Open isn't the same as portable.
Every vendor calls themselves open in 2026. The difference is what they actually let you take with you. We measure portability by what survives a switch — not by what's in the marketing FAQ.
Same manifest, three clouds.
One deployment description. Apply it to an AWS cluster, an Azure cluster, a GCP cluster, or a neo-cloud cluster. The compute, the catalog, the policy plane, the agent runtime — all identical. No cloud-specific application code. Switching clouds is a Helm value, not a rewrite.
Your data, your storage, your keys.
Tables live in your S3 / Azure Blob / GCS, encrypted with your KMS keys — fail-closed by design and verified in our live environment down to the key ARN on the objects, with writes refused rather than landing unencrypted if your key can't be honored. Open table format any compliant engine can read. We can't hold your data hostage, because we never hold it.
Open catalog, no proprietary dialect.
We speak the open Iceberg REST catalog spec natively. Snowflake's Open Catalog is real but billed against. Databricks Unity Catalog OSS publishes APIs but the catalog server people actually run is forked. We have no proprietary catalog dialect to defend — your metadata is portable on day one.
Swappable everything.
Query engine, AI model router, agent runtime, deployment substrate, vector index, lineage emitter — each replaceable behind an interface, none structurally tied to a vendor. Here, 'swap in your own X' is a configuration change, not a roadmap conversation.
What 'open' means at each platform.
Seven dimensions a data architect can verify in an hour by reading our docs and the incumbents'. We will not claim a row we cannot demonstrate end-to-end on a customer call.
| Dimension | Eisberg | Snowflake | Databricks |
|---|---|---|---|
| Same manifest runs on AWS / Azure / GCP | Yes — one Helm chart, three clouds | No — each cloud is a separate account | No — workspaces are per-cloud |
| Customer-owned object storage | Yes — your S3 / ABFS / GCS, your KMS keys | Iceberg Tables in your bucket; Native Tables in theirs | Iceberg via Unity now; Delta legacy in theirs |
| Open table format with no proprietary extensions | Apache Iceberg — every byte readable by any compliant engine | Iceberg + their managed catalog (billed) | Iceberg + Delta + Unity (open API, closed server) |
| Swappable query engine | Multi-engine router — embedded + distributed today, accelerator lanes composable | Single engine. Period. | Photon only, in their runtime |
| Swappable AI model | Yes — any provider, any model, any region | Cortex models within their perimeter | Mosaic AI models within their workspace |
| Data records stay out of the vendor control plane | Yes — tables + Iceberg metadata live in your bucket; our plane holds policy + catalog pointers | No — native tables live inside their account | Partial — Delta legacy in their managed storage |
| Export tomorrow with one command | Iceberg tables already in your bucket — nothing to extract | Iceberg tables yes; Native Tables require unload | Unity Iceberg yes; Delta requires conversion |
Open-format pricing is a margin problem, not a roadmap problem.
Both major incumbents have moved toward Iceberg under competitive pressure. Neither can credibly match true portability without cannibalizing their own consumption model.
Snowflake
The catalog is the lock.
Iceberg Tables solve format portability but only when you pay for the Snowflake-managed catalog and the Snowflake-priced compute that reads it. The day you bring your own engine, you stop paying the warehouse meter — and the warehouse meter is the business. Genuine open catalog cannibalizes the consumption model that funds the company.
Databricks
The runtime is the lock.
Tabular got acquired, Unity Catalog now reads Iceberg, the framing is "open lakehouse." But the actually-shipped Unity server people run in production is forked closed, and the Photon engine that justifies the platform pricing is single- vendor. Genuine portability would mean a customer running Spark + their own catalog server against their own Iceberg — at which point they no longer need a Databricks workspace.
Want the deployment manifest?
One set of Kubernetes manifests, written cloud-agnostic by design — no cloud-specific application code anywhere in the stack. We share the charts under NDA on every architecture call — read them before the demo, verify the claim before the contract.