Each Snowflake and Databricks spent the 12 months racing to assist the identical open desk format, open catalog protocol, and cross-engine governance mannequin. Snowflake shipped Iceberg v3 to normal availability at its June Summit and rebuilt its Horizon Catalog on Apache Polaris for two-way Iceberg interoperability. Databricks pushed Managed, International, and v3 Iceberg by way of Unity Catalog and launched cross-engine entry controls enforced over the Iceberg REST APIs. Two platforms that compete on virtually every thing converged on one concept: the info, catalog, and governance need to be open, as a result of brokers want to achieve throughout techniques.
This convergence is vital and in consequence, the closed platform is shedding, and the businesses that constructed the closed platforms are those telling you so with their roadmaps.
Agentic analytics wants 4 issues, and the way lock-in blocks the final one
Strip the agentic AI hype down and an agent that solutions enterprise questions over your information wants 4 concrete issues:
- Ruled context, so it is aware of which numbers are reliable and who’s allowed to see them.
- Reusable semantics, so “income” means the identical factor whether or not the agent reads it or a dashboard does.
- Quick question entry, as a result of an agent that waits 30 seconds per query is ineffective in a dialog.
- Portability, so the identical information serves the mannequin you employ right now and the one you turn to subsequent quarter.
A single closed platform can provide the first three inside its personal partitions, however it could actually’t provide the fourth, and that decides who wins. Fashions change each few months so the lab with the very best mannequin in March is just not all the time the very best in September. In case your ruled, semantically wealthy information lives in a format just one platform can learn, each time you turn a mannequin, it turns into a migration. Additionally, open structure turns that migration right into a configuration change.
The information format now not owns the client
The clearest sign got here from analysts summarizing Snowflake’s personal Summit bulletins. One framed it bluntly and said that Snowflake is getting ready for a world the place it now not must personal the info format to personal the client relationship. From that perspective, Iceberg v3 assist is desk stakes. The market already moved to open codecs, so the battle shifted up the stack to context, governance, and id.
Each distributors now say the identical factor in another way. Snowflake describes a future the place metadata, lineage, id, and coverage journey with the agent slightly than staying locked contained in the platform the place the info began. Databricks markets Unity Catalog on “write as soon as, learn wherever” and bidirectional federation throughout Snowflake, Glue, and different catalogs. Learn these two positions facet by facet and the conclusion writes itself. The worth is now not the storage. The worth is ruled and transportable entry to information that lives in open codecs.
Why “open inside one vendor” is just not open
Each platforms now wrap open codecs in language that sounds absolutely open whereas holding the gravity inside their very own partitions. A managed Iceberg desk that solely performs nicely by way of one vendor’s engine is open in identify and closed in observe. Bidirectional federation that routes every thing again by way of a single catalog nonetheless concentrates management in a single place. The open desk format is important nevertheless it’s not adequate. What issues is whether or not you’ll be able to run your governance, semantics, and quick queries throughout engines with out one platform sitting in the midst of each path.
Can a second engine learn your information, apply the identical entry insurance policies, and return outcomes at interactive pace with out copying something? If the reply requires routing by way of the platform that saved the info, you purchased open codecs and saved the lock-in.
Image the setup that survives three mannequin generations. Your information sits in Iceberg by yourself object storage. An open catalog, Apache Polaris or one thing appropriate, tracks tables and enforces coverage by way of the Iceberg REST APIs that each critical engine now speaks. A semantic layer defines your metrics as soon as, so brokers and dashboards learn the identical definitions. Any engine and any AI agent connecting by way of a protocol like MCP, can attain that information beneath constant governance.
In that situation, switching fashions prices nothing structural and neither does including an engine as a result of the info by no means strikes. Governance doesn’t fragment throughout copies both. That is the design each Snowflake and Databricks now gesture towards, and it’s the design that open-first platforms had been constructed upon.
The open-first platforms bought there first
The platforms including interoperability in 2026 are reacting to a thesis that open-first distributors shipped years earlier. Apache Arrow, Apache Iceberg, and Apache Polaris didn’t come from the closed platforms. They got here from a contributor neighborhood that guess on open requirements earlier than the agentic second made the guess look apparent.
The rationale this issues is positioning, not branding. A platform designed round open requirements doesn’t need to stroll again lock-in to chase brokers. Its caveats are fewer by development: no proprietary storage emigrate off, no single catalog each question should traverse, no format that just one engine reads nicely. The closed platforms can copy the format and the protocol, but they can not simply copy the absence of gravity.
Wager on the structure, not the mannequin
The temptation is to pick the platform with the very best AI demo, however that’s the mistaken guess. Demos age in months and the mannequin you marry right now will get outclassed by subsequent 12 months. The price of that divorce relies upon fully on how open your information was whenever you signed up.
So, decide platforms by a unique query. Not “whose agent is smartest right now” however “how cheaply can I alter my thoughts.” Open codecs, open catalogs, transportable governance, and question entry that doesn’t rely upon one vendor’s engine all push that value towards zero. Closed platforms, nonetheless polished their AI, push it again up.
The distributors already voted with their roadmaps. Snowflake and Databricks spent 2026 making their walled gardens appear to be open fields, as a result of their prospects demanded information that AI can attain throughout techniques. The lesson is just not that these distributors grew to become open, however that open received, and even the giants needed to observe. To keep away from pricey errors, finest to construct for the structure that gave them no selection.
