Databricks intros OpenSharing, a new standard for sharing AI
The open source protocol modernizes collaboration by enabling teams to share AI assets such as AI models and agent skills across domains and with external partners.
Just as Databricks in 2021 developed an open standard that enabled enterprises to securely share data both internally and with external partners, the vendor has introduced OpenSharing so that organizations can share AI assets.
Five years ago, Databricks built Delta Sharing, a sub-project within the open source Delta Lake project, so that enterprises could share data -- without moving or duplicating it -- to fuel collaboration.
Launched on June 10 and now available on GitHub, OpenSharing, which is hosted by the Linux Foundation, is an extension of Delta Sharing. The new open-source standard allows organizations to further foster collaboration across platforms, departments and with partners by sharing AI models, agent skills with specialized knowledge and workflows, and unstructured data.
In addition, OpenSharing broadens collaborative efforts by adding support for platforms that connect to the Apache Iceberg REST Catalog, which enables sharing between a new set of organizations, and adds partnerships with on-premises storage partners to enable no-movement sharing of on-premises data and AI assets.
"OpenSharing marks a shift from simple data exchange to a unified, governed interface for the AI and data stack," William McKnight, president of McKnight Consulting, told TechTarget. "Beyond traditional tables … this framework provides a blueprint for studying and scaling how autonomous agents interact with distributed data. This could be quite significant for data sharing."
Stephen Catanzano, an analyst at Omdia, a division of Informa TechTarget, similarly called OpenSharing an important development.
"OpenSharing is a solid development in the AI infrastructure landscape," he told TechTarget. "What makes it particularly important is that it extends secure, zero-copy sharing beyond structured data to include agent skills and AI models -- assets that are becoming critical in the agentic era. Previously, organizations had no standardized way to share these AI components across platforms."
Based in San Francisco, Databricks was a pioneer of the data lakehouse format for storing data. Like many data management vendors, Databricks, which from its founding in 2013 included machine learning capabilities, in recent years has focused much of its product development on enabling customers to build generative AI and agentic AI capabilities.
Sharing AI
Collaboration, whether within a single department, across organizational domains or with third-party partners, is an important means of speeding innovation and improving productivity.
OpenSharing marks a shift from simple data exchange to a unified, governed interface for the AI and data stack. Beyond traditional tables … this framework provides a blueprint for studying and scaling how autonomous agents interact with distributed data. This could be quite significant for data sharing.
William McKnightPresident, McKnight Consulting
Historically, collaborative efforts were informed by data. As a result, particularly during and in the immediate aftermath of the COVID-19 pandemic, many data management and analytics vendors built collaboration capabilities into their platforms to enable workers in remote locations to work together in a virtual hub.
Now, as more enterprises build agents and other AI tools to assist employees and execute certain business processes, collaborative efforts include AI. Not only are collaborative projects built on data and data products such as reports and dashboards, but they include agents and other AI assets.
However, without a standard means of sharing AI assets, organizations are forced to piece together their own way of doing so, which must be done over and over again each time a department or business tries to collaborate with a department or business that doesn't use the exact same tools.
OpenSharing was developed to provide a standard, repeatable means of sharing AI to enable collaboration, according to Akram Chetibi, director of product management at Databricks.
"AI created a problem nobody had really solved yet," he told TechTarget. "When you look at how organizations are starting to share things across company boundaries -- not just data tables, but AI models, agent skills, prompts, and tools -- there was no standard way to do it. Everyone was cobbling together their own approach."
Given Databricks' experience developing Delta Sharing, OpenSharing was a logical extension, Chetibi continued.
"We'd already seen this problem play out with structured data before Delta Sharing existed, and we didn't want the AI world to repeat it," he said. "OpenSharing fixes that by giving organizations a single open protocol for publishing and consuming AI assets … regardless of which platform either side runs on."
Specific benefits of OpenSharing include the following:
An open protocol for publishing and sharing data and AI assets so files don't need to be manually copied for each collaborative initiative.
APIs for discovery, authorization and access irrespective of the platforms that different departments and organizations use for managing data and AI tools.
Support for Apache Iceberg APIs, expanding the reach of Delta Sharing to organizations that use Iceberg-native tools.
Integration with on-premises platforms and private clouds to provide enterprises electing to keep their AI operations in more secure environments than public clouds with the same collaborative capabilities as those using public clouds.
As agentic AI systems proliferate, the sharing requirements to collaborate are evolving away from just structured datasets to include new, complex assets, Catanzano noted. OpenSharing addresses some of the challenges that have risen as a result, including the multi-cloud and hybrid infrastructures across which AI assets are spread.
"OpenSharing addresses this new complexity by providing a unified protocol that works across these fragmented environments, enabling the kind of seamless AI collaboration that the current landscape demands but couldn't previously support," Catanzano said.
In addition, given that agents themselves are becoming some of the primary consumers of data, executors of workloads and collaborators across domains, a standardized way to share data and AI assets is beneficial, according to McKnight.
"These systems read data and the underlying metadata like model weights and execution code. OpenSharing treats this like a single, governed package," he said. "Furthermore, modern enterprises are running fast and consequently have many lakehouses and open data formats, so organizations need a vendor-neutral, open standard to share assets without the cost and lag of copying data."
Standardizing documents
Beyond OpenSharing, in other open source development news, the LF AI & Data Foundation revealed the formation of the DocLang Specification Working Group. The group was formed to develop DocLang, an open, AI-native document format to standardize preparing, exchanging and governing document data for AI systems.
Catanzano noted that most enterprise knowledge is stored in documents such as PDFs, Word files and presentation slides. Extracting such document data and operationalizing it for AI can be difficult and time-consuming, slowing AI initiatives. The DocLang working group, therefore, is addressing a significant bottleneck in AI adoption, according to Catanzano.
"The combination of DocLang -- the standard -- with the open-source processing toolkit Docling creates a complete stack for document AI, which could accelerate enterprise AI adoption by making document understanding more deterministic and interoperable across systems," he said. "This is particularly timely as agentic AI systems increasingly need to work with unstructured enterprise documents at scale."
McKnight, meanwhile, theorized that formation of a working group signals a rising emphasis on AI-native document data.
"The launch … is where an industry shifts from fragmentation toward fragmentation-killing collaboration," he said. "It's the beginning of a shift in the foundations to AI-native documents and interactions."
While both OpenSharing and the effort to build DocLang are individually valuable, they continue an ongoing trend of developing open standards to enable AI development, deployment and management.
The Model Context Protocol, an open source set of code released by Anthropic in November 2024 that standardizes how AI models connect to an organization's proprietary data sources, has become widely adopted. The Agent2Agent Protocol, introduced by Google Cloud in May 2025, provides a standard for agent interactions.
Building and managing agents and other AI tools using open-source capabilities is valuable because it allows enterprises to remain flexible, according to Chetibi. That's why Databricks elected to make OpenSharing open source rather than keep it a proprietary feature within the broader Databricks platform.
"Customers and partners don't want to collaborate with their data and AI assets while being confined to a single vendor proprietary ecosystem," Chetibi said. "It simply doesn't stick because it constrains innovation."
Eric Avidon is a senior news writer for Informa TechTarget and a journalist with more than three decades of experience. He covers analytics and data management.
