliberately skipping the wet lab was a real bet. Most previous AI biotech companies built labs because they felt they needed data control. Chai decided to get data control through partnerships instead. Whether that turns out to be right depends on whether pharma actually shares the right training data, not just access to the platform. That is the key variable I would be watching in every new deal announcement.
Chai Discovery is one of the strongest tests of whether AI drug discovery can become a product business before it becomes an asset business. If Chai becomes the default tool scientists use to design antibodies, it could be one of the defining AI biotech companies of this decade. If the value keeps flowing mostly to pharma assets, or if competitors match the model performance, Chai may be forced to move deeper into asset development, like many platform biotechs before it. The next two to three years of clinical data from Chai-assisted programs, and the next handful of pharma deals, will start to tell us which path this is.