Azure Databricks Pricing: DBU Rates and Costs (2026)
Nov 5, 202614 min read
Azure Databricks charges per DBU-hour at rates from $0.30 to $1 depending on compute type, and classic compute adds a separate Azure VM bill on top, which is why teams that wire up job triggers through tools like Activepieces still end up reconciling two invoices.
Ten DBU-hours cost $3.00 on Jobs compute and $9.50 on Serverless All-Purpose, so the compute choice triples the rate before any VM is counted.
Azure Databricks Pricing: The Plans and DBU Rates
What a DBU is and what it is not
Azure Databricks is a managed Spark and SQL analytics platform that runs inside your Azure subscription. A DBU, or Databricks Unit, is a unit of processing capability, billed per second. It's not a unit of time, a VM or a dollar amount.
Price table by compute type, checked 5 November 2026
| Compute type | Unit billed | Price per DBU | Inclusions | Commitment allowance |
|---|---|---|---|---|
| Jobs Compute | DBU-hour | $0.30 | Azure VMs billed separately | Counts toward a pre-purchase plan |
| DLT Core | DBU-hour | $0.30 | Azure VMs billed separately | Counts toward a pre-purchase plan |
| All-Purpose | DBU-hour | $0.55 | Azure VMs billed separately | Counts toward a pre-purchase plan |
| SQL Pro | DBU-hour | $0.55 | Azure VMs billed separately | Counts toward a pre-purchase plan |
| Serverless Jobs | DBU-hour | $0.47 | Infrastructure bundled into the rate | Counts toward a pre-purchase plan |
| Serverless SQL | DBU-hour | $0.70 | Infrastructure bundled into the rate | Counts toward a pre-purchase plan |
The rates come from DBU Watch. Jobs classic is $0.30 per DBU-hour, which makes it the cheapest rate.
According to Dbuwatch, Serverless Jobs is $0.47, which means infrastructure management overhead is priced into every DBU processed. SQL Pro is $0.55.
Per DBU Watch, Serverless SQL is $0.70, and Serverless All-Purpose is $1, the highest rate listed, which means teams avoid it unless all-purpose features are required.
Pay-as-you-go bills the rates above with no commitment.
If you are running this arithmetic for your own team, see what the same workload costs on Activepieces.
Job on the wrong compute
Nobody chose the cluster. The nightly pipeline at this five-person team ran on an all-purpose cluster, which Databricks bills at the interactive rate rather than the lower jobs rate.
It was the one the pipeline was built on, and it stayed there.
Azure Databricks workload setup and cost per run
With no dedicated platform engineer on the team, the setup was one pipeline, run nightly. Cost per run is nodes × hours × (DBU rate + VM rate), and every term leaned the wrong way.
Because the VM charge lands as a separate Azure line, the Databricks page showed one number and the subscription showed another.
Nightly pipeline timeline on Azure Databricks
- On day 1, the team built and tested the pipeline on an all-purpose cluster, because that cluster was already open.
- On day 2, the team scheduled the pipeline for 02:10 and pointed it at the same cluster, which they left running afterward.
- On day 3, the 02:10 run executes and bills at the interactive rate, which means the job costs more than expected from jobs compute. The cluster keeps running after it finishes, and the team notices only when an alert arrives.
We have no confirmed node count or hour total for these three days, so we won't put a number on them.
The run itself is short, and the hours the cluster sat idle around it are what you pay for, at the most expensive rate, twice (once in DBUs, once in VMs).
When a Databricks cost alert should fire
Day 2, not Day 3, was when it should have fired. That was the moment a cluster with no active work stayed up after the build, and an auto-termination setting would have shut it down then.
Workload Cost for a Month
On all-purpose compute ($0.55), the same job costs $0.25 more per DBU than on jobs compute ($0.30). The mistaken setup therefore pays about 83% more on the line the pricing page actually shows. That's before a single idle hour is counted.
| Mistaken (All-Purpose $0.55) | Corrected (Jobs $0.30) | Dollar gap | |
|---|---|---|---|
| DBU charge | $0.55 per DBU | $0.30 per DBU | $0.25 per DBU |
| Azure VM hours | D4s v3 at $0.192/hr, DS3 v2 at $0.229/hr | Same VMs, same rates | None per hour, but every idle hour is billed |
| Storage | Same | Same | None |
| Total | DBU line plus VM line, both running through idle time | DBU line plus VM line, only while the job runs | $0.25 per DBU, plus the VM rate on each idle hour |
| Wasted hours | Hours the cluster sat up waiting | None; the cluster ends with the job | Every hour removed saves both lines |
Whoever reads only the Databricks rate card sees roughly half the cost, so budget both lines together or the monthly total will surprise you.
Once VM cost is folded into the DBU rate, as serverless does, the separate Azure line disappears and the idle risk moves to startup and billing rules. That can be cleaner for short, bursty jobs.
For steady scheduled work, price it against the corrected classic column, not the mistaken one, or the comparison flatters it.
The workflow that launches the job should not add a meter of its own, either, and that is where per-step pricing goes wrong: it charges more for splitting a launcher into small, retryable steps, which is the habit that keeps a nightly job reliable.
Activepieces meters the run, at 1 credit however many steps it has, so a ten-step launcher costs the same as a two-step shortcut. Zapier's per-task and Make's per-module billing, both on their pricing pages, charge more for the split.
The Databricks bill then still turns on compute type and hours alone.
Hidden costs and billing limits
Even in the corrected classic column, five charges are still missing. They appear on other lines of the invoice or as limits in your subscription. Each is documented by Microsoft or Databricks, which means each can be estimated before the bill arrives:
- Virtual machine and managed disk charges, billed by Azure separately from DBUs (Azure Databricks pricing page on Microsoft Learn)
- Idle minutes before auto-termination (Databricks compute configuration docs)
- Network egress (Azure bandwidth pricing)
- Storage for data and logs (Azure Storage pricing)
- Regional core quota per VM family (Azure subscription limits docs)
Only the first two scale with how carelessly clusters are configured, so start there.
Until it terminates, a cluster keeps billing both VM and DBU. The auto-termination timer is therefore a per-run cost: every idle minute is paid at the full node rate, multiplied by the number of nodes.
A pipeline that exports nightly pays for egress nightly, because data leaving Azure or crossing regions is billed by Azure, not Databricks.
Azure Databricks quotas and API limits
Watch the API limits too. According to APIs.io, Jobs create and run-now each allow 20 requests per second, so an orchestrator launching 200 jobs at once needs about ten seconds and must retry.
Apis reports that DBFS allows 30 requests per second, which throttles file-heavy uploads.
Jobs runs/submit allows 35 per second. Runs/get allows 100 per second.
Worth checking against a plan that does not meter every step: one credit covers a whole run on Activepieces.
Azure Databricks Versus Its Closest Alternatives on Cost
On a 10 TB monthly BI workload, Databricks SQL costs $896 for compute against $115 on Fabric and $221 on BigQuery. Snowflake sits in between at $640 (Big Data Explained).
The table shows what each platform meters and whether the machines underneath are on a second invoice.
| Platform | Billing unit | Infrastructure billed separately | Idle behaviour | Commitment discount |
|---|---|---|---|---|
| Azure Databricks | DBU per second | Yes, Azure VMs | Bills until auto-termination | Pre-purchase of DBUs |
| Databricks on AWS | DBU per second | Yes, EC2 instances | Bills until auto-termination | Committed-use contract |
| Azure Synapse | Per hour of provisioned capacity (dedicated pool) | No | Bills until you pause it | Reservations |
| Snowflake | Credits per second of warehouse time | No, bundled into credits | Suspends automatically | Capacity purchase |
Idle behaviour and minimum charges
Idle is where the gap opens. Synapse compute came to $2,416 a month against $208 for storage (Big Data Explained), which means a pool left running costs more than ten times its own storage.
Automatic suspension is why Snowflake's compute stays at $640 against $230 for storage, since compute is not billed while the warehouse is suspended. Databricks SQL carries the same $230 storage figure, which shows storage isn't what separates it from Snowflake; the $256 compute difference is.
Which fits a small team's nightly workload
If a job runs for minutes each night, the platform that stops billing when the work ends wins. Snowflake's auto-suspend does this by default.
If a job runs for minutes each night, the platform that stops billing when the work ends wins.
Databricks does it only if you set an auto-termination timeout, and even then the VM line runs until the cluster is gone.
If the workload is mostly BI queries, Fabric's $115 compute with $208 storage totals $323. BigQuery's $221 with $200 totals $421. Either undercuts Databricks SQL, whose $896 and $230 total $1,126.
Databricks earns its price when you need notebooks, pipelines and machine learning on the same data. For dashboards alone, it's the dearer choice.
Azure Databricks cost controls checklist for clusters
The team stopped relying on judgment and wrote the choice into controls that apply to every cluster.
The list below is what's in place now; each item closes one of the leaks that produced the overrun.
- Cluster policy capping node size
- Forced auto-termination
- Jobs compute for scheduled work
- Azure budget alerts
- Team tags
- Weekly billing-table review
The first three stop money leaving. The last three tell you quickly if it still is.
Databricks cluster policies and auto-termination
A cluster policy is a template that limits what a user can choose when creating compute. Here it caps node size, so nobody can launch the largest virtual machine because it felt safer.
It also sets auto-termination so a cluster shuts itself down when idle. Neither can be overridden at creation.
Scheduled work moves to Jobs compute, which exists for the length of the run and then disappears. Interactive clusters stay for people actually typing.
Azure budget alerts and tagging for Databricks
When spend on a subscription crosses a threshold you set, Azure budget alerts fire. They watch the whole bill, including the separate virtual machine line the Databricks pricing table never adds up for you.
Tags matter for the same reason: each team's compute carries a tag, so the bill can be split by owner rather than argued over as one lump.
Once a week, someone reads Databricks' billing table, which records usage by tag, and compares it with what the team expected. That review catches what thresholds miss, such as a slow climb that never trips an alert.
Automate weekly cost report delivery
Activepieces runs this report as one scheduled workflow, under an MIT-licensed core with unlimited flows on every plan, including free. The workflow queries the billing table, compares each tag's usage with the agreed threshold, and posts any breach to the team's chat channel.
That's all it does.
What Activepieces does about this
The leak in this article is not the DBU rate. It is a cluster nobody owned, billed on two invoices, that nobody added together. Activepieces doesn't fix that for you, and it doesn't touch your clusters.
What it does is make the checking routine, so the review doesn't depend on someone remembering to run a query.
The core is MIT-licensed, and every plan, including free, has unlimited flows. A cost-watch workflow therefore doesn't cost more to run for being thorough.
A schedule trigger fires weekly, a step queries Databricks' billing table grouped by tag, and the next step compares each tag's usage with the threshold you agreed. A breach goes to the team's chat channel with the tag, the figure and the cluster name attached.
The same pattern covers the Day 2 alert. A scheduled flow can list running all-purpose clusters and flag any with no active work, which is the moment the team in this article missed. The launcher that starts the nightly job can also live in Activepieces.
It meters the run at 1 credit however many steps it has, so splitting it into retryable steps adds nothing to the bill.
What it won't do is read the Azure VM line for you unless you point a step at Azure cost data, and it won't stop a cluster unless you build that step and decide it should. Start with the report.
Once the Monday review opens on a finished table, the idle hours are the first thing anyone sees.
Frequently asked questions
Does Azure Databricks have a free tier or trial?
Azure Databricks has offered a time-limited trial that waives the DBU charge. There's no permanent free tier, and the Azure VMs under your clusters still bill from the first minute.
A trial therefore tests your workload, not your budget. Run it with auto-termination set, because an idle trial cluster still racks up the VM line. Confirm the current terms on the Azure pricing page before you plan around it.
Are DBU prices the same in every Azure region?
No. Both the DBU rate and the VM rate underneath it vary by region, so the same job can cost different amounts in two places.
The VM line is usually the larger of the two, which means a region that looks cheap on the DBU table can lose on the full bill.
Price the whole stack for each candidate region, using the VM sizes you'll actually run. Then weigh the result against data residency and latency, which often decide the region before cost does.
Is serverless compute cheaper than classic clusters?
It depends on how much of the time your compute sits idle. Serverless carries a higher effective rate because the VM cost's folded into it, but it starts quickly and stops billing when work ends.
If your classic clusters spend long stretches waiting between jobs, serverless can cost less overall. If they run hot all day, the premium for that bundled VM cost is pure overhead, and a well-sized classic cluster wins.
Measure your utilization over a typical week and compare the two rates against it. Don't pick on the headline DBU number.
Can I get a discount for committing to usage?
Yes. You buy Databricks Commit Units (DBCUs) through an Azure reservation, paying up front for a term in exchange for a lower effective DBU rate. Larger commitments and longer terms earn deeper discounts.
A DBCU covers only the DBU portion of the bill, and that is the catch. The VM line needs its own reservation or savings plan, and an unused commitment is money spent on nothing.
Commit only to the baseline you already run every month, and leave the spiky remainder on pay-as-you-go.
Related reading
References
Running the numbers
See what the same workload costs here.
Free forever plan, and every paid plan self-hosts at no extra cost.
See pricing Talk to sales
