Delta Sharing becomes OpenSharing: the open protocol now carries agent skills, AI models and unstructured data without copying

The object range of the open data-sharing protocol expands from tables to the assets AI teams depend on, including agent skills, AI models and unstructured data. Previously the open protocol centered on tables, while views, unstructured files, notebooks and registered ML models moved only through Databricks-to-Databricks sharing, so open cross-platform sharing could not reach these assets. At announcement Delta Sharing had more than 28,000 data recipients, with roughly a third of shares flowing across platforms through open connectors rather than staying inside Databricks-to-Databricks sharing.

Sharing stays zero-copy: providers do not move data, and the sharing server issues temporary, scoped credentials while sitting outside the data path. This model follows Delta Sharing's approach; what changes is that the specification itself now brings asset types beyond tables into the same sharing hierarchy. The OpenSharing specification is Apache 2.0 licensed on GitHub and developed as a project under the Linux Foundation AI & Data Foundation, currently at the sandbox stage.

Sharing reach also extends to the Iceberg ecosystem and to on-premises or private-cloud storage. Previously a table-centric open protocol could not directly serve organizations standardized on Iceberg tooling, nor govern data on premises or in a private cloud without migrating it first. OpenSharing adds support for Apache Iceberg REST Catalog compatible clients; through the Databricks Storage Ecosystem, storage vendors can run an OpenSharing server connected directly to Unity Catalog, MinIO's AIStor integration is generally available, and other storage partners are in preview or committed for later in 2026.

The demonstrated scope is two layers: the protocol specification, Apache 2.0 licensed on GitHub, and the Databricks enterprise implementation integrated with Unity Catalog, Databricks Marketplace and Clean Rooms; existing Delta Sharing clients and integrations remain compatible, and enabling OpenSharing on a Unity Catalog metastore is a configuration step rather than a migration.

Some capabilities were early stage at announcement: foreign Iceberg table sharing was in Public Preview with general availability described as coming soon, and email-based sharing was also listed as coming soon. Evaluation dataset and benchmark sharing, synthetic dataset sharing with provenance metadata, and a proposed precomputed vector index asset type are community proposals on the public GitHub roadmap, so timelines and final scope can change.