Chai Discovery raises $400 million to design antibodies with AI at a $3.8 billion valuation

Chai Discovery said Monday it raised a $400 million Series C at a $3.8 billion valuation, led by Index Ventures with Kleiner Perkins, Sequoia Capital, and Dimension. Existing backers OpenAI and Thrive Capital came along. The round roughly triples the valuation the San Francisco company carried in December, when it raised $130 million at $1.3 billion, and brings total funding to around $630 million since it was founded in 2024.

The number that matters more than the valuation is a different one, and it comes with an asterisk.

What the models actually do

Chai builds AI models for the design step of drug discovery, specifically antibodies. An antibody is a protein the immune system uses to grab onto a target, and the design problem is brutal: there are on the order of a quintillion possible sequences that might bind a given target, and the traditional approach is to screen millions of physical candidates in a lab and test them one at a time.

Chai's pitch is that you can skip most of that. Its 2025 model, Chai-2, generates full antibody sequences and structures from scratch, conditioned only on a target protein and the spot on it you want to hit. The company calls it zero-shot, meaning it designs candidates for a target it was not specifically trained on. Chai-2 was, by the company's account, the first generative model of its kind to clear a double-digit experimental success rate, meaning at least ten percent of its computer-designed candidates actually bound the target when someone made them and tested them in a lab.

That is the metric to hold onto, because "success rate" here is doing specific work. It measures whether a designed molecule binds its target in a dish. It does not measure whether that molecule becomes a drug. The newer model, Chai-3, is described by the company as a step change over Chai-2, and SiliconANGLE, working from the company's figures, put the improved binding hit rate somewhere around 35 to 40 percent. Take that as the company's claim rather than an independently reproduced benchmark, because no outside lab has published a head-to-head.

The real signal is the customer list

Funding rounds are easy to inflate. Contracts with people who have their own chemists are harder. Chai has three worth naming: a collaboration with Eli Lilly announced in January that included a model trained on Lilly's proprietary data, a licensing deal with Pfizer in June that gave Pfizer access to Chai-3, and a collaboration with Novartis disclosed the day before this raise. Those are companies that can run their own wet-lab validation and would walk if the designs did not bind. That they are paying is a stronger data point than the valuation.

"AI drug discovery has moved from promise to deployment," CEO Joshua Meier said in the announcement. On the design step, that is close to fair. The catch is where deployment stops.

The part nobody has done yet

Designing a binder is the front of a very long pipeline, and the field has a clean record of not finishing it. By one industry tally from the analytics firm Excelra, roughly $20 billion has gone into generative-AI drug discovery, and the number of AI-designed drugs approved for use so far is zero. There are now more than 170 AI-originated drug programs in clinical trials, up from a couple dozen in 2023, so the approvals question is genuinely still open rather than settled in the negative.

What the early clinical data shows is where the AI advantage thins out. AI-originated candidates clear Phase I at something like 80 to 90 percent, which is high, but Phase I mostly tests safety in healthy people. In Phase II, where a drug has to actually work in patients, the pass rate drops to around 40 percent, which is about what traditional discovery gets. Being better at designing a molecule that binds a target does not yet translate into being better at picking a target that matters, and Phase II is where that bill comes due.

What to watch

Chai is good at the thing it claims to be good at, and the pharma contracts suggest the designs hold up under someone else's microscope. That is a real advance in a field with plenty of vaporware. The open question is not whether the models design better binders. It is whether a better binder, this far upstream, changes the odds on the only number that counts, a drug that gets approved and works. That answer is a Phase II readout away, not a funding round, and no amount of valuation moves it up.