Both Snowflake and Databricks spent a year racing to support the same open table format, open catalog protocol, and multi-engine control model. Snowflake sent Iceberg v3 into general availability at its June Conference and rebuilt the Horizon Catalog on Apache Polaris to use Iceberg’s two modes. Databricks has pushed Managed, Foreign, and v3 Iceberg through the Unity Catalog and introduced cross-engine access controls enforced through the Iceberg REST APIs. The two platforms that compete in almost everything converge in one vision: data, catalog, and management should be open, because agents need access to all systems.
This integration is important and as a result, the closed platform is lost, and the companies that build closed platforms are the ones who tell you so with their roadmaps.
Agenttic statistics require four things, and how the lock blocks the last one
Strip the AI hype to the ground and an agent that answers business questions with your data needs four concrete things:
- It’s a controlled environment, so it knows which numbers are trustworthy and who is allowed to see them.
- It’s reusable semantics, so “revenue” means the same thing whether an agent reads it or a dashboard reads it.
- Quick access to the question, because the agent who waits 30 seconds for each question is useless in the conversation.
- Portability, so the same data is provided for the model you use today and the one you switch to next quarter.
One closed platform can give you the first three within its walls, but it cannot give you the fourth, and that determines who wins. Models change every few months so the lab with the best model in March is not always the best in September. If your dominant, mathematically rich data resides in only one format that the platform can read, every time you change the model, it becomes a migration. And, open architecture turns that migration into a configuration change.
The data format is no longer owned by the customer
The clearest signal came from commentators summarizing Snowflake’s Summit announcements. One put it bluntly and said that Snowflake is preparing for a world where you no longer need to own the data format to own the customer relationship. In that sense, support for Iceberg v3 is table stakes. The market has already moved to open formats, so the fight has shifted the stack to context, management, and ownership.
Both sellers are now saying the same thing differently. Snowflake describes a future where metadata, lineage, ownership, and policy move with the agent rather than remaining locked within the platform where the data originated. Databricks markets the Unity Catalog as a “write once, read anywhere” and bidirectional federation across Snowflake, Glue, and other catalogs. Read those two positions together and the conclusion writes itself. Value is no longer storage. Value is dominated by physical access to data that resides in open formats.
Why is “opening within the same seller” not open
Both platforms now wrap open formats in a language that feels fully open while maintaining gravity within their walls. A managed Iceberg table that only works well with a single vendor engine is open in name and closed in practice. A two-way alliance that brings everything back into one catalog still centralizes control. An open table format is necessary but not sufficient. What matters is whether you can apply your governance, semantics, and quick queries across engines without a single platform sitting in the middle of it all.
Can a second engine read your data, apply the same access policies, and return results at interactive speeds without copying anything? If the answer requires a route to a platform that has stored data, buy open formats and keep the lock.
Consider the heavy-duty setup for three generations of models. Your data is sitting on Iceberg in your object storage. Open Catalog, Apache Polaris or equivalent, tracks tables and enforces policy through the Iceberg REST APIs that every serious engine is talking about now. The semantic layer defines your metrics once, so agents and dashboards learn the same definitions. Any engine or any AI agent that connects through a protocol like MCP, can access that data under consistent governance.
In that case, changing models is structurally inexpensive and doesn’t mean adding an engine because the data doesn’t move. Management does not differ between copies either. This is the design that both Snowflake and Databricks now reflect, and it’s the design that the first platforms were built on.
The first open fields got there first
The platforms that add collaboration in 2026 are responding to the idea that the first open sellers were sent years ago. Apache Arrow, Apache Iceberg, and Apache Polaris did not appear on closed platforms. They came from a donor community that bet on open rates before the agency minute made the bet seem obvious.
The reason this is important is to replace, not to seal. A platform designed with open standards doesn’t need to be locked down to chase agents. Its caveats are few in design: no proprietary storage to migrate, no single catalog all queries must pass through, no format readable by only one engine. Closed platforms can copy the format and protocol, however they cannot easily copy the absence of gravity.
Bet on the architecture, not the model
The temptation is to choose the platform with the best AI demo, but that’s the wrong bet. Demos grow by months and the model you marry today will pass the standard next year. The cost of that split depends entirely on how open your data was when you signed up.
Therefore, judge the platforms with a different question. Not “whose agent is the smartest today” but “How can I change my mind at a low cost.” Open formats, open catalogs, physical governance, and query access independent of a single vendor engine all push those costs down to zero. Closed platforms, however polish their AI, push it to the top.
Vendors have already voted on their roadmaps. Snowflake and Databricks spent 2026 making their walled gardens look like open fields, because their customers wanted data that AI could access from every system. The lesson is not that those vendors opened, but that what was opened won, and the giants had to follow. To avoid costly mistakes, it’s best to build structures that don’t offer choices.
