The cloud sells elasticity: scale up when you need it, scale down when you don't, and pay for what you use. For data platforms, the scale-down half of that promise has a floor. Every Databricks workspace on AWS costs a minimum of about $33 per month, even when nobody runs a single job on it. Here is where that money goes, and what you can actually do about it.
We fielded this question from a smaller academic Databricks user. We have seen multi-workspace deployments with six-figure monthly bills, but there is also a long tail of smaller organizations running Databricks for teaching and research. Universities have predictable busy periods (student onboarding, final projects) and predictable idle periods (between semesters). The question was reasonable:
Over the summer when nobody is using the Databricks environment, I would expect about $0 in cost, but every month these accounts each cost around $33, what's going on?
The AWS bill and Cost Explorer showed roughly $33 of spend under "Elastic Compute Cloud" with zero workloads running. Drilling in, the charge came from one line item: "Amazon Elastic Compute Cloud NatGateway."
NAT Gateways are how private subnets in a VPC reach the public internet, and they are standard practice for keeping a data platform off the public network. They are also not free: $0.045 per hour just to exist. Every Databricks workspace provisions a VPC with a NAT Gateway, so a workspace with zero clusters and zero jobs still accrues $0.045 every hour, 744 hours a month. That is your $33.
Oct. 2026 update: fully serverless workspaces in Databricks no longer require a NAT Gateway and can be run for effectively zero cost outside of the DBUs themselves.
Every AWS customer running private subnets pays some version of this, so Databricks is not unusual here. But the practical conclusion stands: there is no configuration that takes an idle Databricks workspace to $0.
Your options:
- Accept the floor. $33 per month per workspace is the price of keeping the environment provisioned and ready.
- Decommission during idle periods. If a workspace is only needed a few months of the year, define it with the Databricks Terraform provider and destroy it when not in use. This works, but it is a blunt instrument: you are rebuilding the environment, not pausing it.
- Right-size the tool for the workload. For genuinely small use cases, serverless query options against Delta tables in S3 may carry no idle floor at all. That is an architecture decision, not a tuning exercise.
The headline number is $33 a month, entirely owed to AWS. Everything above that line is usage, and usage can be optimized. If your bill is uncomfortable, the NAT Gateway is not your problem; the spend above the floor is, and that is the part worth attacking.
Cost discipline starts with knowing where the floor is. If you want a second set of eyes on your platform bill, drop us a note.
