Home › AWS › AWS data services › Cost Explorer, CUR, Budgets
AWS data services · Observability, audit and cost · about 10 min read
AWS Cost Explorer, Cost and Usage Report (CUR 2.0 in AWS Data Exports) and AWS Budgets
AWS's cost tools: charts of spend, a line-by-line bill exported to S3, and budgets that warn before spend runs away.
In plain words
What it is
AWS bills you for thousands of small things: an hour of a server, a gigabyte stored for a month, a million API requests. Three tools help you see and control that bill. Cost Explorer is a chart tool in the console. It shows spend by service, account, region or tag over time, and makes a simple forecast. The Cost and Usage Report, now CUR 2.0 inside a service called AWS Data Exports, is the bill in full detail. It has one row per resource, per usage type and per hour or day, written as Parquet files to your S3 bucket up to three times a day. AWS Budgets sets a limit, actual or forecast, and sends an alert, or even takes an action, when spend crosses it.
Tags make all three useful. A tag is a label such as data_product=sales_orders or team=quality on a resource. Once you turn a tag on as a cost allocation tag, it appears as a column in CUR and as a filter in Cost Explorer and Budgets. It applies only from that day on, not to past spend.
A concrete example: a tyre maker runs its data platform on AWS with Databricks. In March the cloud bill jumps by a third. Cost Explorer shows EC2 grew, but not why. CUR 2.0 shows the EC2 rows carry Databricks' own tags, including the job id and cluster id, plus the team tag the platform adds. Databricks billing system tables show the DBUs for the same jobs. Joined together, they show that one quality data product's nightly job moved to a larger instance type and ran three times a night after a config change. The fix takes ten minutes. Finding it took one query, because the join was already built.
FOCUS, the FinOps Open Cost and Usage Specification, is a common column layout for cloud bills. AWS Data Exports can write the bill in FOCUS 1.2 format, which makes it easier to put AWS costs next to other clouds or tools in one table.
Cost Explorer is the monthly statement, CUR is every receipt in a shoebox sorted by date, and Budgets is the bank text that says you are near your limit.
Big picture
Where it fits
The lit lane is where Cost Explorer, CUR, Budgets works. Boxes on the left feed it; boxes on the right use what it produces.
These tools sit in the band across the bottom of the picture, with health and audit. On the reference platform the cost of a data product has two halves. Databricks charges DBUs for compute, recorded in the billing system tables, and for serverless compute that is the whole cost. AWS charges for the EC2 instances behind classic compute, S3 storage and requests, KMS, NAT gateway data, DMS, AppFlow, Kinesis, Firehose and the rest, recorded in CUR.
A cost view in prod_ops joins the two. CUR 2.0 lands in an S3 bucket and Auto Loader reads it into a table. Databricks billing system tables are already tables. The join keys are tags. Databricks puts its default tags and any custom tags on the EC2 instances it starts, and the platform tags each AWS mover and bucket with its data product. The result is cost per data product, per pipeline run and per team, with findings such as idle clusters or a mover that is far too large. The site's cost guide covers this design in depth.
Cost Explorer stays useful for quick questions in the console and for Savings Plans and Reserved Instance advice. Budgets give early warnings: one budget per account and per large data product, on forecast spend, emailing the owner and posting to the platform channel. Databricks budget policies do the same job on the Databricks side.
How it works
The ideas you need
| Idea | What it means |
|---|---|
| Cost Explorer | A console tool that charts spend and usage over time, groups it by service, account or tag, and forecasts the coming months. |
| AWS Data Exports | The service that writes billing data to S3: CUR 2.0, FOCUS, cost optimisation recommendations, carbon emissions and the legacy CUR. |
| CUR 2.0 | The detailed bill as a table with a fixed set of columns, one row per resource and usage type per time period, delivered as Parquet. |
| FOCUS | The FinOps Open Cost and Usage Specification, a common bill layout across clouds. Data Exports supports FOCUS 1.2 with AWS columns. |
| Cost allocation tag | A resource tag you turn on in Billing so that it appears in CUR, Cost Explorer and Budgets. It applies only from activation forward. |
| Amortised cost | Cost with upfront Savings Plan and Reserved Instance fees spread over the hours they cover, which is fairer for chargeback. |
| Budget | A limit on cost or usage, actual or forecast, with alerts by email or SNS and optional actions such as applying a restrictive policy. |
| Cost Anomaly Detection | A free model that watches spend by service, account or tag and alerts on unusual jumps. |
| Databricks default tags | Tags such as Vendor, ClusterId and JobId that Databricks adds to the EC2 instances behind classic compute. |
Choosing
When to use it, and when not
Use it when
- You need cost per data product, per team or per pipeline run that includes both AWS resources and Databricks DBUs.
- You want an early warning when an account, team or data product is on track to overspend this month.
- You need chargeback or showback to business units, with amortised Savings Plan costs.
- You want to decide on Savings Plans or Reserved Instances for steady EC2 usage behind classic compute.
- You run many AWS accounts and need one detailed bill in the management or billing account.
Avoid it when
- You need real-time cost; CUR updates up to three times a day and Cost Explorer at least daily, so neither stops a runaway job in minutes. Use compute policies, job timeouts and auto-stop.
- You only want Databricks serverless costs; those are in Databricks billing system tables, not in CUR.
- You plan to give every engineer Cost Explorer API access in a tool that calls it on each page load; the API is charged per request.
Production grade
What goes wrong in production
| Problem | What breaks | What to do |
|---|---|---|
| Untagged spend | A third of the bill has no data product tag, so the cost view cannot assign it. | Enforce tags with tag policies and SCPs, add custom tags through Databricks compute policies, and report untagged spend as its own line every week. |
| Tags not active in CUR | A tag exists on resources but was never turned on as a cost allocation tag, so CUR has no column for it. | Activate tags in the management account as soon as they are agreed; they apply only from activation, not backwards. |
| Shared costs | NAT gateway data, KMS requests and the SQL warehouse are shared by many data products and carry no single owner. | Agree a split rule, such as by bytes read or by run time, record it in the cost view's code, and show shared cost separately. |
| Two halves that do not join | DBUs come from Databricks billing tables and EC2 from CUR, with different time grains and ids. | Join on the Databricks cluster or job tags present in both, at hourly or daily grain, and reconcile monthly totals against both bills. |
| CUR 2.0 migration | Reports built on the legacy CUR's changing columns break when moved to CUR 2.0's fixed schema. | Create a CUR 2.0 export next to the legacy one, compare a month of totals, then move the readers and retire the legacy export. |
| Budget alerts too late | A budget on actual monthly spend fires on the 28th, after the money is gone. | Use forecast-based budgets and daily anomaly detection, plus Databricks budget policies and job timeouts for fast-moving compute. |
| Marketplace billing | If Databricks is bought through AWS Marketplace, DBU charges also appear in CUR, and naive sums count them twice. | Mark Marketplace Databricks rows in the cost view and take DBUs from one source only. |
Alternatives
What to compare it with
| Alternative | Pick the alternative when | Pick Cost Explorer, CUR, Budgets when |
|---|---|---|
| Databricks billing system tables and budget policies | You want DBU cost per job, warehouse, user or tag, including serverless. | You need EC2, S3, NAT, KMS and AWS mover costs; those are only in the AWS bill. |
| Microsoft Cost Management | You are on Azure. | You are on AWS. |
| Google Cloud Billing export to BigQuery | You are on Google Cloud. | You are on AWS. |
| A FinOps tool such as Vantage, CloudZero or Apptio Cloudability | You want ready-made allocation, many clouds and SaaS bills, and you will pay a percentage or licence for it. | You already have Databricks and want cost joined with run logs and data products in your own tables. |
| OpenCost | You mostly run Kubernetes and need cost per namespace or pod. | Your compute is Databricks and managed AWS services. |
Cost
How you pay
Cost Explorer's console use is free. Its API is charged per request, and hourly and resource-level granularity is a paid option. Data Exports itself is free; you pay S3 storage for the files and the query cost of whatever reads them. Budgets are free to monitor, and action-enabled budgets are charged per budget per day after a small free allowance. Cost Anomaly Detection is free. Check the AWS Cost Management pricing page for current list prices.
What drives the bill
- Cost Explorer API calls from tools and scripts.
- Hourly and resource-level granularity in Cost Explorer, if turned on.
- Number of action-enabled budgets beyond the free allowance.
- S3 storage and query compute for years of CUR files.
Cost traps
- A dashboard that calls the Cost Explorer API on every page load; read CUR tables instead.
- Keeping both legacy CUR and CUR 2.0 exports with hourly, resource-level detail forever.
- Querying raw CUR files directly each time instead of a small, partitioned silver table.
Prices change; check the AWS pricing page for your region before deciding.
Security and governance
Keeping it safe
- Billing data reveals contracts, discounts and the shape of the business; restrict Cost Explorer and the CUR bucket to finance and platform leads.
- Create the export in the management or billing account, deliver to a bucket with a KMS key and a bucket policy that allows only the billing service to write.
- Use Data Exports' column selection and row filters to remove sensitive columns or accounts before sharing a cost table with teams.
- Grant read access to the cost view in Unity Catalog by group, with row filters so each team sees its own data products.
- Record changes to budgets, exports and cost allocation tags in CloudTrail.
Operating it
What to watch, and the limits
| Watch | Why |
|---|---|
| Untagged spend share | Percentage of each day's cost with no data product tag; target near zero. |
| Budget forecast against limit | Per account and per large data product, from AWS Budgets and Databricks budget policies. |
| Cost anomalies | Alerts from Cost Anomaly Detection by service and by cost allocation tag. |
| CUR delivery freshness | Time since the last CUR 2.0 file landed; alarm if more than a day. |
| Monthly reconciliation | Cost view totals compared with the AWS invoice and the Databricks invoice. |
| Cost per run | From the run log joined with billing; a sudden jump points to a config or data-volume change. |
| Limit | Value and notes |
|---|---|
| Cost Explorer history | 13 months by default; up to 38 months at monthly granularity when multi-year data is enabled |
| Hourly and resource-level data | Past 14 days only |
| CUR refresh | Updated up to three times a day |
| Concurrent exports | Up to 5 CUR 2.0 exports and 10 legacy CUR reports at the same time |
Limits and quotas checked 2026-11-11; many can be raised through Service Quotas.
Decisions
Situations and verdicts
| Situation | Choice | Verdict |
|---|---|---|
| Cost per data product across AWS and Databricks | CUR 2.0 to S3, read by Auto Loader, joined with billing system tables on tags in prod_ops | Pursue: the only way to see the whole cost of a product. |
| Existing legacy CUR feeding a finance report | Add a CUR 2.0 export, run both for a month, compare, then move | Pursue with care: CUR 2.0 is the recommended path; check schema differences first. |
| Stop a runaway nightly job within minutes | Job timeouts, compute policies and auto-stop in Databricks; Budgets as a backstop | Pursue: billing data is too slow to be the first guard. |
| Buy a third-party FinOps tool | Only if many clouds and SaaS bills must be combined | Keep if you have it: for one AWS plus Databricks platform, the own cost view joins better with run logs. |
FAQ
Questions people ask
What is the difference between CUR and CUR 2.0?
CUR 2.0 is delivered by AWS Data Exports with a fixed set of columns, and you can choose columns and filter rows with SQL. The legacy CUR had columns that changed with your usage. AWS recommends CUR 2.0, and the legacy report still works for now.
Where do Databricks costs appear?
DBUs appear in Databricks billing system tables, and in CUR too if you buy Databricks through AWS Marketplace. The EC2 instances behind classic compute, S3 and networking appear in CUR. Serverless compute appears only on the Databricks side.
Can AWS Budgets stop spending?
Budget actions can apply an IAM or SCP policy or stop some EC2 and RDS instances when a threshold is reached. They act on billing data that lags by hours, so use Databricks policies and timeouts for fast protection.
How far back does Cost Explorer go?
Thirteen months by default, and up to 38 months at monthly granularity once multi-year data is enabled in the management account. Hourly and resource-level detail covers the past 14 days.
Should I use FOCUS or CUR 2.0?
Use CUR 2.0 when you need every AWS-specific column. Use FOCUS 1.2 when you want one layout across clouds and tools. Both come from Data Exports, so you can create both.
Sources
Checked against
- AWS docs: What is AWS Data Exports
- AWS docs: Migrating from CUR to Data Exports CUR 2.0
- AWS docs: Configuring multi-year and granular data in Cost Explorer
- AWS docs: Troubleshooting CUR 2.0
- AWS docs: Services in maintenance (AWS Usage Report)
Checked 2026-11-11.
Related
Where it connects
Cost calculator · AWS and Databricks reference platform · Amazon EC2 for Databricks compute · AWS Organizations, Control Tower and SCPs · Amazon CloudWatch · All AWS data services