On June 10th, 2026, the Linux Foundation announced the OpenSharing Project — an open protocol contributed by Databricks that aims to do for AI assets what HTTP did for web documents: create a universal exchange layer that any platform can implement and any organization can use.
The Problem It Solves
Enterprise AI deployments in 2026 involve a sprawling mix of platforms: data lakehouses, model registries, vector stores, agent frameworks, and proprietary marketplaces from every major cloud provider. Moving a fine-tuned model from one organization's infrastructure to another's, or sharing a curated dataset across a joint venture, currently requires either point-to-point custom integrations or accepting the lock-in of a single vendor's tooling.
OpenSharing is a single protocol designed to cover all of it. It extends Databricks' existing Delta Sharing standard — already widely deployed for structured data exchange — to cover the full range of AI-era assets: agent skills, trained models, unstructured data volumes, and evaluation datasets.
How It Works
The protocol is built around a zero-copy sharing model. Rather than copying data between systems, OpenSharing issues access tokens that let the receiving party read directly from the source storage. This is the same architecture that made Delta Sharing efficient at scale, and extending it to model weights and agent artifacts means organizations can share a 70-billion-parameter model without physically transferring hundreds of gigabytes.
Cross-platform interoperability is built in from the start. OpenSharing supports both Delta Lake and Apache Iceberg as table formats, which covers the two dominant open lakehouse standards. Recipients using either format can consume shared assets without format conversion.
Security is handled through short-lived, scoped credentials rather than persistent API keys. The sharing server issues a credential tied to a specific asset and time window, which limits the blast radius of any compromise.
Why the Linux Foundation Matters Here
Databricks could have kept OpenSharing proprietary or hosted it under their own governance. Donating it to the Linux Foundation does two things: it provides a neutral home that competitors can trust enough to implement (Microsoft, Google, and AWS are all more likely to support a foundation-backed spec than a Databricks-owned one), and it opens the governance process to the broader community.
The Linux Foundation has a strong track record of hosting standards that become infrastructure: Kubernetes, OpenTelemetry, and ONNX all followed similar paths from single-company contribution to industry-wide adoption.
What Comes Next
The project is accepting founding members and the specification is already public on GitHub under the Linux Foundation's GitHub organization. Early reference implementations are available for Python and Java. If the trajectory follows the Delta Sharing pattern, expect major cloud providers to add native support within the next 12 to 18 months.
For anyone building agentic AI systems that need to move assets between organizations — whether for federated training, model marketplaces, or multi-party pipelines — OpenSharing is worth watching closely.