Comparison
Two ways to take data and AI all the way to production. This page compares graal and Databricks criterion by criterion: what each one publishes, in its own words, dated and sourced.
At a glance
graal
With you: your Kubernetes or OpenShift, on-premises or with the hosting provider of your choice. No third-party service is needed to run it.
Databricks
AWS, Azure and GCP. The control plane is in the Databricks account: “The control plane is located in the Databricks account, not your cloud account.”
graal
Two open source products, under the Apache 2.0 license: graal Data Platform and graal Agentic Platform.
Databricks
A service under contract: “Databricks retains all rights, title and interest in and to the Databricks Services”. Spark, Delta Lake, Iceberg, Unity Catalog and MLflow are open source projects.
graal
Annual vendor subscription per vCPU, unlimited users and agents, no usage-based billing.
Databricks
Usage-based, per second, in DBUs: “Only pay for the products you use at per second granularity.” Usage commitments give access to discounts.
graal
An MCP server drives the Data Platform, under a service account scoped to one project and with human approval. The Agentic Platform puts any agent to work on your processes.
Databricks
The Genie One MCP server has been generally available since 25 September 2026; the dbsql, sandbox and web_search servers require the Unity Gateway beta. Agent Bricks: “the Databricks agent developer platform”.
Criterion by criterion
Every Databricks cell links to its source and to the date we accessed it. Quotes stay in their original language.
| Criterion | graal | Databricks |
|---|---|---|
| Where the platform runs | Your Kubernetes or OpenShift: on-premises, in a private cloud or with the hosting provider of your choice. Installed with Helm, without cluster administration rights; offline installation is possible. | AWS, Azure and GCP. On AWS, classic compute runs in your own account: “For classic Databricks compute, the compute resources are in your AWS account in what is called the classic compute plane.” Serverless compute runs in the Databricks account. |
| Who operates the control plane | You do. No SaaS and no telemetry: no third-party service is needed to run it. Operations by graal Services is a separate service that you choose. | “The control plane includes the backend services that Databricks manages in your Databricks account. The control plane is located in the Databricks account, not your cloud account.” |
| Vendor, headquarters, governing law | GRAAL.SYSTEMS SAS, a French-law vendor, Paris. | Databricks, Inc., headquartered in San Francisco. For a customer domiciled in Europe, the contract (MCSA) sets Irish law and the courts of Dublin. |
| Software license | Open source, Apache 2.0 license, for the Data Platform and the Agentic Platform alike. | A platform under contract: “Databricks retains all rights, title and interest in and to the Databricks Services”. It builds on open source projects: “Apache Spark™, Delta Lake, Apache Iceberg™, Unity Catalog, MLflow, Delta Sharing and Redash”, several of them created by Databricks. |
| Open formats | Parquet and Apache Iceberg on your S3 storage, standard SQL with Trino; low-code pipelines exported as Pandas or PySpark code. | Delta Lake is the default format: “Delta Lake is the default format for all operations on Databricks.” Iceberg tables are supported and reachable through the Iceberg REST API of Unity Catalog. |
| Published pricing model | Annual vendor subscription per vCPU, unlimited users and agents, no usage-based billing. 8-week pilot; graal Services priced separately. | “Databricks offers you a pay-as-you-go approach with no up-front costs.” The unit is the DBU, “a normalized unit of processing power”. Storage and networking depend on the cloud you choose. Usage commitments give access to discounts. |
| Generative AI and agents | MCP server built into the Data Platform, one service account per project, per-agent policies, human approval, revocation, audit. LLM gateway and RAG over your documents. Agentic Platform: workflows of agents and people, any agent, measured guarantees. | “The Genie One MCP server is now generally available” (25 September 2026). The servers “system.ai.dbsql”, “system.ai.sandbox” and “system.ai.web_search” require the Unity Gateway beta. Each MCP in the system.ai schema is a Unity Catalog object, governed by grants and policies. Agent Bricks: “the Databricks agent developer platform” (page updated on 13 October 2026). |
A cell older than 90 days is re-checked or removed. Has something changed? Write to contact@graal.systems and we will correct it.
When to choose graal
graal installs on your Kubernetes or OpenShift, on-premises or with the hosting provider of your choice. The platform, the data and the traces stay in the infrastructure you picked.
The vendor subscription is computed per vCPU, with unlimited users and agents and no usage-based billing: the bill varies neither with the number of queries nor with the number of people or agents.
Both platforms are open source, under the Apache 2.0 license. Low-code pipelines export as Pandas or PySpark code, and the data stays in Parquet and Iceberg on your storage.
The Data Platform’s MCP server opens its tools under a service account scoped to one project, with human approval, immediate revocation and an audit log. The agents of the Agentic Platform operate it the same way, and every call is traced on both sides.
When Databricks suits you
Databricks operates the control plane and offers “serverless” compute: “In the serverless compute plane, Databricks compute resources run in a compute layer within your Databricks account.”
The price follows usage: “Only pay for the products you use at per second granularity.” Usage commitments give access to discounts and other benefits, including across several clouds.
Databricks is built on these projects, several of which were created there, and Delta Lake is its default format.
Agent Bricks is “the Databricks agent developer platform”, and each MCP in the system.ai schema is a Unity Catalog object, governed by grants and policies.
Moving from Databricks to graal
graal Services support the move: inventory of your jobs, cut-over order, team training.
Step 01
A Python script or a Spark job is dropped as is into a graal job. Spark runs on your Kubernetes, distributed when the size calls for it.
Step 02
Your exported notebooks reopen in Jupyter or VS Code, behind your SSO. Experiments and models go through an MLflow client.
Step 03
You export your tables as Parquet; a graal job rewrites them as Iceberg on your S3 storage, and Trino queries them in SQL.
Step 04
Chains, cron and triggers are recreated in graal’s orchestration, with retries, backfills and alerts.
Frequently asked questions
graal installs on your Kubernetes or OpenShift, on-premises or with the hosting provider of your choice, and covers the chain from source to model in production: connectors, low-code pipelines exported as code, orchestration and compute, Iceberg lakehouse and Trino SQL, notebooks, ML and generative AI, governance.
No. The two can coexist: graal takes the workloads whose data must stay with you, and the rest can stay elsewhere. The 8-week pilot compares on a real use case, with your data.
An annual vendor subscription per vCPU, with unlimited users and agents and no usage-based billing. The amount depends on your vCPU quota and the services you choose: you receive it after a scoping conversation.
Yes, through the Data Platform’s MCP server, under a service account scoped to one project. Sensitive actions wait for a manager’s approval, and every call is visible live and in the audit log.
The 8-week pilot puts graal to the test on your use case, in your infrastructure.