AI-Designed Antibodies Move From Demo to the Pharma Supply Chain

For a decade, the promise of AI in drug discovery lived mostly in demos and press releases. That is starting to change. AI models that design proteins and antibodies — the lineage that began with DeepMind's AlphaFold and has since spawned a wave of successors — are no longer just predicting how molecules fold. They are proposing viable drug candidates outright, and big pharmaceutical companies are signing real money to license them.

The clearest signal came this week. Chai Discovery, a two-year-old startup building AI models that design antibodies, announced a $400 million Series C that nearly tripled its valuation to about $3.8 billion. The round was led by Index Ventures alongside Kleiner Perkins, Sequoia Capital and Dimension, with OpenAI among the existing backers carried over from earlier rounds. That valuation jump — from roughly $1.3 billion just seven months earlier — is not the interesting part. The interesting part is what the money is chasing.

From screening millions to curating the space

Traditional drug discovery has always been a brute-force search problem. To find an antibody that binds tightly and specifically to a disease target, researchers screen enormous libraries — millions of candidates — hoping a handful survive the gauntlet of binding, stability and manufacturability tests. It is slow, expensive, and mostly a numbers game.

AI protein design flips the logic. Instead of screening the space, a model curates it, proposing a small set of candidates it predicts will work before anyone touches a lab bench. Chai's newest model, Chai-3, is the case study the industry is pointing to. The company says it roughly doubles the "hit rate" on molecular targets over its predecessor Chai-2 — from around 20% to a reported 35–40% — and produces antibodies that bind far more tightly to their intended targets. When the shortlist a model hands you is more than a third good, the economics of discovery change.

The licensing deals are the real story

Valuations are easy to inflate; licensing deals are harder to fake, because they mean a pharmaceutical company with its own scientists decided the outside model was worth paying for. Chai has three that check out:

• Pfizer signed a licensing agreement in June 2026 granting access to Chai-3 plus a custom model trained on Pfizer's proprietary data.
• Eli Lilly entered a research collaboration announced in January 2026, including a bespoke model trained on Lilly's data and access tied to Lilly's TuneLab drug-discovery platform.
• Novartis established a formal collaboration to use Chai's latest models, including Chai-3, to discover antibodies across multiple therapeutic programs.

This is what "moving from demo to the supply chain" actually looks like: not a splashy paper, but three of the largest drugmakers in the world routing part of their earliest, hardest discovery step through an external model.

The last mile still runs at old-world odds

Here is the honest caveat, and it matters. Designing a promising candidate on a screen is not the same as proving it works in a human body. That last mile — preclinical validation, then Phase I, II and III clinical trials — still runs at roughly the same brutal odds it always has.

The numbers make the point. Despite an estimated $20 billion poured into generative-AI drug discovery, no AI-designed drug has yet won regulatory approval. AI-originated candidates reportedly clear Phase I trials at high rates — 80 to 90% — but that advantage largely evaporates by Phase II, where pass rates fall back to around 40%, roughly matching drugs discovered the old way. Phase II is where a molecule has to prove it actually treats the disease in patients, and no amount of upstream design elegance guarantees that.

In other words, AI is compressing the front of the pipeline — the search for a plausible candidate — dramatically. It has not yet moved the part of the funnel where most drugs die.

Why this still counts as a turning point

It would be easy to read the clinical bottleneck as a reason to stay skeptical, and skepticism is warranted on any single candidate. But the shift underneath is real regardless. The most expensive, luck-dependent stage of discovery — finding a molecule worth developing at all — is being turned from a lottery into something closer to engineering. That compresses timelines and cost at the exact point where pharma has historically burned the most resources on dead ends.

The trend does not depend on Chai specifically. It sits on a broader foundation of AI structure-prediction and protein-design work that has matured from research curiosity into infrastructure. Chai is simply the sharpest current example of the pattern: models that design biology, and incumbents willing to pay for the output.

The demo phase is ending. The proof phase — measured in clinical trials, not funding rounds — is where the story goes next.