Nvidia and Eli Lilly partner to build a $1 billion AI drug discovery lab in San Francisco

The failure dataset argument is genuinely interesting, but it conflates two different problems. Knowing why a molecule failed in screening is valuable for discriminative tools, the ones that filter known chemical space. It does less work for a generative platform that is writing sequence space evolution never sampled, because the failure modes in that territory are structurally different from anything in Lilly's 3 million compound graveyard.

The moat question I kept returning to when writing about the Profluent deal at https://www.onhealthcare.tech/p/profluents-225b-lilly-deal-and-why?utm_source=x&utm_medium=reply&utm_content=2071606615682449636&utm_campaign=profluents-225b-lilly-deal-and-why is not who has the largest historical failure library but who is running the tightest closed-loop pipeline, design to synthesis to wet-lab test to retrain. That loop generates proprietary signal on non-natural proteins that no historical dataset contains, and it compounds with each cycle in a way Lilly's archive structurally cannot.

Proprietary failure data is a discriminative moat. Closed-loop synthesis-test-retrain is a generative moat. They are not the same asset.

The deeper question the Lilly-Profluent deal raises is whether big pharma is starting to recognize that distinction by moving toward platform-fee structures rather than per-asset milestones. If Ricks believes historical failure data is the edge, the deal structure his team signed suggests someone inside Lilly is betting on something else entirely.