DataQube vs Databricks AI/BI Genie
Databricks offers genuinely strong network isolation. It does not offer a deployment without Databricks in it — the control plane is always operated by Databricks, and that is where Genie runs.
Updated 25 July 2026
The short answer
PrivateLink removes the public internet, not the dependency
Databricks splits into a control plane and a compute plane. Only one of them can run in your cloud account, and Genie is not in it.
The control plane is “the backend services that Databricks manages in your Databricks account” — the web application, job scheduler, Unity Catalog metadata services and model serving. It is always operated by Databricks, in Databricks' own cloud account.
Documentationdocs.databricks.com — Architecture overviewaccessed 2026-07-25
“Genie Agent responses and approved AI-generated comments are stored in the Databricks control plane database.”
Documentationlearn.microsoft.com — Databricks AI trust and safety, page dated 2026-07-17accessed 2026-07-25
Back-end PrivateLink “prevents your Databricks clusters from communicating with the control plane over the public internet”. The cluster must still reach that control plane; the traffic simply travels over a private backbone. A genuinely disconnected deployment is not a supported configuration.
Documentationdocs.databricks.com — PrivateLink conceptsaccessed 2026-07-25
Deployment
“On-premise” describes four different things
The term is used loosely enough to be meaningless in a procurement conversation. These are four distinct states, and the useful question in an evaluation is simple: if we cut the network cable, does it still answer questions?
SaaS
Vendor-operated, in the vendor's own cloud account.
Customer VPC
Runs in your cloud subscription, but a vendor-operated control plane is still required. Private networking removes the public internet, not the dependency.
True on-premise
Runs on hardware you own and operate.
Air-gapped
No outbound path exists at all — including for licensing, telemetry, updates and model inference.
Query execution can be placed in your own VPC. Genie's orchestration, model calls and conversation state cannot — they run in the Databricks-operated control plane regardless of warehouse type.
Pricing
Two meters, and one of them is currently unpublished
Genie bills LLM consumption separately from the warehouse compute that executes the SQL it generates — and the introductory pricing period has now closed.
Genie One and Genie Agents were free only through 31 July 2026. Pricing after that window is not published. Any figure quoted for Genie today should be confirmed against a current order form.
Contractdatabricks.com — Pricing; Genie release notesaccessed 2026-07-25
The free allowance is stated inconsistently across Databricks' own sources — “$10 free for every user every month” in the announcement against $10.50 on the pricing page.
Contractdatabricks.com — Pricing; Genie announcementaccessed 2026-07-25
Warehouse compute is billed separately and is explicitly excluded from Genie budgets, so the cost of actually running generated SQL sits outside any Genie figure. Service principals receive no free allowance at all.
Contractdocs.databricks.com — Genie billingaccessed 2026-07-25
DataQube licenses by concurrent active session, with model inference on hardware you already own. There is no second meter for the compute that answers the question.
Provenance
Where the conversation is stored
Databricks has a mature governance stack in Unity Catalog. The question for a regulated buyer is narrower: which side of the boundary the analytical record sits on.
DataQube
- The thread is an append-only record held entirely inside your perimeter — conversation, generated query, permission decisions and artefacts.
- Compaction appends an event to the log. It never rewrites what was recorded.
- A notebook is a projection of a thread. Reopening it shows what happened, not a fresh run against today's data.
- Workflows are rerunnable artefacts with parameters and run history, appended back to the thread.
Databricks
- Genie Agent responses and approved AI-generated comments are stored in the Databricks control plane database, by Databricks' own documentation.
- Unity Catalog lineage is genuinely strong for data and models. It is not a per-answer record binding a figure to the query, the rows, and the permission decision behind it.
- The only official Databricks use of “air-gapped” refers to running deployment bundles from a pre-downloaded Docker image — a CLI tooling scenario, not the platform running offline.
Fair comparison
Where Databricks is the better choice
Databricks is a serious platform and the closest genuine competitor for the analytical use case. There are several situations where it is clearly the right answer.
Choose Databricks when
- You already run a lakehouse on Databricks and Unity Catalog is your governance system of record.
- You need the surrounding platform — pipelines, ML training, model serving, streaming — not only a conversational analyst.
- Network isolation to a customer-managed VPC with PrivateLink satisfies your regulator, and a vendor-operated control plane is acceptable.
- Your data engineering team wants to build on the platform rather than adopt a finished product.
Choose DataQube when
- A vendor-operated control plane is not acceptable, and “no route out” has to be literally true.
- The analytical record has to stay inside your perimeter, not in a vendor's database.
- You want a finished analyst experience rather than components to assemble.
- Cost has to be forecastable a year ahead, without a second meter for query compute.
Questions we get in evaluations
Can Databricks Genie run air-gapped?
No. Genie's orchestration, model calls and conversation state run in the Databricks-operated control plane regardless of warehouse type. Customer-managed VPC and PrivateLink remove the public internet, but the cluster must still reach that control plane, so a genuinely disconnected deployment is not a supported configuration.
Does a customer-managed VPC mean Databricks runs on our infrastructure?
Partly. The classic compute plane runs in your cloud account, and Databricks documents that customer VPCs have no open ports and no public IP addresses. The control plane — including Genie — remains in Databricks' own account.
Databricks announced on-premises support in June 2026. Does that change this?
That announcement covers a storage partner ecosystem allowing compute to read on-premises datasets without copying files. The Databricks control plane, and Genie with it, continues to run in Databricks' cloud. Read the announcement carefully before treating it as an on-premise deployment option.
Can DataQube query data held in a lakehouse?
Yes. DataQube connects to approved sources as governed connections. What changes is where the orchestration, the model inference and the durable record sit.