Stanford MS&E435: Chai and Claude - Who Will Capture the Value in AI Biotech?

The final lecture of Stanford MS&E435, which I wrote about, focuses on life sciences. It asks where the economics sits in the AI stack. In biotech, one successful result can be worth tens of billions of dollars, but years of biology, clinical trials, and regulatory work lie between a demo and sales. Host Apoorv Agrawal talks with Joshua Meier, co-founder of Chai Discovery, and Eric Kauderer-Abrams, Anthropic's Head of Life Sciences. They propose not “a medicine from one prompt,” but a new architecture for R&D.

What the conventional process looks like

Compressed into one line: disease and patient population → biological target → therapeutic modality → initial hit → optimized lead → drug candidate → preclinical studies → Phases I–III → approval and manufacturing

The path often takes 10–15 years, and a program can die at any stage: the wrong target, a toxic or unmanufacturable molecule, or no effect in clinical trials. Therefore: binder ≠ drug candidate ≠ medicine.

What Chai and Anthropic propose

1️⃣ Chai is building “CAD for molecules”—the inner loop. Its model takes a target structure and designs antibodies intended to bind to it. In Chai-2's own non-peer-reviewed preprint, the company reports a hit rate of roughly 16%: it tested no more than 20 designs against each of 52 targets and found at least one binder for half of them in under two weeks. It is a strong discovery result, but not yet evidence that a medicine works.

2️⃣ Claude is the outer loop: read the literature, choose a tool, run a specialist model, prepare an experiment, interpret the data, and propose the next step. Chai designs the component, Claude coordinates the process, and the laboratory remains the source of truth.

Where the acceleration comes from

- Conventional discovery screens large libraries and repeats the “synthesis → test → analysis → new design” cycle. Here the model narrows the search to dozens of designs, while the agent reduces handoffs among people, software, and laboratories. The entire feedback loop becomes faster.

- AI may later help with targets and clinical trials. Eric suggests that the full cycle could shrink to about five years and eventually to only a few. That remains a forecast: human biology, long observation periods, and regulators cannot simply be removed from the pipeline.

- The amount of laboratory work may actually increase. When every experiment becomes cheaper and more informative, it makes sense to test more hypotheses—Jevons' paradox in a lab coat.

Where the money will be

- Pharma owns the candidate, clinical infrastructure, manufacturing, and sales. This is where most of the value of a successful medicine remains for now.

- Toolmakers—Chai, Anthropic, lab-automation providers, and CROs—sell access, compute, and experiments. To capture more value, they must prove a higher probability of program success. Eric nevertheless warns that selling tools to big pharma is a difficult business.

- AI-native biotech may enable a small team to own molecules and license them after early proof. This offers the maximum upside together with clinical and capital risk. If design becomes broadly available, the moat shifts to target selection, data, the laboratory feedback loop, and rights to the medicine.

How large is the market?

IQVIA estimates that the global life-sciences industry generated about $1.94 trillion in revenue in 2025, while biopharma spent about $199 billion on drug development. These are different circles: $1.94 trillion is not the TAM (Total Addressable Market) for Chai or Claude. Tools compete for part of the R&D budget; the owner of an approved medicine competes in the end market.

One molecule illustrates the gap: tirzepatide, sold as Mounjaro and Zepbound, generated 36.5 billion in 2025 sales for Lilly. The central question is therefore not who first generates an impressive molecule, but who closes the “hypothesis → design → experiment → data” loop, retains the rights to the result, and carries it all the way to the patient.

#AI #Biotech #DrugDiscovery #LifeSciences #Engineering #Economics