A four-part series on the hard biological and clinical infrastructure that will determine where value accrues in AI-bio.
The first wave of AI-bio capital has been priced around the model. I think that is increasingly the wrong place to look. The model matters, but the harder question is what it sits on top of: whether the biology is predictive, whether the clinic can absorb more candidates, and whether the company can learn quickly enough from human data.
AI inherits the biology, the experimental systems, the clinical infrastructure, and the regulatory frame it operates in. Capital flows have so far concentrated in the layer that looks like software; meanwhile, the adjacent layers that determine whether the software actually produces medicines have been under-priced. The companies that win may be the ones that build the infrastructure the model depends on. That is the primary thesis of this series.
Why Better Biology Beats Bigger Models. AI-bio's data problem is not mainly data volume. It is biological fidelity. Models trained on weak experimental systems learn weak proxies for human biology. The investor question is what the predictive validity of the system generating the training data actually is.
PublishedWhy More Candidates Won't Mean More Drugs. Even if AI generates better candidates, the clinical system cannot automatically absorb them. The bottleneck moves from candidate generation to first-in-human translation. The investor question is how a company gets from candidate to human readout faster than the market.
PublishedAI rewards iteration, but the iteration loop is most valuable when it reaches patients. Examples include Recursion's underperformance and Hengrui's outlicensing pattern. The investor question is how many independent human learning loops a company can run per year.
Coming soonEurope should not imitate the US-China speed race. It should build compound assets around fidelity, regulatory niches, and clinical-data products. The investor question is whether a European company is exploiting a structural advantage, or just running a slower version of a US or Chinese comparable.
Coming soon