Why Pharma Finally Bought AI Software

William Palmer

Four billion dollars. That is what a two year old startup named Chai Discovery was valued at around the time of the JP Morgan Healthcare Conference in San Francisco this past January. And if you know anything about biotech, you know that number sounds completely absurd on paper. Historically, if an AI biotech startup showed up at JP Morgan, pharma execs would nod politely, look at their predictive models, and say, cool story, go build your own drug pipeline, bring us clinical trial results, and maybe we will pay you some biobucks in five years. It was the classic partner or build your own pipeline trap. You had to prove your software worked by spending a decade taking actual physical molecules into clinical trials yourself, because nobody trusted pure software predictions across a whole portfolio.

William Palmer

So what shifted in January? Why did giants like Eli Lilly, Novartis, and argenx suddenly flip the script and sign massive pure software tooling deals directly with an OpenAI backed startup? Well, um, it comes down to a fundamental phase shift in what the models are actually predicting. We went from pure structural prediction, like just predicting what a protein folds into, to accurate binding affinity models. That means predicting exactly how tightly and specifically a molecule will cling to a target, and crossing what the industry calls the good enough to trust threshold. And look, when you cross that threshold, you are not just optimizing existing workflows by five percent. You are suddenly enabling drug design teams to engineer mechanisms that lab based trial and error could literally never touch. Think about bispecific antibodies that need to bind two completely different targets simultaneously, or complex GPCR agonism. Trying to find those in a physical wet lab with mice or high throughput screening is like looking for a needle in a sandstorm.

William Palmer

I, I, I actually think about this a lot in terms of race car tuning, which is a bit of a personal obsession of mine. Decades ago, if you wanted to setup a chassis or dial in an engine, you sent a driver out on track, waited for them to lap for twenty minutes, and then listened to them say, well, uh, it feels a little loose in turn four. It was purely empirical, slow, and full of noise. But modern telemetry changed everything. Now, sensors give you real time telemetry on damper velocity, tire surface temperatures, downforce load, every millisecond. You stop guessing and you move to closed loop engineering iterations in the garage before the car even touches the asphalt. That is what binding affinity models are doing for molecular design. They give chemists real time molecular telemetry, turning blind trial runs into true structural design.

William Palmer

When you talk to Matthew McPartlon, the cofounder of Chai Discovery, and Neil Patil, their product lead, you notice something right away. They did not build a chat bot. They did not chase the conversational AI hype where you type a prompt and pray a drug comes out the other side. Instead, they built what is essentially a CAD program, a Photoshop for molecules, built around a vector style editor tailored to how wet lab chemists actually manipulate three dimensional protein geometry. You move a side chain here, adjust a binding pocket there, and the physics aware model updates the binding dynamics in real time.

William Palmer

This is where biology stops feeling like a unpredictable empirical science and starts behaving like deterministic engineering. In a recent bioRxiv preprint on their Chai 2 platform, they demonstrated over eighty six percent strong developability profiles in de novo full length monoclonal antibody designs. Think about that number for a second. Over eighty six percent of their de novo full length monoclonal antibodies were deemed highly developable right out of the gate, without needing months or years of high throughput screening loops to fix aggregation or stability issues. You are effectively one shotting candidates straight toward animal trials.

William Palmer

And that leads to a wild framing that Neil and Matt touched on, which has been echoing in my head ever since. In the world of Large Language Models, everyone talks about the economic value per token. But protein tokens might actually have the highest downstream economic value of any token in human history. A single sequence of amino acid tokens, if designed correctly, can generate billions of dollars in enterprise value and save hundreds of thousands of lives. When you view capital allocation through that lens, pharma investment starts looking less like traditional research and development and more like pure portfolio optimization across compute, candidate volume, and clinical throughput.

William Palmer

Which leaves us with a really fascinating tension on the horizon. Will incumbent pharma giants maintain their dominance by simply absorbing these software layers to turbocharge their existing commercial engines? Or will one shot AI design tools get so good that the hard part of pharma shifts entirely to manufacturing and distribution, turning legacy drug pipelines into commodity infrastructure? I, I, I do not think we know the answer yet, but watching the software layer take over biology is easily the most thrilling race in tech right now. Alright, that is it for today. Talk to you soon.