Databricks pricing 2026: DBU rates, discount bands & levers.

01 Key findings

1. The DBU is a matrix, not a price. The same workload prices differently across four cluster types (Light, Classic, Photon, Serverless), three SKU tiers (Standard, Premium, Enterprise), and three clouds. Enterprise tier charges up to 25% more than Standard for identical compute.

2. Cluster type beats negotiation. Moving ETL from All-Purpose Classic ($0.55/DBU) to Jobs Compute Photon ($0.30/DBU plus a favourable consumption multiplier) cuts cost 40–60% with no commercial conversation at all.

3. Every DBU carries a second bill. Underlying EC2/VM compute, storage, and networking add 30–60% on top of the Databricks DBU charge. Discounting only the DBU half leaves the larger optimisation on the table.

4. Discount scales with commit, non-linearly. Bands move from 10–18% at sub-$1M to 35–48% at $10M+. Three-year terms add 4–8 points but transfer carry-forward risk to the buyer.

5. Cloud placement is a commercial lever. Azure Databricks draws down MACC; GCP Marketplace billing counts toward CUD commits. For buyers holding large cloud commitments, placement changes the breakeven independent of DBU rate.

02 The DBU pricing model

Databricks bills in Databricks Units (DBUs) — one unit of compute capability consumed per hour. The rate varies by workload type (Jobs, All-Purpose, SQL, Model Serving), cluster type (Light, Classic, Photon, Serverless), SKU tier (Standard, Premium, Enterprise), and cloud. The customer pays the DBU rate to Databricks plus the underlying cloud compute, storage, and network cost to the cloud provider.

Three consequences buyers underestimate: the DBU price for one workload differs up to 4x across cluster types, so the wrong cluster type doubles or quadruples cost without touching the SKU; Enterprise tier charges up to 25% more than Standard for identical compute; and the DBU sits on top of underlying cloud cost that typically equals 30–60% of the DBU charge on AWS and Azure.

03 DBU rates by SKU & cloud

Databricks DBU rates on AWS, as published on the public pricing page in May 2026. Rates on Azure Databricks are similar but slightly higher, reflecting Azure's standard product premium; on GCP they align with AWS at the workload level.

A typical lakehouse mix (35% Jobs Photon, 25% SQL Serverless, 20% All-Purpose Classic, 10% SQL Pro, 10% Model Serving) averages ~$0.52/DBU at Enterprise tier before discount. At 1M DBUs/month that lists at $520,000/month — $6.24M/year — before negotiated discount and before the EC2/S3 cost underneath.

Azure Databricks folds Enterprise-grade features (Private Link, customer-managed keys, IP access lists, audit logging) into Premium tier rather than a separate line, and consumption counts toward the Azure MACC commitment. GCP's lever is Marketplace billing against a Google Cloud CUD commitment. For buyers with large cloud commits, either changes the breakeven versus AWS.

04 Photon, Serverless & Mosaic AI

Photon is Databricks' vectorised query engine (Premium and Enterprise tiers). It does not change the DBU price but changes the consumption rate, typically 3x–8x faster for SQL and ETL, so the same workload completes in a fraction of the wall-clock time and total DBU consumption drops even after the Photon multiplier. SQL Serverless carries the highest DBU rate ($0.70) but eliminates idle and spin-up cost; for bursty BI it lands 20–35% below SQL Pro, with the break-even at roughly 6–8 hours/day of sustained utilisation.

Mosaic AI — the generative AI suite — splits across pay-per-token Foundation Model APIs and provisioned-throughput DBUs for hosted models. Unity Catalog governance is included at no additional DBU cost on Premium and Enterprise, though the metastore charges per operation (watch a 10x–100x spike during Hive-to-Unity migration cutover).

For Mosaic AI weighed against direct OpenAI or Anthropic deployment, see our enterprise LLM cost comparison and AI vendor selection framework. Where the data already lives in Databricks the platform argument (no egress, no data movement) is strong; where it does not, the cost case is weaker.

05 Commit tiers & discount bands

Databricks does not publish discount bands. Realised discounts observed in advisor-led Databricks negotiations, 2024–2026:

Midpoint DBU discount by commit band — the curve is steep early, then flattens:

Contracts run 1, 2, or 3 years. Three-year terms add 4–8 points over one-year but add carry-forward risk if cloud strategy shifts. Unused commit at the end of year one or two carries forward as a credit (capped at ~25–35% of the year's commit); year-three unused commit is forfeit. Widening the carry-forward cap and extending it to year three are core negotiation levers.

06 The double-cost trap

The DBU charge is not the total cost. Underlying cloud charges — EC2 or Azure VM compute, EBS/managed disk, S3/ADLS storage, NAT and PrivateLink fees, cross-AZ traffic — typically add 30–60% on top of the Databricks DBU bill.

07 Cost optimisation patterns

Five patterns deliver most of the savings in Databricks estates. Independent cost reviews typically identify 22–35% of total spend as recoverable through architecture alone — no vendor agreement required.

1. Move ETL from All-Purpose to Jobs Compute. The same cluster runs ETL at $0.30/DBU instead of $0.55. Savings: 30–40% of ETL DBU spend.

2. Enable Photon for SQL and ETL. Multiplies the DBU rate but divides wall-clock time by 3x–8x. Net savings: 40–60% on Photon-eligible workloads.

3. Right-size with auto-scaling and tight idle timeouts. Idle clusters still burn the underlying VM cost. Cutting idle timeout from 60 to 15 minutes saves 10–20% of compute.

4. Migrate bursty BI to SQL Serverless. When sustained load is under ~6 hrs/day, Serverless lands 20–35% below SQL Pro by eliminating idle hours.

5. Split dev/test/prod and apply Standard tier to dev. Where Enterprise features are not required, dev compute drops 15–25%.

08 Commit-sizing framework

Four factors set the right commit band and term. Weight them to your own trajectory before signing.

09 Commit or stay on-demand

10 Contract levers

Eight commercial levers are worth negotiating beyond the headline DBU discount — typically conceded selectively, so rank them by your own architecture risk and trade only what matters.

For the full framework see our cloud contracts guide and cloud cost optimization; to engage, our cloud contract negotiation and SaaS license optimisation services. For peer comparisons, see the Snowflake pricing pillar, Snowflake vs Databricks vs BigQuery, and our AWS vendor hub.