The OpenAI Foundation is putting money where the data gap is. It has launched Public Data for Health, a new initiative designed to break the bottleneck stifling medical AI progress. The flagship move? Funding Ruxandra Teslo’s proposal to harvest regulatory filings and safety data from bankrupt biotech firms. Teslo called these documents “biotech’s lost archive,” arguing they hold the key to training AI copilots for drug approval. The foundation awarded $500,000 to 1Day Sooner, an advocacy group Teslo advises, to prove this model works.
Money isn’t the only resource flowing here. The foundation also committed $40 million to the University of North Carolina, Chapel Hill, for novel cancer vaccine data and backed OpenAdmet, which runs drug effect prediction competitions. Morgan Levine, former VP at Altos Labs, notes that everyone recognizes data as the primary hurdle in applying AI to biology. Without high-quality scientific datasets, models remain blunt instruments rather than precise diagnostic tools. This grant round signals a shift from theoretical potential to practical infrastructure building.
The financial backdrop is staggering. With a 26% equity stake in OpenAI, the foundation could soon hold $250 billion in stock value, potentially surpassing the Gates Foundation as the world’s richest charity. Jacob Trefethen, a foundation executive, aims to distribute $1 billion by year-end. Yet the organization is still hiring key roles and refining its strategy. Its largest gift so far was $100 million to the Common Health Coalition for hepatitis C access. This scale allows for aggressive experimentation, but it also raises questions about oversight and priority-setting in such a young operation.
Acquiring corporate remnants is becoming a competitive sport. Google recently won a bid for Spirit Airlines’ data, including 100 million emails, sparking privacy concerns. Similarly, 1Day Sooner president Josh Morrison reports mixed results in bankruptcy auctions, securing two datasets from Lumen Bioscience while losing others. He estimates nonexclusive copies could cost just tens of thousands each. As AI firms race to feed their models, failed companies’ secrets may become the next great land grab. Whether this yields cures or just cleaner training sets remains to be seen.
