The Hidden Cost of Moving Data for AI

Architects and CIOs often face conflicting demands to deploy AI immediately while following a data-first playbook that requires migration, centralization, and governance before AI can begin. Historically, this meant physically moving data into a single system, which incurred significant costs and risks, as each transfer required proprietary compute tied to the source system. This approach, once standard, is no longer technically necessary but persists due to outdated assumptions about data interoperability and governance.

Open table formats, particularly Apache Iceberg, now decouple storage from compute, enabling any compatible engine to read data without copying or rebuilding pipelines. This allows organizations to query data across disparate systems—such as ERP, CRM, or supply chain platforms—as a unified asset, while leaving the data in place. The shift means AI workloads can start on existing data immediately, while modernization efforts proceed independently, preserving prior investments and accelerating AI timelines.

Traditional governance models required data to be centralized before consistent policies could be applied, creating bottlenecks that delayed AI deployment. Snowflake Horizon Catalog now extends governance policies—such as role-based access control and masking—to external engines like Apache Spark, Trino, and PyIceberg in public preview. These policies are enforced at query time, ensuring users and AI agents only access permitted data while maintaining a full audit trail of all interactions.

For European enterprises, this architecture aligns with regulatory demands under GDPR, data residency rules, and the EU AI Act, as it enables governance and auditing without forcing data consolidation into a single region or system. Snowflake’s approach allows AI to operate on data where it resides, supported by a unified governed layer spanning compute, catalog, and storage, eliminating the need for upfront replatforming.