Foundries vs Navigators: Lowering the Cost of Science

Anthropic may be opening a wet lab, but the more immediate shift in AI and science is less glamorous. Adrian Sanborn argues that AI is already changing front-line research by making the thinking around experiments much cheaper, even when the physical work still takes days or weeks to verify. That matters because science does not move when code gets faster alone; it moves when the whole loop gets faster.

His split is simple. “Foundries” shrink the cost of doing by industrializing measurement, using technologies like next-generation sequencing, multiplexing, high-throughput microscopy, and physical automation. The names in that camp include Xaira, NewLimit, Octant, Tahoe, Endura, Insitro, Eikon, Noetik, Lila, and Periodic Labs. “Navigators,” by contrast, spend the surplus of thinking inside the company itself. The model is not the product. It shapes decisions, tools, and which questions deserve an experiment.

At Endura Therapeutics, that shows up in the annoying bits that used to slow research down. When a protocol changes, the analysis code has to change with it. In the old setup, that work often sat with a separate computational person, which created friction and pushed teams toward fewer changes. Now the same adjustments can be handled in an afternoon, and an internal dashboard can be built in a few hours. The scientist who ran the experiment can present the results instead of waiting in line behind someone else’s Python notebooks.

The bigger effect is on how labs organize themselves. A data portal built in-house can match the quirks of a team’s own work instead of forcing every lab into the same commercial template. That takes a day to implement, but the debate over what belongs on one screen can take weeks, because it forces people to decide what actually drives decisions. The old rule was never build what you can buy. Sanborn’s version is sharper: build the tools that shape how you think.

Endura is also using that logic for one of the biggest bets in drug discovery: choosing where to focus. Because the company is developing CRISPR in a pill, it had to build a new sequencing method and then triage the entire map of disease instead of picking from a neat shortlist. It used two stages of LLM research agents across about 500 disease targets, then about 100, to produce report-level analysis that would have taken about one person-year in the first pass and nearly a century of expert time in the second. The human checks still happen, but the search is suddenly broad enough to be useful.

That is the real story here. AI in science is not just about fancy models trained on lab data. It is about turning Monday morning judgment into something faster, broader, and less dependent on whatever happened to fit in one person’s head.