Databricks DBU Pricing: Every Rate, Every Workload Type
A complete reference for understanding Databricks Units. What a DBU actually measures, how rates differ by compute type, instance-to-DBU mapping, and step-by-step examples showing how to calculate your monthly cost.
What Is a DBU?
A Databricks Unit (DBU) is a normalized measure of processing capability. It abstracts away hardware differences so Databricks can price compute consistently across instance types and cloud providers. Think of DBUs as compute currency: when you run a cluster, each node consumes a certain number of DBUs per hour based on its size. Your Databricks platform bill is the total DBUs consumed multiplied by the per-DBU rate for your workload type.
The key insight is that DBU charges are only the Databricks portion of your bill. You also pay your cloud provider (AWS, Azure, or GCP) for the underlying virtual machines, storage, and network transfer. The DBU charge is the fee for the Databricks platform layer: managed Spark, Delta Lake, Unity Catalog, notebooks, workflows, and SQL warehouses.
The Two-Layer Formula
Total Cost = (DBUs consumed x DBU rate) + Cloud infrastructure cost
Complete DBU Rate Reference
| Compute Type | Rate/DBU | Description | Includes Compute? |
|---|---|---|---|
| Jobs Light Compute | $0.07 | Lightweight automated batch jobs | No |
| Jobs Compute | $0.15 | Standard batch and scheduled pipelines | No |
| Delta Live Tables Core | $0.20 | Declarative ETL with core features | No |
| Delta Live Tables Pro | $0.25 | ETL with CDC, expectations, advanced monitoring | No |
| All-Purpose Compute | $0.40 | Interactive notebooks and development | No |
| SQL Classic | $0.22 | Classic SQL warehouse endpoints | No |
| SQL Pro | $0.55 | Advanced SQL with query profiling | No |
| SQL Serverless | $0.70 | Fully managed SQL, instant startup | Yes |
| Model Training | $0.65 | ML model training workloads | No |
| Model Serving | $0.07 | Real-time model inference endpoints | Yes |
AWS Premium tier rates shown. Azure rates are 10-30% higher for compute-intensive types. Verify current rates at databricks.com.
How DBUs Are Consumed
Each cloud instance type has a defined DBU consumption rate. Larger instances consume more DBUs per hour. The rate depends on the instance family and size, not on what your code does. A node consuming 2 DBU/hr burns 2 DBU/hr whether it is processing data or sitting idle.
| Instance Type (AWS) | vCPUs | Memory | DBU/hr | Use Case |
|---|---|---|---|---|
| m5.xlarge | 4 | 16 GB | 0.75 | Light workloads, dev |
| m5.2xlarge | 8 | 32 GB | 1.50 | General ETL |
| i3.xlarge | 4 | 30.5 GB | 1.00 | Storage-optimized |
| i3.2xlarge | 8 | 61 GB | 2.00 | Default production |
| i3.4xlarge | 16 | 122 GB | 4.00 | Large pipelines |
| i3.8xlarge | 32 | 244 GB | 8.00 | Heavy processing |
| r5.xlarge | 4 | 32 GB | 1.00 | Memory-intensive |
| r5.2xlarge | 8 | 64 GB | 2.00 | Large datasets in memory |
| p3.2xlarge (GPU) | 8 | 61 GB | 5.50 | ML model training |
Compute Type Deep Dives
Jobs Compute ($0.15/DBU)▼
Jobs Compute is the workhorse of most Databricks deployments. It runs non-interactive batch jobs: ETL pipelines, data transformations, scheduled notebooks, and orchestrated workflows. At $0.15/DBU, it is the second cheapest compute type (after Jobs Light at $0.07/DBU).
Jobs clusters start when a job begins and automatically terminate when the job finishes. There is no idle cost. This makes Jobs Compute the most cost-efficient option for production workloads that do not require interactive access.
Best for: Production ETL, scheduled data pipelines, batch ML inference, dbt jobs.
All-Purpose Compute ($0.40/DBU)▼
All-Purpose Compute powers interactive notebooks where data engineers and scientists explore data, develop pipelines, and run ad-hoc queries. The cluster stays running between cell executions, which is why the rate is 2.7x higher than Jobs Compute.
The major cost risk with All-Purpose Compute is idle time. A developer starts a cluster in the morning, runs a few notebooks, goes to lunch, and the cluster burns compute for 2 hours doing nothing. Always set auto-termination to 10 to 15 minutes.
Best for: Development, data exploration, ad-hoc analysis, notebook prototyping. Never use for production jobs.
SQL Warehouses ($0.22-$0.70/DBU)▼
SQL warehouses come in three flavors: Classic ($0.22/DBU), Pro ($0.55/DBU), and Serverless ($0.70/DBU). All provide JDBC/ODBC endpoints for BI tools and SQL clients. The difference is in features and management overhead.
Classic: Manual cluster management, cheapest per-DBU rate but requires infrastructure configuration. Pro: Adds query profiling, intelligent workload management, and Photon acceleration. Serverless: No infrastructure to manage, starts in under 10 seconds, and includes compute cost in the DBU rate.
Best for: BI dashboards, SQL analytics, reporting. Choose Pro for medium teams, Serverless for bursty usage.
Delta Live Tables ($0.20-$0.25/DBU)▼
DLT is a declarative ETL framework where you define the desired state of your data and Databricks manages the pipeline execution. Core ($0.20/DBU) handles standard ETL. Pro ($0.25/DBU) adds change data capture (CDC), data quality expectations, and advanced monitoring.
DLT pricing is on top of the underlying compute. A DLT pipeline running on a Jobs cluster pays both the DLT DBU rate and the Jobs Compute DBU rate for the same compute hours. This effectively doubles the per-DBU cost compared to running equivalent Spark code directly.
Best for: Teams that value pipeline reliability and automation over raw cost. DLT reduces engineering time at the expense of higher per-hour costs.
Model Training & Serving ($0.65 / $0.07)▼
Model Training at $0.65/DBU covers ML model training workloads, typically on GPU instances. The high DBU rate reflects the premium compute (V100, A100 GPUs) and MLflow integration for experiment tracking, model registry, and artifact management.
Model Serving at $0.07/DBU is the cheapest rate and is serverless-only. It runs inference endpoints that scale to zero when not receiving requests. You pay only for the compute seconds your model actually uses to process requests. This makes it cost-effective even at low inference volumes.
Best for: Training: deep learning, LLM fine-tuning, XGBoost at scale. Serving: REST API endpoints for model predictions in production applications.
Step-by-Step Cost Calculation
Scenario: Daily ETL Pipeline on AWS
Premium vs Enterprise Tier
Databricks offers Premium and Enterprise tiers. The per-DBU rate is the same for both tiers on most clouds. The difference is in features, not unit pricing. Enterprise adds advanced security (customer-managed keys, HIPAA, FedRAMP), enhanced governance, and dedicated support.
| Capability | Premium | Enterprise |
|---|---|---|
| DBU rates | Standard published rates | Same rates (discount via commit) |
| Unity Catalog | Yes | Yes |
| Customer-managed keys | No | Yes |
| HIPAA / FedRAMP | No | Yes |
| Enhanced audit logging | Basic | Verbose |
| Committed-use discounts | Available | Better rates at scale |
| Support SLA | Standard | Premium + TAM |
Most organizations start on Premium and move to Enterprise when compliance requirements demand it. See the Enterprise pricing page for negotiation details.