Linux Foundation Launches OpenSharing to Standardize AI Asset and Data Exchange

The Linux Foundation announced the OpenSharing Project on June 10, 2026, adding another entry to its growing stack of AI governance and infrastructure efforts. The pitch is straightforward: build an open, vendor-neutral protocol for exchanging AI assets and data across different organizations and different platforms, so that sharing a model or a dataset doesn't require committing to a single vendor's marketplace.

OpenSharing isn't a clean-room design. Databricks contributed its existing Delta Sharing protocol as the foundation, which is a meaningful head start. Delta Sharing has been around for a few years as Databricks' answer to the problem of moving large datasets between parties without copying them into proprietary silos first. By handing the protocol to the Linux Foundation, Databricks is betting that broad adoption matters more than keeping it as a differentiator, which is the same calculation that drove projects like Delta Lake into open governance.

What the protocol actually moves

The interesting expansion here is scope. Delta Sharing started as a data-sharing protocol focused on analytics tables. OpenSharing broadens that to cover what the announcement calls AI assets: model weights, agent skills, and other unstructured data, alongside the structured tabular data Delta Sharing already handled.

That matters because the way you ship a 700GB model checkpoint is not the way you ship a Parquet dataset, and it's definitely not the way you ship an agent skill definition. A protocol that wants to cover all three needs to handle:

Asset type	Typical size	Access pattern
Tabular data	GB to TB	Range reads, predicate pushdown
Model weights	GB to hundreds of GB	Bulk transfer, integrity verification
Agent skills	KB to MB	Frequent small fetches, versioning
Unstructured blobs	Variable	Streaming, partial reads

Delta Sharing's core mechanism is a REST protocol that hands out short-lived, pre-signed URLs pointing directly at object storage. The recipient never talks to the provider's compute layer for the actual bytes. They get a credential, then pull data straight from S3, GCS, Azure Blob, or whatever backs the share. That design is what makes it cheap to operate, since the provider isn't proxying terabytes through its own servers, and it's the same design that should scale reasonably well to model artifacts.

Open table formats, plural

One detail worth pulling out: OpenSharing supports multiple open table formats rather than locking to one. In practice that means it isn't tied exclusively to Delta Lake. The format wars between Delta, Apache Iceberg, and Apache Hudi have been one of the more contentious corners of the data ecosystem, and a sharing protocol that picks a single winner would have limited reach. Keeping it format-agnostic is the right call if the goal is genuinely cross-organization exchange.

{{IMAGE:2}}

Why eliminate proprietary marketplaces

The stated goal is to kill off proprietary marketplaces and custom integrators. Today, if you want to consume a commercial dataset or a hosted model, you frequently go through a vendor-specific exchange that dictates the format, the authentication, and the billing. Every new provider relationship means another custom integration. An open protocol collapses that into one client implementation that talks to any compliant provider.

For anyone running their own infrastructure, this is the part that actually changes the math. If OpenSharing endpoints become common, a self-hosted setup can consume external models and datasets through one standard client instead of maintaining a pile of vendor SDKs. The same pre-signed URL approach means you pull bytes directly into your own storage, which keeps egress predictable and avoids routing large transfers through an intermediary you don't control.

Where this fits in the Linux Foundation's AI portfolio

The Linux Foundation has been steadily collecting AI projects, from PyTorch governance to the various efforts under the LF AI & Data umbrella. OpenSharing slots in as the transport and interchange layer, the plumbing that sits underneath model registries and data catalogs. It's less glamorous than a new model architecture, but interoperability plumbing tends to outlast the flashier components above it.

The project is early. A contributed protocol and a press release are a starting point, not a thriving ecosystem, and the real test is whether providers beyond Databricks ship OpenSharing endpoints and whether client implementations show up across the major data and ML platforms. The full details are in the Linux Foundation's announcement, and the project now lives at OpenSharing.io. If you already run Delta Sharing, the practical question is how much of your existing setup carries forward unchanged, and the answer should become clear as the project publishes its governance and roadmap.