Part Two · The Stack · Layer 9 of 9

The Discoverers — The Tenants Who Run the Loop to Find Drugs

At the top of the stack sit the companies everyone actually reads about: the AI-native drug hunters. They are the tenants of the machine this primer has described, assembling the models, the synthesis, the measurement, and the data into a single engine pointed at one goal, finding medicines faster and cheaper than the industry's dismal historical odds. They are also, for an investor, the most seductive and the most dangerous names in the entire field, because they wrap the binary risk of biotech inside the margins and narrative of software. Understanding why requires understanding how they actually make money, which is where we start.

Two Business Models Wearing the Same Clothes

Almost every company in this layer is one of two things, or an uneasy blend of both, and the distinction determines everything about how it should be valued.

The first archetype is the platform company. It builds a discovery engine and monetizes the engine itself, licensing the software, running discovery for many partners, collecting fees and milestones across a portfolio of programs it does not fully own. Revenue is recurring and diversified, closer to software or tooling than to biotech. A platform can be judged on adoption, partnerships, and the breadth of its pipeline, and it does not live or die on any single molecule.

The second archetype is the asset company. It uses its AI to build its own proprietary drug pipeline and captures value the old-fashioned way: by owning drugs that, if approved, are worth billions, and if they fail, are worth nothing. This is biotech economics, binary, lumpy, and brutal. The AI is a means; the drug is the product.

The trouble is that most companies in this layer present as platforms while being valued, ultimately, on their assets. A brilliant discovery engine still has to produce a molecule that survives Phase 2, and as the earlier chapter on the drug pipeline made clear, roughly nine in ten candidates that reach the clinic fail, most of them in Phase 2, for lack of efficacy, after most of the money is spent. No model yet built has repealed that statistic. An AI-designed drug that binds its target perfectly still fails if the target turns out not to drive the disease. This is the hard truth that anchors the whole layer: a platform can be genuinely excellent and its lead asset can still die in the clinic, and the market will punish the stock as if the platform never worked.

The Data Flywheel, One More Time

The most durable edge a discoverer can have is not its model, models diffuse, as the Design chapter argued, but its proprietary data. A company that owns a way to generate experimental data no one else has, at a scale no one else can match, feeds a model that competitors cannot replicate no matter how good their algorithms are. The measurement produces the data, the data trains the model, the model prioritizes the next measurements, and the loop compounds. This is why the strongest discoverers are the ones that own a piece of the measurement layer, not just the design layer. The company that owns the "hardware tested them" step owns the flywheel.

Deal Economics: How Partnerships Fund and Validate

Before a discoverer has an approved drug, which may be never, it survives on partnerships with large pharma. The structure is standard: an upfront payment when the deal is signed, a series of milestone payments as a program hits development and regulatory markers, and royalties on eventual sales. These deals do two things at once. They fund the discoverer with non-dilutive cash, and they serve as third-party validation, when a Lilly or a Novartis writes a large check and stakes a program on a startup's engine, it is a credible signal that the engine is real. The size and quality of a discoverer's partnerships are often a better read on its platform than any self-reported metric.

The Roster, Public and Private

The public pure-plays are where investors can actually participate, and they span the platform-to-asset spectrum:

• Schrödinger (SDGR), met in the Design layer, is the most platform-like: a physics-and-ML software business with an owned pipeline bolted on as upside.
• Recursion (RXRX) is the highest-profile data-flywheel bet, an industrialized wet lab generating vast phenomics datasets (images of how cells respond to compounds) to train its models, strengthened by its merger with Exscientia and backed by NVIDIA. It is the purest listed expression of "platform, not a drug," and it trades with the volatility that implies.
• AbCellera (ABCL) is an antibody-discovery engine with a royalty-bearing model, optionality across many partnered programs rather than a bet on one asset, which makes it structurally more platform than asset. Absci (ABSI) uses generative AI to design antibodies and validate them in its own wet lab.
• Tempus AI (TEM) is a data-layer play more than a drug hunter, a multimodal clinical-and-molecular dataset that trains diagnostic and discovery models, monetized across many customers. In July 2026 it agreed to acquire Personalis (PSNL) for a $1.5B enterprise value, buying the ultrasensitive NeXT Personal MRD test to own the cancer-monitoring measurement outright, a data-layer name paying up for the instrument that generates the data, which is this primer's thesis compressed into one deal.
• Two 2026 IPOs extended the public set: Eikon Therapeutics (EIKN), built on single-molecule imaging to watch drugs act on proteins in living cells (founded by a Nobel laureate), and Generate:Biomedicines (GENB), whose Chroma generative model designs proteins across modalities.

The private leaders are, frustratingly, where the frontier and the best science mostly live:

• Isomorphic Labs is the flagship, the DeepMind spinout built on the AlphaFold lineage, which raised roughly $2.1 billion and signed target-rich partnerships with Eli Lilly, Novartis, and Johnson & Johnson. It is the clearest "AI biolab" tenant in existence, and it cannot be bought publicly. It also just got materially more serious: when DeepMind broke up the AlphaFold team in July 2026, part of that talent went to Isomorphic, and on August 5 Demis Hassabis stepped back from running Google DeepMind day to day, taking its chair and Alphabet's chief-scientist role, and said he would lean further into Isomorphic to work on curing disease. The man who won a Nobel for predicting protein structure has chosen the drug pipeline over the model lab. Read that as a statement about which layer he thinks compounds.
• Terray Therapeutics is the discoverer that owns its own measurement layer rather than renting it, which is this primer's thesis embodied in a single company. Its EMMI platform (Experimentation Meets Machine Intelligence) runs a proprietary, ultra-dense microarray that physically measures billions of small-molecule binding interactions — on the company’s figures, on the order of a billion measurements a quarter — a dataset no competitor can buy, and feeds its COATI foundation model; its newer TerraSynth capability closes the loop on the make side with AI-designed synthesis. Because Terray owns the data, the instruments, the models, and the harness that runs the loop, its edge compounds instead of diffusing, unlike a design shop whose model leaks within months. NVIDIA-backed, it is pointed at an internal pipeline led by a brain-penetrant small-molecule inhibitor for multiple sclerosis, alongside partnered programs. (Disclosure: I am an angel investor.)
• Xaira (launched with around $1 billion), insitro (machine learning grounded in large in-house functional-genomics data), Genesis Therapeutics (deep learning for hard small-molecule targets), Iambic (physics-informed models, with a candidate already in Phase 1), Chai Discovery, and Cradle round out a private cohort collectively holding many billions in capital.

China has produced two names worth knowing. Insilico Medicine (2179.HK) runs an end-to-end platform (Pharma.AI) and reached a milestone the company describes as a first: Rentosertib, which Insilico says is the first drug with both an AI-generated target and an AI-generated molecule to reach a Phase IIa clinical trial. Treat the "first" as the company’s framing, but the underlying fact, an AI-originated candidate in mid-stage trials, is real evidence that the full loop can produce a clinical candidate. XtalPi (2228.HK) pairs quantum-physics simulation with AI and lab robotics. Both carry the China regulatory overhang discussed elsewhere.

The incumbents are not spectators, and this is the most important point for a risk-conscious investor. Eli Lilly (LLY) is the most aggressive of the large pharmas, running federated AI programs on NVIDIA's stack (its TuneLab platform) and, as of February 2026, its own LillyPod supercomputer, a 1,016-GPU NVIDIA Blackwell DGX SuperPOD built to train proprietary models on data from more than a billion dollars of internal experiments, while its GLP-1 franchise generates the cash to fund it. Note what Lilly chose to build versus buy: it rented nothing on the data side and spent its own capital standing up LillyPod, because the GPUs are a purchase order and the billion-dollars-of-experiments dataset behind them is not. That combination, an AI program backed by one of the strongest cash engines in the industry, makes Lilly the least binary way to own the theme, because you are not betting the company on a single readout. Novartis (NVS) and Johnson & Johnson (JNJ) are renting the leading external engines (both are Isomorphic partners), and Roche/Genentech (RHHBY) runs deep internal efforts. The incumbents have what the startups lack: the data, the clinical infrastructure, the balance sheet to absorb failures, and the distribution to monetize a win.

The Read on the Tenants

The discoverers carry the field's asymmetric upside, a genuinely AI-designed blockbuster would reprice the whole category, and its genuine binary risk, since most of them are, underneath the software gloss, betting on drugs that will probably fail. Three practical conclusions follow. The AI-armed incumbent, Lilly above all, is the least-binary way to own the theme, because a diversified pharma with a cash engine survives the failures that would sink a single-asset startup. The public pure-plays (Recursion, AbCellera, Absci, Tempus) are the higher-beta "platform, not a drug" bets, appropriate in small size and with clear eyes about the clinical odds. And the best privates, Isomorphic, Xaira, Terray, are simply not investable on the public tape, which is the recurring lesson of this entire primer. When the most exciting tenants are private and the ones you can buy are binary, the durable trade is to own the layers every tenant has to rent from: the measurement, the data, the reagents, and the compute that get consumed whether the drug works or not. You would rather sell picks to all the miners than bet on which one strikes gold.

That is the machine, end to end. Below the paywall: the investment framework that falls out of it, and the AI-Science Basket, a proposed twenty-name book with weights, tiered from the measurement core to the tenant optionality, the one-line case for each, the concentration risk the tiers hide, the bear case that could break the whole thesis (including the one that inverts it), and the specific catalysts I am watching. If you take one idea from the free section, take this: the money is not in guessing the drug, it is in owning the bench that every guess has to run on.

© BEP Holdings · Ben Pouladian. Research and commentary, not investment or medical advice.