The Linux Foundation has officially announced the launch of OpenSharing, a vendor-neutral protocol designed to facilitate the secure and standardized exchange of artificial intelligence assets between organizations. Contributed by Databricks, OpenSharing represents a significant milestone in the maturation of the global AI ecosystem, offering a technical solution to a problem that has historically required either prohibitively expensive custom integration work or reliance on restrictive, single-vendor marketplaces. By extending the capabilities of the widely adopted Delta Sharing protocol, OpenSharing aims to streamline how companies share not only raw data but also AI models, agent skills, and unstructured data across disparate platforms and organizational boundaries.
The introduction of OpenSharing comes at a critical juncture for the enterprise AI sector. As businesses move beyond experimental pilots and into large-scale production, the need for a collaborative framework that preserves security while enabling interoperability has become paramount. Currently, when two companies—such as a financial institution and a specialized fraud-detection vendor—wish to share an AI model, they are often forced into a "lock-in" scenario. They must either use the same cloud provider’s marketplace or spend weeks developing custom APIs and credentialing systems. OpenSharing seeks to eliminate these friction points by providing a common language for discovery and authorization.
The Foundation: Building on Delta Sharing’s Proven Success
The decision to base OpenSharing on the existing Delta Sharing protocol was a strategic move intended to capitalize on an established user base and a robust ecosystem of connectors. Delta Sharing is already utilized by thousands of enterprise customers, including industry leaders such as Stripe, SAP, the London Stock Exchange Group (LSEG), and Atlassian. This existing infrastructure allows OpenSharing to bypass the "cold start" problem that often plagues new technical standards.
According to Akram Chetibi, Director of Product Management at Databricks, the transition from data sharing to AI asset sharing was a natural evolution driven by customer demand. Enterprise partners who were already collaborating on tabular data began requesting similar mechanisms for sharing more complex assets, including generative AI models and the specialized "skills" required for autonomous agents. By leveraging the Delta Sharing model, OpenSharing inherits a proven security framework and a set of connectors that are already operational in high-stakes production environments.
The protocol also draws interest from major AI researchers and providers, including OpenAI. While OpenAI is currently positioned as an interested partner rather than a primary contributor, its involvement signals a broader industry consensus: the future of AI development depends on a standardized, secure method for authorizing access to high-value assets without the need for manual, one-off integrations.
Positioning OpenSharing Within the Open AI Stack
To understand the impact of OpenSharing, it is necessary to view it within the context of the Linux Foundation’s broader efforts to build a comprehensive, open-source AI stack. Over the past 18 months, the Foundation has systematically addressed various layers of AI governance and technical communication.
In May 2024, the Linux Foundation released OpenMDW 1.1, a permissive license framework specifically tailored for AI model distributions. In December 2023, it launched the Agentic AI Foundation (AAIF), which now governs the Model Context Protocol (MCP)—an open standard developed by Anthropic for connecting AI models to external tools and data sources. The AAIF also oversees OpenAI’s AGENTS.md convention and Block’s Goose agent framework.
OpenSharing sits at a distinct layer above these communication protocols. While MCP and the Agent-to-Agent (A2A) protocol define how agents talk to one another or interact with tools, OpenSharing provides the APIs for discovering and authorizing access to the stored assets themselves. It functions at the storage layer, vending temporary credentials for secure access to proprietary models or agent skills.
This hierarchy ensures that different technical challenges are handled by specialized protocols:
- The Licensing Layer (OpenMDW): Defines the legal rights and restrictions for using a model.
- The Access Layer (OpenSharing): Manages who can see and retrieve the asset from storage.
- The Communication Layer (MCP/A2A): Dictates how the model or agent operates and interacts with other systems once access is granted.
Technical Mechanics: The Digital Hotel Key Analogy
The technical architecture of OpenSharing is designed to provide "asymmetric openness." While the protocol itself is open and vendor-neutral, the assets it carries are often highly proprietary and sensitive. The protocol manages this through a system of signed, time-stamped tokens.
In practice, when a recipient requests access to an AI model hosted by a provider, the OpenSharing protocol issues a credential that functions much like a digital hotel key card. The key allows the recipient to enter a specific "room" (the storage location of the asset) for a strictly defined period. Once the time expires, the key is automatically invalidated. This mechanism allows the provider to maintain full control over their storage environment. They can see exactly which recipient accessed which asset and at what time, all without ever having to copy the asset into the recipient’s environment or provide long-lived API keys that could pose a security risk.
This approach addresses the "six-week integration" problem. Currently, an insurance firm wanting to test a vendor’s new model might spend over a month on security audits and custom API development. With OpenSharing, the vendor simply publishes the model to an access endpoint, and the firm’s data scientists can request access through a standardized interface, reducing the onboarding time from weeks to minutes.
Chronology of Development and Industry Adoption
The development of OpenSharing is part of an accelerated timeline of open-source AI standard-setting:
- Late 2021: Databricks introduces Delta Sharing as the industry’s first open protocol for secure data sharing.
- December 2023: The Linux Foundation establishes the Agentic AI Foundation to focus on agent interoperability.
- May 2024: OpenMDW 1.1 is released, providing a legal framework for AI model weights and distributions.
- June 10, 2024: The Linux Foundation officially announces the OpenSharing project, extending the Delta Sharing model to AI assets.
- June 2024: The MCP Developer Summit in Mumbai sees significant engagement from major corporations like UnitedHealth Group and HDFC Bank, signaling enterprise readiness for standardized AI protocols.
While the data-sharing components of the protocol are already mature, with connectors available for platforms like Oracle, SAP, and Power BI, the AI asset-sharing components are in the early adoption phase. However, the Linux Foundation reports strong interest from a wide array of vendors looking to avoid the high costs of maintaining proprietary sharing ecosystems.
Broader Implications and Strategic Analysis
The strategic significance of OpenSharing lies in its ability to prevent "vendor capture." In the early days of cloud computing, many enterprises found themselves locked into specific ecosystems because the cost of moving data or changing integrations was too high. By establishing a vendor-neutral protocol for AI assets, the Linux Foundation is ensuring that the same mistake is not repeated in the AI era.
Industry analysts point out that OpenSharing deliberately stops at the point of access. It does not attempt to enforce downstream lineage or monitor how a model is used once it has been retrieved. This distinction is vital for enterprise adoption, as it separates technical authorization from legal contracting. The protocol handles the "how" of access, while the licensing agreement (potentially using the OpenMDW framework) handles the "what" of usage. This separation of concerns allows legal and engineering teams to work in parallel rather than becoming bottlenecks for each other.
Furthermore, the "asymmetric openness" of OpenSharing reflects a realistic understanding of the modern economy. Companies are willing to use open standards for "plumbing"—the infrastructure of exchange—but they remain protective of the "water" flowing through those pipes—their proprietary AI IP. By providing a secure, open pipe for private assets, OpenSharing creates a middle ground that satisfies both the need for collaboration and the requirement for competitive secrecy.
As the AI stack continues to solidify, the pieces are falling into place for a more transparent and efficient market. With OpenSharing managing the distribution layer, the Agentic AI Foundation handling the communication layer, and OpenMDW providing the legal scaffolding, the barriers to cross-organizational AI collaboration are falling. For enterprise AI teams, the arrival of these protocols means that the focus can finally shift from the "how" of integration to the "what" of innovation. The rapid delivery of these layers suggests that a unified, open-source framework for enterprise AI is no longer a distant possibility, but a rapidly approaching reality.
