OpenAI Foundation launches Public Data for Health with $125M+ in grants

The OpenAI Foundation said it launched Public Data for Health, a program to fund high-quality scientific datasets that researchers and AI tools can use across many diseases. Initial support totals more than $125 million across nonprofits and universities, including OpenADMET drug-property data, CTD Commons regulatory archives from failed programs, and UNC’s personalized cancer-vaccine data work.

Same-day HEALTH primary that a major AI lab’s nonprofit parent is launching a second life-sciences program to fund open scientific datasets — more than $125 million in initial grants, with named first projects. This is a program launch, not a claim the data is already public or that any trial succeeded.

Goal, as the Foundation tells it: enable progress across life sciences and many diseases by funding the creation and preservation of high-quality scientific datasets made broadly available to researchers. File that job as the Foundation’s. This desk is not inventing a public catalog that is already online.

Initial tranche, still Foundation: supporting more than $125 million in grants across nonprofits and universities, spanning molecules to epidemiology to regulatory knowledge. File the more-than-$125-million floor and that span as the Foundation’s. This desk is not inventing a per-grantee table from this page.

OpenADMET, Foundation: open datasets, benchmarks, and blinded competitions to test whether AI models can predict how small molecules are absorbed and distributed in the body — their ADMET properties — aimed at making drug development more predictable and reducing clinical failure rates. The Foundation says 90% of drug candidates fail in clinical trials, often because those properties are hard to predict. James Fraser, PhD, OpenADMET Governing Board member and Professor and Chair of Bioengineering and Therapeutic Sciences at UCSF, is quoted on the same page. File the OpenADMET job and the 90% line as the Foundation’s. His quote stays in Sources. This desk did not run an ADMET contest.

CTD Commons, Foundation: preserve and publish Common Technical Document regulatory knowledge from failed or shelved drug programs so future teams can learn from precedent usually locked away. A CTD is the full FDA-style dossier for an investigational drug. Josh Morrison, leading organizer of CTD Commons and President of 1Day Sooner, is named on the Foundation page. File that preserve-and-publish job and those names as the Foundation’s. This desk did not acquire a CTD.

UNC Initiative for Generative Immunotherapy, short: the Foundation says UNC will create public multimodal data to improve personalized cancer vaccines — shots designed from a patient’s tumor genetics and proteins, also called neoantigen vaccines. UNC’s newsroom already printed a $40 million figure for that grant. That UNC-only pack is already filed as unc-openai-foundation-40m-cancer-vaccines. One line here is enough. File the Foundation’s UNC description as the Foundation’s, and the $40 million as UNC’s. This desk is not retelling that grant.

Strategy, Foundation: focus on connected data, scarce data, and direct data — datasets that follow biology across scales, preserve one-way-door observations that cannot be recreated, or measure states closest to clinical truth. File those three labels as the Foundation’s. This desk is not ranking future grant applications.

Accessibility, still Foundation: datasets should be broadly available while protecting patient privacy and consent. The Foundation encourages preprints and regular data sharing, not only a dump at the end. File that access-and-privacy line as the Foundation’s. This desk is not claiming the first datasets are already public, and it is not saying OpenAI Group PBC models are trained on them.

Plain English for the rest of the card: ADMET = how a drug is absorbed, distributed, metabolized, excreted, and whether it is toxic. CTD = Common Technical Document, the full FDA-style dossier for an investigational drug. Neoantigen vaccine = a personalized cancer shot designed from a patient’s tumor genetics and proteins. Open dataset = research data other labs can reuse. Multimodal data = more than one kind of measurement on the same biology, such as images plus lab reads.

PRIMARY here: OpenAI Foundation news post, 15 Sep 2026 — Tier A PRIMARY original record for the program. Clear to file. Discovery only: MIT Technology Review’s same-day wrap — Tier C corroboration, not the sole primary and not a source for extra grant amounts, endowment estimates, or IPO talk. The Public Data for Health launch, the second-program / AI-for-Alzheimer’s sequence, the more-than-$125-million initial tranche, OpenADMET, CTD Commons, the UNC generative-immunotherapy description, the connected / scarce / direct strategy, and the access-plus-privacy line are Foundation-attributed. The UNC $40 million stays UNC’s, via the already-filed unc-openai-foundation-40m-cancer-vaccines. NOT claimed: that the datasets are already public, clinical-trial success, FDA approvals, per-grantee dollar amounts beyond what the Foundation states, that OpenAI Group PBC models are trained on these datasets, a ChatGPT product launch, endowment or IPO figures, that this desk reviewed grant files, a stock tip, or investment advice. Distinct from the already-filed unc-openai-foundation-40m-cancer-vaccines, insilico-longevity-vaccines, openevidence-penn-medicine, and gates-foundation-1b-ai-access.