Databricks says the recipient relationship—not the asset type—should choose your sharing boundary

Databricks now has three overlapping ways to distribute data and AI assets: OpenSharing, Clean Rooms and Marketplace. Its new decision guide makes the useful distinction that teams should choose among them based on who is on the other side and what that party may see, not simply whether the asset is a table, model or agent. Databricks

Known recipient, readable asset: OpenSharing

OpenSharing gives a named recipient direct, read-only access to a live asset. Databricks says recipients do not need to run Databricks, and BI tools including Power BI and Tableau can read shared tables directly. The supported examples extend beyond Delta and Iceberg tables to views, volumes, notebooks, UDFs and AI models. Because the recipient can read the shared object, this is the broadest exposure model of the three; revocation is the operational control, not mutual approval of each computation. Databricks

For architects, that makes OpenSharing appropriate when the consumer is known and the provider is comfortable granting object-level read access. It removes per-consumer copy pipelines, but it does not turn an over-broad share into a privacy-preserving collaboration. Databricks

Known counterparties, hidden inputs: Clean Rooms

Clean Rooms target a different contract. Up to ten businesses can contribute data while keeping their full underlying datasets hidden; code runs only after every party approves it, and participants receive results rather than direct access to one another’s objects. Databricks describes parties sharing table names and columns, collaborating on a notebook and approving execution before it runs. Databricks

That boundary suits joint measurement, overlap analysis or cross-company model work where the value comes from combining inputs but no participant should receive the others’ source data. The practical review question is therefore not “can this notebook run?” but “does the approved output reveal more than the parties intended?” The platform’s approval gate controls execution; teams still need to review the information released by the result. Databricks

Unknown audience: Marketplace

Marketplace changes the counterparty model again. It is public and discoverable rather than addressed to a pre-existing partner. Databricks lists files, tables, models, notebooks, apps, SaaS software and MCP servers as publishable offerings, with examples that include Genie agents and MCP servers intended to plug into customer agents. Databricks

That means discovery, packaging and consumer evaluation become part of the security boundary. A Marketplace listing can solve distribution, but it should not inherit the trust assumptions of a bilateral OpenSharing recipient or an approval-gated Clean Room.

The clean deployment rule is simple: use OpenSharing when a known recipient may read the asset, Clean Rooms when known parties may approve computations but not inspect one another’s inputs, and Marketplace when the audience is not known in advance. The three can coexist, but substituting one for another silently changes who can discover, read or compute over the asset. Databricks