Snowflake says you can start AI without finishing migration, by querying open-format data where it sits.
I think the bill just moves.
Henrik Nielsen, Field CTO for Enterprise Architecture at Snowflake, lays out the obvious case fairly. For 10 to 20 years centralization was required because data in Oracle needed Oracle and data in Teradata needed Teradata. Apache Iceberg breaks that tie by decoupling storage from compute, so any compatible engine can read the same tables. A manufacturer can query ERP plus other systems as one asset, a retailer can join CRM, point of sale and supply chain data, with tables stored as Parquet in its own cloud storage and readable in both directions. Iceberg itself is Apache 2.0-licensed, community-driven and open source, per AWS, which is why the portability claim has teeth.
Governance is the second half of the pitch. Snowflake Horizon Catalog applies role-based access, masking, row access policies, lineage and access history inside and outside the platform. Horizon now extends row access and masking to external engines such as Apache Spark, Trino and PyIceberg, but that support is in public preview. Horizon plans each scan and hands the engine only allowed rows, with each call logged. For agents, Snowflake adds an agent identity that is separate from the user, with policies able to target agent sessions and actions attributed to both user and agent.
That preview status is the tell. For European firms facing GDPR, residency rules and the EU AI Act, Snowflake frames in-place governance with end to end lineage as the foundation. It is a fair point, and Nick Cheesman, CIO at NTT DATA Inc., backs the direction in the post, saying the work is about making data easier to access, govern and consume across a complex global estate without adding complexity. Still, no price, no latency cost for central policy checks, and no date for general availability are given. The model still needs Horizon to sit in the scan path for every caller, human, BI tool or agent. That is less a migration than a control-plane dependency, a familiar theme after Snowflake said tokens lie and context pays.
To its credit, Snowflake warns against a wild west of disconnected AI tools and tells teams to start with one focused pilot that can fund the next step. Good advice. If you want AI this quarter, pilot where the data lives, but budget for the governance integration Snowflake leaves in preview.
