over the past month or so, i’ve struggled to find the time and headspace to write properly. i’ve been knee-deep in a new and exciting project, hopefully more on that soon. i’ve still been reading, but mostly things that didn’t feel like they belonged here. one exception was Shelby’s “The Sirens of Data Sales”, which deserves a thought of its own.
her argument, as i understand it, is roughly this: frontier AI labs have emerged as attractive buyers of biological data, particularly when compared with pharmaceutical companies. but selling to them can distract bio companies from their larger vision. there is also no guarantee that biology will remain a priority for the labs over the five-to-ten-year horizon that biotech companies must consider. data sales therefore make the most sense when they fund, or directly support, the company’s core mission.
it’s an interesting take. many bio companies have been starved for a meaningful way to participate in recent technological advances: to use AI in service of their vision (rather than merely add it to a pitch deck). Shelby does a good job of explaining why frontier labs are attractive, especially when contrasted with pharma’s long sales cycles, organizational inertia, and limited discretionary budgets.
still, calling data sales a siren call may be a little strong in some cases. the field is difficult, capital-intensive, and short on viable business models. for some companies, selling data to frontier labs may be a pragmatic way to fund the mission. there is real opportunity here, even if part of that opportunity is, bluntly, to “sell out.”
the harder, and hopefully much more rewarding, path is to hold on to the original raison d’être (as Shelby puts it). the real question is whether data sales finance that path or replace it. this is one reason i find VC perspectives useful: they offer a structurally different view from that of founders and operators.
one point i think the piece leaves underexplored is the commercial half-life of the data itself. a static dataset cannot necessarily be sold on the same terms forever. if it is incorporated into model training, or models are repeatedly optimized against an evaluation built from it, the data may eventually lose value as an independent benchmark. after one or two model generations, a lab may need fresh or more difficult data rather than the same dataset again.
the data may retain scientific or training value, while its ability to generate recurring revenue diminishes quickly. that seems like one more reason for caution: even a lucrative data deal may be temporary, while the strategic distraction it creates can last longer.