Why Chai Is Worth $3.8 Billion

Last week, Chai Discovery raised $400 million at a $3.8 billion valuation. The day before, it announced that Novartis would use its newest AI model, Chai-3, to design therapeutic antibodies across several targets. Pfizer and Eli Lilly already have access as well. So, on the surface, the story is pretty simple. Big pharma likes Chai, and investors are willing to pay a lot for it.

But there is something very interesting about this business model. Chai is selling a powerful discovery tool to companies that compete directly with one another. Pfizer can use it, Lilly can use it, and now Novartis can use it too.

If they all have access to roughly the same frontier technology, then access to Chai alone cannot be the thing that separates the winner from everyone else.

So what is it then? I’d like to compare their model to the use of Bloomberg Terminal.

Almost every major bank, hedge fund, and asset manager uses Bloomberg. They see much of the same market data and use many of the same analytical tools. Yet they do not make the same trades or earn the same returns. Bloomberg gives them a common operating system. The advantage comes from what each firm brings to it, including its strategy, private information, judgment, and ability to act.

Chai may be building something similar for drug discovery. Simply, Chai helps scientists design molecules. A pharmaceutical company can give the model a biological target and a set of desired properties, and the model can propose antibodies that may bind to it.

Novartis is now applying Chai-3 across multiple antibody programs. Pfizer will receive both the shared model and a custom version tied to Pfizer’s proprietary data and workflows. Lilly has a similar arrangement, including a purpose-built model trained on Lilly’s own data.

That custom layer is probably the most important part. The common model may become widely available, but every company still begins with different biology. They choose different diseases, targets, patient groups and experiments. They have different internal datasets and different ideas about which mechanism is worth pursuing.

Once a molecule is designed, they still need to test it, decide whether the assay is meaningful, manufacture it, run clinical trials and know when to end the program. So, even if Chai makes molecular design much easier, it does not make every pharma company equal. It may do the opposite.

When one part of a process becomes cheaper and more widely available, the remaining scarce parts become more valuable. If good antibody designs become easier to generate, then knowing which antibody should be designed becomes the harder problem. The advantage moves toward disease biology, proprietary data, experimental systems, and development judgment.

This may also help explain Chai’s valuation. Investors may not be valuing it like a normal biotech with a few drugs in development. They may be valuing it as shared infrastructure that can sit inside many pharmaceutical companies at once.

Bloomberg does not need to make the best trade. It sells the platform used by thousands of firms making trades. Chai may not need to own the winning drug. It may need to become the place where many winning drugs begin.

The irony is that a company built around AI molecular design could end up proving how valuable old-fashioned pharmaceutical knowledge still is. Chai may make the design engine common. The real competition then moves to everything each customer knows that the others do not.

Insilico dosed the first patients in a Phase III trial of rentosertib, its oral TNIK inhibitor for idiopathic pulmonary fibrosis. The study will enroll 320 patients across 47 sites in China and is randomized and placebo-controlled. It is also the first AI-designed drug against an AI-identified target to reach a registrational trial. Most of the evidence for AI drug discovery still sits in preclinical work, where candidates are cheaper to make and failure is easier to absorb. Rentosertib now has a trial large enough, and a readout clear enough, to test the claim properly.[Link]

The Boston Globe told general readers that the bench scientist’s job is shifting toward curation. Its July 13 newsletter walks through in-silico synthesis and the idea of the scientist as a “curator,” citing Anthropic’s Claude Science. The interesting part is where the argument appeared. This was a metro newspaper, not Endpoints or STAT. Once a framing moves out of the trade press and into a general-interest publication, it starts to harden into an assumption. What remains unclear is whether laboratories are actually reorganizing work around it. [Link]

BCC Research announced that AI has cut drug development timelines by 70%, to 12–18 months, on $2 billion of investment. It's a vendor promotional release with self-reported figures, the firm's second in five days. Worth knowing anyway, because it's the number the trade press will quote as baseline all year. Set it against BCG's: AI candidates roughly double trial success, 5–10% to 9–18%, nearly all of it preclinical. Both can't be describing the same industry. [Link]