Amazon Redshift Pricing Explained: Provisioned vs Serverless Costs
Amazon Redshift has two pricing models that work very differently. Provisioned clusters charge per node-hour whether or not queries are running. Serverless charges only while queries are actively executing. Picking the wrong model for your workload is one of the most common causes of an unexpectedly large Redshift bill.
Simple explanation
A provisioned cluster is like renting office space. You sign a lease, and the rent is due every month whether your team shows up or not. A slow Tuesday costs the same as a packed Friday. Redshift Serverless is like booking hot-desk time. You only pay for the hours someone is actually sitting down and working. Empty desks cost nothing.
With a provisioned cluster, you rent a fixed set of compute nodes. The billing clock runs every hour those nodes exist, idle or not. To understand why this matters, read the Redshift architecture guide, which explains how nodes and slices map to cost. Provisioned pricing is predictable (you know the exact hourly rate) but wasteful if your warehouse sits unused for large parts of the day.
Redshift Serverless flips this model entirely. You don’t manage nodes. Redshift spins up compute automatically when a query arrives and stops it when the query finishes. You’re billed in RPU-hours (Redshift Processing Units), and idle time costs nothing. If your data warehouse sits unused for 20 hours a day, those 20 hours are free.
The right choice comes down to how many hours per day you actually query. Sporadic or unpredictable usage almost always favours Serverless. Heavy, consistent workloads like daily ETL pipelines or BI dashboards with all-day traffic typically favour a provisioned cluster with Reserved Nodes. This guide breaks down every cost component and gives you worked examples so you can make a concrete decision.
How Redshift pricing works
Provisioned cluster pricing
A provisioned Redshift cluster is a fixed set of nodes you rent by the hour. You are billed for every node-hour the cluster is running, regardless of query activity. The two main variables are the node type and the number of nodes.
There are two current node families. DC2 nodes store data on local SSDs attached to the compute node. Storage is included in the hourly price but cannot be scaled independently of compute. RA3 nodes separate compute from storage: compute is billed per node-hour, and data lives in Redshift Managed Storage (RMS), an S3-backed layer that scales independently. RA3 is the recommended choice for any dataset likely to grow beyond a few terabytes.
DC2 is a laptop with a fixed internal SSD: fast, but when the drive is full, you need a bigger laptop. RA3 is a laptop connected to an unlimited cloud drive: you pay separately for storage, but you never have to buy a bigger machine just to get more space.
| Node type | Storage | Approx. on-demand (us-east-1) | Best for |
|---|---|---|---|
| dc2.large | 160 GB SSD (included) | ~$0.25/node-hr | Small datasets up to ~2 TB compressed |
| dc2.8xlarge | 2.56 TB SSD (included) | ~$4.80/node-hr | Larger fixed-storage workloads |
| ra3.xlplus | RMS (billed separately) | ~$1.086/node-hr | Datasets where storage exceeds DC2 limits |
| ra3.4xlarge | RMS (billed separately) | ~$3.26/node-hr | Mid-to-large warehouses |
| ra3.16xlarge | RMS (billed separately) | ~$13.04/node-hr | Enterprise-scale warehouses |
Node prices above are approximate figures for us-east-1. Always verify current on-demand and Reserved Node rates on the official AWS Redshift pricing page before committing to a node type or cluster size.
Reserved Nodes for provisioned clusters
If you plan to run a provisioned cluster consistently for a year or more, Reserved Nodes reduce your effective hourly rate in exchange for a term commitment. You commit to a specific node type and region, and AWS bills you for the reservation whether or not the cluster is actively running. See AWS Pricing Models for a broader explanation of how commitment-based pricing works across AWS services.
| Commitment | Payment option | Approximate discount vs on-demand |
|---|---|---|
| 1-year | No upfront | ~20% |
| 1-year | Partial upfront | ~25–30% |
| 1-year | All upfront | ~30–35% |
| 3-year | All upfront | ~50–60% |
If you buy a 1-year reservation and then pause the cluster for three months, you still pay the reserved fee for those three months. A reservation is a financial commitment, not a usage plan. Don’t buy Reserved Nodes until you have weeks of real on-demand usage data showing consistent, heavy demand. Also note: a reservation locks to a specific node type and region. A dc2.large reservation in us-east-1 does not apply to dc2.large in eu-west-1 and does not transfer to a different node type.
Redshift Serverless pricing
Redshift Serverless charges by the RPU-hour: the number of Redshift Processing Units multiplied by the hours they were active. Charges accumulate only while queries are executing. When no queries are running, compute cost is zero.
RPU-hours are like a taxi fare. You only pay while the cab is moving (a query is running). The moment you get out (the query finishes), the meter stops. A provisioned cluster is more like leasing a car: the monthly payment is due whether you drove it or not.
- RPU-hour rate: $0.375/RPU-hour in us-east-1, billed per second with a 60-second minimum per query
- RPU range: 8 RPUs by default, scaling up to your configured maximum cap per query
- Minimum charge per query: 60 seconds at whatever RPU level the query ran. A query completing in 5 seconds still costs 60 seconds of RPU time.
Redshift Serverless has no default cost ceiling. A single analyst running a poorly written query against a large unoptimised table can trigger hundreds of RPUs and a large bill in minutes. Always configure a maximum RPU limit before anyone starts querying. Setting it to something sensible like 32 or 64 RPUs gives you a meaningful ceiling without sacrificing performance for normal workloads.
Serverless Reservations are available for teams with predictable Serverless usage. Committing for 1 year saves approximately 24% off on-demand RPU-hour rates. A 3-year commitment saves approximately 45%. These are separate from Reserved Nodes, which only apply to provisioned clusters.
If you’re new to Redshift, read the Amazon Redshift Overview before making pricing decisions. Understanding how the service works makes cost estimates significantly more accurate.
Storage, snapshots, transfer, and hidden costs
Redshift Managed Storage (RA3 and Serverless)
RA3 nodes and Redshift Serverless both store data in Redshift Managed Storage, which sits in S3 with an SSD cache on the compute layer for frequently accessed data. You pay approximately $0.024/GB-month for all data in RMS. This lets you scale storage independently of compute, which is useful when your dataset grows faster than your query load does.
DC2 storage
DC2 nodes include local SSD storage in the hourly price. There is no separate storage charge, but you cannot scale storage without adding nodes. If your dataset outgrows the per-node SSD limit, you must add more nodes, which also adds compute you may not need. At that point, migrating to RA3 is usually the better financial decision.
Automated snapshots
Redshift automatically takes incremental snapshots and stores them in S3. You get free snapshot storage up to the size of your provisioned cluster’s total storage. Most teams stay well within the free tier, but high-churn datasets with many inserts and deletes can accumulate snapshot storage faster than expected.
Redshift Spectrum
Redshift Spectrum lets you query data directly in S3 without loading it into Redshift first. This is useful for querying large historical datasets or data lake files alongside your warehouse data. The charge is $5.00 per terabyte of data scanned in S3, on top of normal S3 storage and request fees.
Spectrum scans entire files to find the rows your query needs. A 5 TB uncompressed CSV file costs $25 every time that query runs. At daily frequency, that is $750/month from a single report. Converting to Parquet reduces the scan volume by 75-80% and brings the cost down to roughly $6.25 per run. If you use Spectrum regularly, format conversion is the single most impactful cost change you can make. See Redshift Performance Optimisation for more on query tuning with Spectrum.
Data transfer
- S3 COPY into Redshift (same region): Free. You only pay S3 GET request fees.
- Cross-region COPY: Standard S3 cross-region data transfer rates apply on top of the COPY operation.
- Query results to client (same region): Free. Cross-region or internet egress incurs standard AWS data transfer charges.
- Data sharing across accounts (same region): Free. Different regions incur standard transfer costs.
Provisioned vs Serverless: side-by-side comparison
| Provisioned | Serverless | |
|---|---|---|
| Pricing unit | $/node-hour, billed continuously unless paused | $/RPU-hour, billed only during active query execution |
| Idle cost | Full node-hour charge. Significant if the cluster is mostly idle. | Zero compute cost when no queries are running |
| Cost predictability | Predictable. Fixed nodes mean a fixed hourly rate. | Variable. Depends on query volume and RPU auto-scaling. |
| Best for | Consistent, high-volume workloads (6+ hours/day active) | Sporadic or bursty workloads (a few hours/day or less) |
| Scaling behaviour | Manual. You resize or add nodes when needed. | Automatic. Scales RPUs per query up to your cap. |
| Commitment discounts | Reserved Nodes (1-year or 3-year) | Serverless Reservations (1-year or 3-year) |
| Common mistake | Leaving the cluster running 24/7 when usage is part-time | No max RPU cap set, so one large query spikes the bill |
| When it gets expensive | Large node count with low query utilisation | Heavy daily query hours where cost per hour exceeds provisioned |
When to use a provisioned cluster
Provisioned clusters are the right choice when you have consistent, predictable query load that justifies dedicated compute running for most of the day. Before provisioning, review the Redshift architecture guide to understand how node type and count affect both performance and cost.
Daily ETL pipelines running for several hours. If your data engineering team runs large ETL or ELT transformation jobs every morning for 4 to 8 hours, provisioned compute is almost always cheaper than Serverless at that scale.
BI dashboards with all-day query traffic. A Tableau or Looker instance hitting Redshift throughout the business day means the cluster is rarely idle. You get good utilisation on provisioned nodes, and the fixed cost is easy to budget for.
Large datasets that benefit from local SSD performance. DC2 nodes perform extremely well for datasets that fit within the node storage limits and where query latency is critical.
Cost predictability matters more than idle-time savings. Reserved Nodes give a fixed, predictable monthly cost. Some teams prefer this over Serverless’s variable billing even when it isn’t the absolute cheapest option.
You need fine-grained cluster configuration. Provisioned clusters expose distribution keys, sort keys, and workload management (WLM) queue settings that Serverless handles automatically. If you want manual control over these, provisioned is the right choice.
Estimate your monthly query hours before committing. If your team realistically uses the warehouse for 5 or more hours every working day, provisioned almost certainly costs less. Below that, start with Serverless and migrate later with real usage data to back the decision.
When to use Redshift Serverless
Serverless removes cluster management entirely and is cost-effective when your usage pattern is intermittent, unpredictable, or just getting started.
Ad-hoc analytics with irregular timing. Analysts who run queries a few times a day at random times pay only for those minutes. A provisioned cluster charges for every idle hour between sessions.
Development and test environments. A dev warehouse used for a few hours per week has near-zero compute cost on Serverless, while a provisioned dev cluster bills 24/7 unless someone remembers to pause it.
New workloads where query volume is unknown. Starting on Serverless avoids over-provisioning while you learn actual usage patterns. Once you have real usage data, you can decide whether to migrate to provisioned. See Estimating AWS Costs for how to model those numbers before switching.
Application-triggered queries. If queries are fired by user actions or API calls rather than scheduled jobs, Serverless billing aligns directly with actual demand: the warehouse scales up when it’s needed and costs nothing when it isn’t.
Teams who don’t want to manage cluster sizing. Serverless auto-scales RPUs per query. You don’t pick a node type, decide on node count, or plan resize operations as your data grows.
Once you have Serverless running, Running Your First Query is a practical walkthrough of querying data in Redshift from scratch.
Common beginner mistakes
Running a provisioned cluster 24/7 when usage is part-time. A 2-node dc2.large cluster left running overnight and on weekends costs approximately $360/month on-demand, even when no queries run outside business hours. Either pause the cluster on a schedule or switch to Serverless if usage is genuinely sporadic. See Redshift Cost Optimisation for a full pause/resume automation example.
Not setting a max RPU cap on Serverless. Redshift Serverless scales RPUs automatically for large or concurrent queries. Without a cap, a single runaway query can consume hundreds of RPUs and generate a large bill before you notice. Always configure a maximum RPU limit that reflects your cost ceiling before anyone starts querying.
Choosing DC2 nodes for a dataset that will grow. DC2 storage is local and fixed per node. When you hit the storage limit, you must add more nodes, which also adds compute you may not need. RA3 nodes let you scale storage independently, which is almost always the better long-term choice for growing datasets.
Scanning Spectrum data in uncompressed CSV format. Spectrum charges $5/TB scanned. Querying a 5 TB CSV file costs $25 per query. The same data in Parquet at 75% compression is about 1.25 TB, costing $6.25 per query and saving $18.75 every time that query runs. At daily frequency, that is over $560/month saved on a single query.
Buying Reserved Nodes before validating actual usage. A Reserved Node commits you to paying for that node for 1 to 3 years, running or not. If your workload turns out lighter than expected, or if you switch to Serverless, you still pay the reserved fee. Run on-demand for a few weeks first, then commit based on real usage data. Use Cost Explorer to analyse your actual usage patterns before purchasing.
Ignoring data transfer costs in multi-region setups. Loading data from S3 in the same region as Redshift is free. Cross-region COPY operations and cross-region data sharing incur standard AWS data transfer charges that are easy to overlook when initially estimating costs.
Worked cost examples
All examples use approximate us-east-1 on-demand rates. Actual costs vary by region and change over time. Verify current figures on the AWS Redshift pricing page before making provisioning decisions. These examples illustrate the structure of costs and help you compare options, not serve as exact billing estimates.
Example 1: Light analytics workload (1 hour of queries per day)
Assumptions: Small analytics team, queries run for approximately 1 hour/day on average, 300 GB of data, 30-day month.
Option A — Redshift Serverless (8 RPUs base capacity)
- Active query time: 1 hr/day × 30 days = 30 hours/month
- Compute: 8 RPU × 30 hrs × $0.375/RPU-hr = $90.00/month
- Storage: 300 GB × $0.024/GB-month = $7.20/month
- Total: ~$97/month
Option B — Provisioned 2× dc2.large, on-demand running 24/7
- Compute: 2 nodes × 720 hrs/month × $0.25/node-hr = $360/month
- Storage: included in node price
- Total: ~$360/month
Option C — Provisioned 2× dc2.large, paused nights and weekends
- Running hours: 8 hrs/day × 22 weekdays = 176 hrs/month
- Compute: 2 nodes × 176 hrs × $0.25/node-hr = $88/month
- Storage: included in node price
- Total: ~$88/month
Serverless ($97/month) and paused-provisioned ($88/month) are very close in cost. Serverless wins on operational simplicity: no pause/resume automation to build and maintain. Paused-provisioned is slightly cheaper, but it requires a scheduling mechanism and is vulnerable to someone forgetting to pause it.
Example 2: Business hours data warehouse (8 hours/day, weekdays)
Assumptions: Active data warehouse queried heavily during business hours, approximately 8 hours/day on weekdays, 1 TB of data, 22 working days per month.
Option A — Redshift Serverless (8 RPUs base capacity)
- Active query time: 8 hrs/day × 22 weekdays = 176 hours/month
- Compute: 8 RPU × 176 hrs × $0.375/RPU-hr = $528/month
- Storage: 1,000 GB × $0.024/GB-month = $24/month
- Total: ~$552/month
Option B — Provisioned 2× dc2.large, on-demand running 24/7
- Compute: 2 nodes × 720 hrs/month × $0.25/node-hr = $360/month
- Storage: included in node price
- Total: ~$360/month
Option C — Provisioned 2× dc2.large, 1-year Reserved (approx. 30% discount)
- Effective rate: ~$0.175/node-hr (30% off $0.25)
- Compute: 2 nodes × 720 hrs/month × $0.175/node-hr = ~$252/month
- Note: reserved fee applies whether the cluster is running or not
- Storage: included in node price
- Total: ~$252/month
At heavy daily usage, provisioned is significantly cheaper than Serverless. On-demand provisioned saves roughly $192/month over Serverless. Reserved Nodes save another ~$108/month on top. The tradeoff: Reserved Nodes lock you in for a year and you pay regardless of whether the cluster runs. Validate actual query hours first using Cost Explorer before committing.
Example 3: Redshift Spectrum — CSV vs Parquet scan cost
Assumptions: A nightly report queries 5 TB of data in S3 via Redshift Spectrum, running every day for a month.
- Uncompressed CSV (5 TB): 5 TB × $5.00/TB = $25.00 per query run
- Parquet format (~75% compression = 1.25 TB): 1.25 TB × $5.00/TB = $6.25 per query run
- Monthly saving (30 runs): ($25.00 − $6.25) × 30 = $562.50/month saved
The one-time cost of converting files to Parquet pays back within days for workloads like this. See Loading Data into Redshift for guidance on ingestion formats and the COPY command.
Pausing a provisioned cluster to eliminate idle charges
A paused provisioned cluster stops incurring compute charges immediately. Storage charges continue during the pause. Pausing and resuming takes 2 to 5 minutes. You can trigger it from the console, schedule it with EventBridge, or run it via the CLI:
# Pause the cluster -- compute billing stops immediately
aws redshift pause-cluster --cluster-identifier my-cluster
# Resume the cluster when ready to query again
aws redshift resume-cluster --cluster-identifier my-clusterA cluster running only during business hours (approximately 176 hours/month) saves about 75% on compute costs compared to running 24/7 (720 hours/month). Pair pause automation with billing alerts so you’re notified if charges exceed expectations. This matters especially if a scheduled pause fails to trigger. For a comprehensive approach to cutting Redshift costs beyond pausing, see Redshift Cost Optimisation.
Summary
- Provisioned clusters charge per node-hour whether or not queries run. DC2 includes local SSD storage; RA3 bills storage separately via Redshift Managed Storage at ~$0.024/GB-month.
- Redshift Serverless charges $0.375/RPU-hour (us-east-1) only during active query execution. Idle time costs nothing in compute.
- Serverless has a 60-second minimum charge per query. Always set a max RPU cap to protect against runaway costs.
- Reserved Nodes (provisioned) and Serverless Reservations both offer commitment-based discounts. Verify exact percentages on the AWS pricing page before purchasing.
- Redshift Spectrum charges $5/TB scanned in S3. Converting data to Parquet or ORC cuts scan volume by 70-80% and saves money on every query.
- Rule of thumb: use Serverless for workloads under 4 hours/day of active querying. Use provisioned with Reserved Nodes for consistent heavy use of 6 or more hours/day.
- All prices vary by region. Always verify on the AWS Redshift pricing page before provisioning.
Frequently asked questions
How much does Redshift Serverless cost per month?
It depends entirely on how many hours your queries actually run. Redshift Serverless charges $0.375 per RPU-hour (us-east-1) for active compute only. When no queries are running, compute cost is zero. A workload querying for 2 hours per day at 8 RPUs costs roughly 8 × 2 × $0.375 × 30 = $180/month in compute, plus storage. Rates vary by region, so verify current pricing on the official AWS Redshift pricing page.
Is Redshift Serverless always cheaper than provisioned?
No. Serverless is cheaper for sporadic or low-usage workloads because you only pay during active queries. But if your team queries the warehouse for 6 or more hours per day, a provisioned cluster with Reserved Nodes will typically cost significantly less. The crossover point depends on your query hours and how many RPUs each query consumes.
What is an RPU in Redshift Serverless?
RPU stands for Redshift Processing Unit. It is the unit of compute capacity in Redshift Serverless, combining CPU and memory. You set a base capacity (8 RPUs by default) and a maximum RPU cap. Redshift scales automatically within that range per query. You are billed per RPU-hour of actual usage, with a 60-second minimum charge per query regardless of how fast it completes.
Does Redshift charge when idle?
It depends on the mode. A provisioned cluster charges per node-hour whether or not queries are running. You pay for idle time unless you explicitly pause the cluster. Redshift Serverless charges only when queries are actively executing, so idle time incurs zero compute cost. This is the main reason Serverless is cheaper for sporadic workloads.
Are Reserved Nodes worth it for Redshift?
Reserved Nodes make financial sense when you run a provisioned cluster consistently for a year or more. They lock to a specific node type and region, and you pay the reserved fee whether or not the cluster is running. For unpredictable workloads they can backfire. Redshift Serverless has a separate commitment option called Serverless Reservations for teams on the serverless model who want predictable discounts.