Databricks Serverless: Pricing, How It Works, When It Saves Money
Serverless eliminates cluster management and bundles compute costs into the DBU rate. Higher per-DBU pricing but no separate cloud bill. The math favors serverless for bursty workloads and penalizes sustained usage.
What Is Serverless in Databricks?
Databricks serverless removes the need to provision and manage clusters. You submit work and Databricks handles the infrastructure: spinning up compute, scaling based on demand, and shutting down when idle. There is no cluster configuration, no instance type selection, and no infrastructure cost on your cloud bill.
The trade-off is straightforward: higher per-DBU rate in exchange for zero infrastructure management and zero idle cost. You pay only for the compute seconds your workloads actually use.
Serverless SQL Pricing
$0.70/DBU
Includes all compute infrastructure
vs SQL Classic: $0.22/DBU + cloud
vs SQL Pro: $0.55/DBU + cloud
Serverless SQL Warehouses start in under 10 seconds (vs 5 to 10 minutes for classic), auto-scale elastically, and shut down instantly when idle. The $0.70/DBU rate bundles the Databricks platform fee and the cloud compute cost into one price. No separate EC2 or VM charges appear on your cloud bill.
Serverless Compute Pricing
Serverless compute for notebooks and jobs is priced differently from SQL. DBU rates vary by workload type and are higher than classic equivalents, but include the compute infrastructure cost.
| Workload | Serverless Rate | Classic Rate | Notes |
|---|---|---|---|
| SQL Warehouse | $0.70/DBU | $0.22-$0.55/DBU + cloud | Most mature serverless offering |
| Jobs Compute | $0.35/DBU | $0.15/DBU + cloud | GA for Python and Spark |
| Notebooks | $0.60/DBU | $0.40/DBU + cloud | Interactive development |
| Delta Live Tables | $0.45/DBU | $0.20-$0.25/DBU + cloud | Declarative pipelines |
| Model Serving | $0.07/DBU | N/A | Serverless only |
Classic vs Serverless: Monthly Cost Comparison
Side-by-side cost estimates for three usage levels. Classic assumes i3.2xlarge instances on AWS with spot workers.
| Usage Pattern | Classic (DBU + Cloud) | Serverless | Winner |
|---|---|---|---|
| Bursty: 2 hrs/day, 15 days/mo | $180 + $130 = $310 | $295 | Serverless |
| Moderate: 6 hrs/day, 22 days/mo | $396 + $277 = $673 | $840 | Classic |
| Heavy: 12 hrs/day, 22 days/mo | $792 + $554 = $1,346 | $1,680 | Classic |
| Overnight batch: 4 hrs/night, 30 days | $360 + $252 = $612 | $510 | Serverless |
| Ad-hoc SQL: 20 queries/day avg | $450 + $330 = $780 | $380 | Serverless |
Estimates assume 3 nodes per cluster, Jobs Compute rates. Actual costs vary by workload type and instance selection.
When Serverless Saves Money
- Bursty workloads: Short tasks running a few hours per day. Classic clusters waste money during startup (5 to 10 min) and idle periods. Serverless starts instantly and stops instantly.
- Small teams: Teams with 1 to 5 users running ad-hoc queries. The operational overhead of managing clusters is not worth the cost savings.
- Unpredictable usage: Workloads that spike 3x to 5x during certain periods. Serverless auto-scales without pre-provisioning. Classic clusters require over-provisioning or manual scaling.
- Development and testing: Interactive notebook work where a developer runs a cell, thinks for 10 minutes, runs another cell. Classic clusters burn compute during the thinking time.
When Classic Clusters Are Cheaper
- Sustained workloads: ETL pipelines running 8+ hours daily. The lower classic DBU rate plus spot instances beats serverless pricing at sustained utilization.
- Large teams: 10+ users sharing clusters. The cost per user drops significantly with shared classic clusters.
- Spot instance eligibility: Batch jobs that tolerate interruptions. Spot instances reduce the cloud portion by 60% to 70%, making classic dramatically cheaper.
- GPU workloads: ML model training requiring specific GPU instances. Serverless does not support GPU compute, so classic is the only option.