Databricks list price ranges from $0.07 per DBU for Jobs Light on AWS Standard tier to $1.40 per DBU for SQL Serverless on Azure Enterprise tier, a 20x spread that turns architecture choice into the single largest cost lever. Realised enterprise pricing at $1 million annual commit lands 18 to 28 percent below list. At $3 million the discount band moves to 25 to 38 percent. At $10 million it reaches 35 to 48 percent. This pillar covers every Databricks DBU rate, the discount bands by commit size, and the architecture choices that move spend faster than negotiation does.
Inside This Pillar
Databricks DBU pricing model
Databricks bills usage in Databricks Units (DBUs). A DBU represents one unit of compute capability consumed per hour. The DBU rate varies by workload type (Jobs, All-Purpose, SQL, Model Serving), by cluster type (Light, Classic, Photon, Serverless), by SKU tier (Standard, Premium, Enterprise), and by cloud (AWS, Azure, GCP). The customer pays the DBU rate to Databricks plus the underlying cloud compute, storage, and network cost to the cloud provider.
The pricing model has three consequences buyers underestimate. First, the DBU price for the same workload differs by up to 4x across the four cluster types, so picking the wrong cluster type doubles or quadruples cost without changing the SKU. Second, the DBU price for the same workload differs across Standard, Premium, and Enterprise tiers, with Enterprise charging up to 25 percent more than Standard for identical compute. Third, the Databricks DBU is on top of the underlying cloud cost, which typically equals 30 to 60 percent of the DBU charge on AWS and Azure.
DBU rates on AWS
On AWS, Databricks DBU rates as published on the public pricing page in May 2026:
| Workload | Cluster type | Standard tier | Premium tier | Enterprise tier |
|---|---|---|---|---|
| Jobs Compute | Light | $0.07 / DBU | $0.10 / DBU | $0.13 / DBU |
| Jobs Compute | Classic | $0.15 / DBU | $0.30 / DBU | $0.40 / DBU |
| All-Purpose Compute | Classic | $0.40 / DBU | $0.55 / DBU | $0.65 / DBU |
| All-Purpose Compute | Photon | not available | $0.55 / DBU | $0.65 / DBU |
| SQL Compute | Classic | not available | $0.22 / DBU | $0.22 / DBU |
| SQL Compute | Pro | not available | $0.55 / DBU | $0.55 / DBU |
| SQL Compute | Serverless | not available | $0.70 / DBU | $0.70 / DBU |
| Model Serving | Provisioned | not available | $0.07 / DBU | $0.07 / DBU |
| Model Serving | Serverless | not available | $0.07 / DBU | $0.07 / DBU |
For a typical lakehouse running daily ETL jobs and on-demand SQL on AWS, the DBU mix breaks down roughly as 35 percent Jobs Compute Photon, 25 percent SQL Serverless, 20 percent All-Purpose Classic, 10 percent SQL Pro, 10 percent Model Serving. At Enterprise tier on AWS that mix averages to approximately $0.52 per DBU before discount. A workload consuming 1 million DBUs per month lists at $520,000 per month, $6.24 million per year, before negotiated discount and before the underlying EC2 and S3 cost.
DBU rates on Azure
On Azure, Databricks is sold as a first-party service (Azure Databricks). Rates are similar but slightly higher than AWS, reflecting Azure's standard product premium:
| Workload | Cluster type | Standard tier | Premium tier |
|---|---|---|---|
| Jobs Compute | Light | $0.07 / DBU | $0.10 / DBU |
| Jobs Compute | Classic | $0.15 / DBU | $0.30 / DBU |
| All-Purpose Compute | Classic | $0.40 / DBU | $0.55 / DBU |
| All-Purpose Compute | Photon | not available | $0.55 / DBU |
| SQL Compute | Classic | not available | $0.22 / DBU |
| SQL Compute | Pro | not available | $0.55 / DBU |
| SQL Compute | Serverless | not available | $0.70 / DBU |
Azure Databricks does not surface an Enterprise tier as a separate line item. Enterprise-grade features (Private Link, customer-managed keys, IP access lists, audit logging) are included in Premium tier. Azure-specific commercial structures matter more than tier choice: Databricks consumption on Azure counts toward Microsoft Azure Consumption Commitment (MACC), which means committed Azure spend can be drawn down via Databricks usage. For organisations holding large MACC commits, this is a material reason to favour Azure Databricks over AWS Databricks at otherwise equivalent total cost. See our Azure MACC versus CTP comparison for the MACC commercial framework.
DBU rates on GCP
On GCP, Databricks rates align with AWS at the workload level. Databricks on GCP is the smallest of the three cloud deployments and carries less local sales motion. The GCP-specific commercial lever is the Google Cloud Marketplace billing option, which allows Databricks spend to count toward a Google Cloud Committed Use Discount (CUD) commitment. For organisations with large Google Cloud commits, this changes the breakeven against AWS or Azure deployment. See our Google Cloud CUD versus Flex CUD guide for the commit interaction.
Photon and SQL Serverless
Photon is Databricks' vectorised query engine, available on Premium and Enterprise tiers. Photon does not change the DBU price but it changes the DBU consumption rate, typically by 3x to 8x for SQL and ETL workloads. The net cost effect is favourable: the same workload completes in one third to one eighth the wall-clock time, and total DBU consumption drops by a comparable factor even after accounting for the Photon DBU multiplier.
SQL Serverless is the most aggressively priced option at $0.70 per DBU on AWS Premium, but it eliminates the cluster idle cost and the cluster spin-up cost. For BI-style workloads with bursty query patterns, SQL Serverless typically lands 20 to 35 percent below SQL Pro despite the higher DBU rate because the cluster idle hours disappear. For sustained-load workloads, SQL Pro remains cheaper. The break-even is approximately 6 to 8 hours per day of sustained utilisation.
The cluster type lever: Moving an ETL workload from All-Purpose Classic ($0.55 per DBU on AWS Premium) to Jobs Compute Photon ($0.30 per DBU plus Photon multiplier) typically cuts cost by 40 to 60 percent while improving throughput. This is the single largest Databricks cost lever and it requires no commercial negotiation.
Unity Catalog cost
Unity Catalog is the Databricks governance layer for tables, files, ML models, and now AI tools. It is included at no additional DBU cost on Premium and Enterprise tiers but the underlying storage and metastore queries do consume DBUs. The hidden cost is the cross-workspace governance pattern: enterprises adopting Unity Catalog typically consolidate workspaces, which exposes previously hidden ungoverned spend.
For migration from the legacy Hive Metastore to Unity Catalog, plan 6 to 12 weeks of dual-running cost. The Unity Catalog metastore charges per metastore operation, and the migration scripts can spike metastore operations by 10x to 100x during cutover. This is a four to six figure transient cost that finance teams routinely miss in the migration business case.
Mosaic AI and Foundation Models
Mosaic AI is the Databricks generative AI suite, covering Foundation Model APIs, Vector Search, AI Functions, Model Serving for fine-tuned and provisioned-throughput models, and the Mosaic Pretraining service. Pricing splits across two structures: pay-per-token for Foundation Model APIs, and provisioned-throughput DBUs for hosted models.
| Service | Pricing unit | Rate (Premium tier) |
|---|---|---|
| Foundation Model APIs (Pay per token) | Per 1M input tokens | $0.50 to $15 depending on model |
| Foundation Model APIs (Pay per token) | Per 1M output tokens | $1.50 to $75 depending on model |
| Provisioned Throughput Llama 3 70B | Per DBU hour | $0.07 per DBU at 12 DBUs per hour minimum |
| Vector Search (Direct Access) | Per DBU hour | $0.07 per DBU at 1 endpoint hour |
| Vector Search (Delta Sync) | Per DBU hour | $0.07 per DBU at 1 endpoint hour |
| AI Functions (ai_query) | Per query DBU | Charged at Foundation Model API rate |
| Mosaic Pretraining | Negotiated | From $200,000 for a 7B parameter pretrain run |
For organisations weighing Mosaic AI against direct OpenAI or Anthropic deployment, see our enterprise LLM cost comparison and AI vendor selection framework. The Databricks pitch is that hosting the model on Databricks means the data and the model live in the same platform, removing egress and data-movement cost. For workloads where the data already lives in Databricks, the platform argument is strong. For workloads where the data does not, the cost case is weaker.
Commit tiers and discount bands
Databricks does not publish discount bands. Realised discounts observed in advisor-led Databricks negotiations during 2024 to 2026:
| Annual commit (TCV) | DBU discount range | Photon and Serverless | Mosaic AI |
|---|---|---|---|
| $250K to $1M | 10 to 18 percent | 0 to 10 percent | Not available |
| $1M to $3M | 18 to 28 percent | 10 to 20 percent | 5 to 15 percent |
| $3M to $10M | 25 to 38 percent | 20 to 30 percent | 15 to 25 percent |
| $10M+ | 35 to 48 percent | 25 to 35 percent | 20 to 30 percent |
Databricks commit contracts run 1, 2, or 3 years. Three-year terms typically add 4 to 8 percentage points of discount over one-year terms but add commensurate carry-forward risk if cloud strategy shifts. Term length should be weighed against architecture roadmap, not against the discount in isolation.
Commit shortfall behaviour is contractually permissive. Unused commit at the end of year one or year two carries forward as a credit against the same Databricks SKUs in the following contract year, with a cap (typically 25 to 35 percent of the year's commit). Year-three unused commit is forfeit. The negotiation lever is to widen the carry-forward cap and to extend carry-forward to year-three of a multi-year term.
Underlying cloud cost
Databricks' DBU charge is not the total cost. The underlying cloud charges (EC2 or Azure VM compute, EBS or managed disk storage, S3 or ADLS storage, NAT gateway and PrivateLink fees, cross-AZ traffic) typically add 30 to 60 percent on top of the Databricks DBU bill.
| Cloud cost category | Share of total Databricks bill |
|---|---|
| EC2 / Azure VM compute | 20 to 40 percent |
| Storage (S3 / ADLS) | 3 to 10 percent |
| Networking (NAT, PrivateLink, cross-AZ) | 5 to 15 percent |
| Other (snapshots, audit logs) | 1 to 5 percent |
The negotiation implication is that Databricks DBU discounts are one half of the cost picture. The other half is the AWS or Azure or GCP commit. EC2 Savings Plans, Azure Reservations, and Google Cloud CUDs applied to the underlying VMs can reduce the cloud half by 30 to 55 percent. For the cloud-side negotiation, see our AWS EDP negotiation guide, Azure EA negotiation, and Google Cloud Enterprise Agreement guide.
Cost optimisation patterns
Five optimisation patterns deliver most of the savings in Databricks estates. Independent Databricks cost reviews on enterprise customers typically identify 22 to 35 percent of total spend as recoverable through architecture changes alone.
First, move ETL workloads from All-Purpose to Jobs Compute. The same cluster runs ETL on Jobs Compute at $0.30 per DBU on AWS Premium instead of $0.55 on All-Purpose Classic. Savings: typically 30 to 40 percent of ETL DBU spend.
Second, enable Photon for SQL and ETL workloads. Photon multiplies the DBU rate but divides wall-clock time by 3x to 8x. Net savings: typically 40 to 60 percent on Photon-eligible workloads.
Third, right-size clusters with auto-scaling and aggressive idle timeouts. Idle clusters consume the underlying EC2 or VM cost even when no Databricks workload is running. Reducing idle timeout from 60 minutes to 15 minutes typically saves 10 to 20 percent of compute cost.
Fourth, migrate BI workloads to SQL Serverless when query patterns are bursty (less than 6 hours of sustained load per day). Savings: 20 to 35 percent versus SQL Pro on bursty workloads.
Fifth, separate dev, test, and prod workspaces and apply Standard tier to dev where Enterprise features are not required. Savings: 15 to 25 percent on dev compute.
Contract levers
The Databricks contract has eight commercial levers worth negotiating beyond the headline DBU discount:
Carry-forward cap on unused commit, raise from 25 to 40 percent. Price-protection clause locking DBU rates for the contract term. Termination for convenience after year one with pro-rated refund of unused commit. PrivateLink and customer-managed key inclusion without per-workspace fees. Support tier upgrade (Business Critical) at no premium when annual commit exceeds $3 million. Right to convert legacy workspaces to Unity Catalog without re-pricing. Marketplace credit inclusion against the commit. Right to true-down annual commit by up to 20 percent at the annual anniversary.
These levers are typically conceded selectively. Buyers should rank the eight by their own architecture risk and trade only what matters. For the full Databricks negotiation framework see our cloud contracts guide and cloud cost optimization. To engage on Databricks negotiation, see our cloud contract negotiation service and SaaS license optimisation service. For peer comparisons, see Snowflake pricing pillar, Snowflake vs Databricks vs BigQuery, and our AWS vendor hub.