Data moats in AI biotech are wet, not digital
Public datasets are a starting point, not a fortress. The durable advantage in AI biotech is operational: assays, automation, and the discipline to keep provenance.
Every few months someone argues that AI biotech will be won by whoever has the best dataset. The claim is usually illustrated with a list of public atlases, patent filings, or scraped literature. It sounds plausible. It is mostly wrong.
Digital data moats erode. Public datasets get larger. Models get better at learning from less. Anything that can be downloaded, licensed, or reconstructed from publications will eventually be. The advantage that compounds is not a static corpus. It is the ability to generate new, decision-relevant data and to know what it means.
The copy problem
Consider a company that trains a model on a large public single-cell atlas. The model is useful, but the data is available to everyone. The next competitor trains a slightly better model on the same data and undercuts on price. There is no moat, only a head start.
Now consider a company that runs its own perturbation screens, standardizes the protocols, records failures, and feeds the results back into model design. The data is not public. It is also not merely data — it is a set of operational decisions about what to measure, how to measure it, and when a result is trustworthy. That is much harder to replicate than a weights file.
Wet data is expensive for a reason
It is tempting to treat wet-lab data as a slow, old-fashioned input that AI will eventually make unnecessary. In practice, the expense is the point. A well-designed experiment encodes a causal intervention. It tells you what happens when you change something, not just what correlates with what. That is precisely the information that public observational data lacks.
The bottleneck is not sequencing or imaging. It is the chain of custody: which batch, which passage, which operator, which protocol version, which exclusion criteria. A dataset without provenance is a liability disguised as an asset.
What actually compounds
Three things tend to compound in AI biotech, and none of them are a static dataset:
- Experimental throughput with quality control. The ability to run more informative experiments per dollar, without letting quality drift.
- A data model that captures context. Not just measurements, but the conditions under which they were made and the decisions they informed.
- Institutional memory. A culture that records negative results, revisits old assumptions, and treats evaluation as a living process.
These are operational advantages, and operational advantages are slow to build. That is exactly why they are defensible.
The uncomfortable implication
If the moat is wet, then the companies that win will look less like software companies and more like research organizations with excellent software. They will invest in lab operations, automation, and data hygiene before they invest in a flashy model. They will be boring in the ways that matter.
That is a harder story to tell at a demo day. It is also the story that tends to survive the next model release.