Lakehouse architecture

Can Snowflake run on-premise or air-gapped? The answer, and what to use instead

No. Snowflake runs only in Snowflake-operated accounts on AWS, Azure and Google Cloud. There is no on-premises, bare-metal or air-gapped edition, and Virtual Private Snowflake is dedicated but still operated by Snowflake. Snowflake can query external tables on S3-compatible storage you run, but compute, metadata and results stay in Snowflake's cloud.

Published . Last updated

On this page

Snowflake made the data warehouse easy to adopt by removing infrastructure from the customer's hands entirely. That is also why it cannot follow your data into your own data centre. This guide sets out what Snowflake's deployment options do, where their limits sit for sovereignty, and how teams build a Snowflake-style experience on infrastructure they control.

Is there an on-premises version of Snowflake?

No. Snowflake is software-as-a-service by design. Compute (virtual warehouses), storage, metadata and the cloud services layer all run in Snowflake-operated accounts on AWS, Azure or Google Cloud. Snowflake has not released a self-hosted, bare-metal or disconnected edition.

Does Virtual Private Snowflake count as self-hosted?

No. Virtual Private Snowflake (VPS) is Snowflake's highest isolation tier: a dedicated environment with its own metadata store and compute, not shared with other accounts. It is still provisioned and operated by Snowflake in a public cloud. It answers "can we avoid sharing infrastructure with other customers?" It does not answer "can it run inside our perimeter?"

What about external tables on my own storage?

Snowflake can read data that sits in S3-compatible object storage you operate, including on-premises storage appliances that pass Snowflake's compatibility tests. The data at rest stays with you. But queries run on Snowflake compute, metadata about those tables lives in Snowflake, and results are returned through Snowflake's cloud. If your policy says data may not be processed outside your boundary, querying it from Snowflake does not satisfy that policy.

How do Snowflake's options compare for sovereignty?

OptionData at restComputeMetadata and controlAir-gapped?
Standard Snowflake accountSnowflake's cloudSnowflake's cloudSnowflakeNo
Virtual Private SnowflakeDedicated, Snowflake-operatedDedicated, Snowflake-operatedSnowflakeNo
External tables on your storageYour storageSnowflake's cloudSnowflakeNo
Iceberg tables with an external catalogYour cloud storageSnowflake's cloudSharedNo
DataNature on-premisesYour storageYour hardwareYour deploymentYes, if required

Which Snowflake features have an open, self-hostable equivalent?

Snowflake capabilityDataNature equivalentOpen technology
Virtual warehousesInteractive SQL and serving computeTrino, DN Warehouse
Native tables, Iceberg tablesLakehouse tablesApache Iceberg
Snowpipe and connectorsConnect and StreamHouseJDBC extraction; managed Kafka with CDC for the real-time hot path
Dynamic tables, tasksdbt models and DataNature Jobsdbt on Spark or Trino, scheduled job DAGs
Row access and masking policiesAtlas row filters and column masksPolicies enforced in Trino and the warehouse
Horizon catalogAtlas catalog and lineageDataHub-based catalog, dbt artifacts
Semantic views, Cortex AnalystAtlas semantic models and VoirYAML cubes over REST and MCP, governed agents
Snowpark Container ServicesApps on the DataNature PlatformKubernetes
Hybrid tablesLakebaseManaged PostgreSQL
Time travelIceberg snapshotsApache Iceberg time travel

How do masking policies translate?

Snowflake masking policies are SQL functions attached to columns. In DataNature, an admin chooses a mask type per column and per user or group in the Atlas Data Governance console: null, redact, show first four, show last four, hash, year only, or exempt. Hash masks are consistent, so a masked customer ID still joins across tables. Row access policies become row filters: a SQL predicate such as region = 'ON' applied to a user or group. Both take effect within about a minute and apply to the SQL workbench, JDBC clients, cubes and impersonating Superset connections.

What does migrating off Snowflake look like?

  1. Export tables to Parquet or Iceberg in object storage, or land source systems directly into Bronze through Connect.
  2. Translate SQL. Most ANSI SQL runs on Trino unchanged. Snowflake-specific functions (FLATTEN, VARIANT handling, QUALIFY usage, stored procedures) need rewriting; dbt projects move by changing the adapter target.
  3. Rebuild governance as Atlas grants, row filters and masks, scoped to specific schemas.
  4. Repoint BI. Superset, Power BI and Tableau connect to the lakehouse or warehouse; keep impersonation on so per-user rules apply.
  5. Reconcile and cut over domain by domain, comparing Gold metrics per period.

Frequently asked questions

Can Snowflake run in a private cloud I operate?

No. Snowflake only runs in Snowflake-operated accounts on supported public clouds.

Is Snowflake's Iceberg support a path to on-premises?

It makes your data portable, which helps, but Snowflake compute and services still run in Snowflake's cloud. Running the engines yourself is what moves processing on-premises.

Can DataNature and Snowflake share data?

Yes. Iceberg tables in shared object storage can be read by both, subject to your governance and residency rules.

Sources

See it on your own data.

Book a demo

Demos and quotes are handled by Dlytica, the company behind DataNature.