LillyPod AI factory runs 90% research at Lilly

Eli Lilly's LillyPod, announced in February, represents a significant bet on treating AI as core scientific infrastructure rather than a set of external models. Speaking at CoreWeave's Fully Connected 2026, Brian Lewis, head of machine learning engineering and advanced intelligence, and Smith outlined how the company built one AI operating system that connects data, models and tools into a single system. The approach prioritizes disciplined engineering over chasing every AI advance, with an emphasis on reproducibility, traceability, and integrating with tools scientists already use.

The AI factory's air-gapped, on-premises design allows Lilly to leverage decades of proprietary data, including information about millions of molecules, without internet connectivity. Smith noted that token costs weren't an issue given the on-premises setup, and that intelligent routing decides based on workload and GPU hour where to send a job. This control over models and data has enabled Lilly to save on token expenditures and offset a good portion of cloud model costs with agentic coding models.

Looking ahead, the outcome hinges on whether Lilly's scientists can translate scaled experiments into new medicines. Smith acknowledged that Lilly will not know which experiments improve the human condition until they do, but said the bet has been made. As governance and visibility in the industry mature, Lilly's experience with local compute as a controlled environment for experimentation may offer lessons for other enterprises weighing hybrid and local AI infrastructure.