Are foundation models the new moat?
Back in April, Tempus AI announced a strategic agreement with the pharma company AstraZeneca and a startup called Pathos, to build a Biology Foundation model focused on oncology. While the announcement is appropriately sparse on details, it does mention one important number: $200 million, which is what Tempus is getting from the deal. So this week, I want to speculate about what that number might mean for the future role of foundation models in pharma/biotech business models.
But first, a quick note: I've recently gotten back into working with biotech/pharma companies, particularly on applications of AI. As part of this, I'm partnering with companies that can provide implementation teams and complementary expertise. If you’re interested in exploring a project, or need help finding the right implementation team for an existing project, send me an email at jesse@merelogic.net.
The first of these partners is Arrayo, who provide teams and individuals specializing in transforming scientific innovation into scalable digital solutions. I started working with Arrayo when I was at Cellarity, based on a recommendation from a contact at Moderna. I’ve had many good experiences with them, so I’m excited to kick off this partnership.
The idea of a “moat” is that for a company to be successful, it needs to have something that a competitor can’t easily replicate. When founders and CEOs and VCs and board members talk about strategy, this is one of the things they look for.
The conventional wisdom in biotech/pharma has been that a large dataset, collected over years, makes a pretty good moat. Whether you’re running strategically planned experiments or carefully curating publicly available data, a competitor would need to spend just as much time and money to replicate it. So if you get a head start, you can probably keep it.
Tempus has done exactly this by providing genetic tests to cancer patients and cancer clinics for about a decade. Since 2015, they’ve been compiling the results from these tests into a massive research dataset that will presumably be the basis for this $200 million foundation model.
So Tempus has played this game about as well as possible. But still, the dataset moat idea kind of reminds me of underpants gnomes: Phase 1: Collect data. Phase 2: ??? Phase 3: Profit. Some companies that use this strategy may know what their Phase 2 is. But I suspect that quite a few are hoping to figure it out later.
By signing this deal, Tempus has figured out their Phase 2, which is to sell (rent?) the data in the form of a foundation model. Now AstraZeneca and Pathos have to figure out their own.
The $200 million attached to the deal suggests that all three parties consider a foundation model trained on data to be worth a lot more than just the data.
One possible theory for why is that the extra work involved in building a foundation model makes it a better moat. But I don’t think this works. The software side of a model is relatively easy to replicate, and training is relatively fast, given the data. So a foundation model isn’t any more of a moat than the data itself.
Another possible theory is that a foundation model solves the gnomes problem by connecting the dots between data and profit. But the foundation model on its own is not a Phase 2. For a drug discovery company, a Phase 2 has to involve getting a drug to market, or at least to clinical trials. That’s still a lot of dots to be connected.
So AstraZeneca is making a huge bet by investing in this model. The price tag suggests that they think it will be easier to turn a foundation model into Phase 2, and thus Phase 3: Profit, than just the data itself. But they still may not know exactly how they’re going to do it. (The announcement mentions target identification and alludes to other uses, but that’s it.)
In other words, turning your proprietary dataset moat into a foundation model can give you more options for figuring out Phase 2. But it doesn’t guarantee you won’t turn into underpants gnomes.
Thanks for reading Scaling Biotech! In between posts, I’m working on a guide to evaluating the impact, feasibility and cost of different use cases for AI in drug discovery. If you’d like to read an early copy when it’s ready, send me an email at jesse@merelogic.net or fill out the form at merelogic.net.