Skip to main content
Cloud DevOps Solutions
Get a quote Talk to an engineer
Cloud Cost 2026-11-12 16 min read

How to Reduce AWS Costs: A Complete Guide for Engineering Teams

Saddique Khan

DevOps / SRE Engineer & CEO, Cloud DevOps Solutions

AWS pricing is designed to be flexible. That same flexibility makes it easy to overspend. A team that provisions generously to avoid performance issues ends up paying for capacity it never uses. A data pipeline that moves data across regions adds thousands in transfer fees. A developer spins up a GPU instance for a quick test and forgets to terminate it.

The result: most AWS accounts are spending 30-40% more than they need to. Not because AWS is overcharging, but because the default path, On-Demand pricing with generous provisioning, is the most expensive way to run.

This guide covers the strategies that consistently deliver the largest savings, from quick wins you can implement this week to structural changes that compound over months.

1. Start with visibility: know where the money goes

You cannot reduce costs you cannot see. Before optimising anything, set up the baseline:

AWS Cost Explorer

Enable Cost Explorer (it is free) and create saved views for:

  • Monthly spend by service: identifies which services dominate your bill
  • Daily spend trend: catches anomalies early
  • Spend by linked account: shows which team or environment costs the most
  • Spend by tag: maps costs to business units, products, or environments (requires tagging, covered in section 11)

Cost and Usage Report (CUR)

The CUR is the most granular billing data AWS provides. It contains every line item, including resource IDs, usage types, and pricing details. Export it to S3, query it with Athena, and build dashboards in QuickSight or Grafana.

For teams that want a simpler setup, cloud cost optimisation tools like Vantage or Kubecost provide ready-made dashboards without the Athena setup.

AWS Cost Anomaly Detection

Enable Cost Anomaly Detection (free) to get alerts when spending deviates from normal patterns. Configure it to monitor by service, linked account, or cost allocation tag. Route alerts to Slack or email via SNS.

2. Right-size EC2 instances

EC2 compute is typically 40-60% of an AWS bill. Most instances are oversized because teams provision for peak load and never revisit.

Finding oversized instances

AWS Compute Optimizer analyses 14 days of CloudWatch metrics (CPU, memory via CloudWatch Agent, network, disk) and recommends optimal instance types. It is free and requires no agent for CPU and network metrics.

AWS Trusted Advisor flags instances with average CPU below 10% over 14 days.

The right-sizing process

  1. Export Compute Optimizer recommendations
  2. Filter to instances where the recommended size is smaller and the projected performance risk is “very low” or “low”
  3. Group by application or team for batch changes
  4. Resize during maintenance windows (requires a stop/start for most changes)
  5. Monitor for 7 days after resizing to confirm no performance degradation

Graviton migration

AWS Graviton processors (arm64) deliver up to 40% better price-performance than equivalent x86 instances. The migration path:

  • Containerised workloads: build multi-arch images, deploy to Graviton node groups. Most applications work without code changes.
  • Managed services: RDS, ElastiCache, and OpenSearch all support Graviton-based instance types. Switch during a maintenance window.
  • EC2 directly: test your application on a Graviton instance. Most Linux workloads, Java, Python, Node.js, Go, and .NET 6+ run without modification.

A migration from m5.xlarge ($140/month) to m7g.xlarge ($118/month) saves 16% with equal or better performance. Across a fleet of 50 instances, that is $1,100/month.

3. Use Savings Plans and Reserved Instances

On-Demand pricing is the rack rate. Commitment-based pricing is where the real savings are.

Savings Plans

Savings Plans come in two types:

  • Compute Savings Plans: apply to any EC2, Fargate, and Lambda usage across all regions. The most flexible option. Savings: 20-40% for 1-year, 40-60% for 3-year.
  • EC2 Instance Savings Plans: locked to a specific instance family and region but offer deeper discounts. Savings: 30-45% for 1-year, 50-72% for 3-year.

How to calculate commitment

  1. In Cost Explorer, open the Savings Plans recommendations
  2. Look at your steady-state hourly spend (the flat baseline, not the peaks)
  3. Commit to 70-80% of that baseline as Savings Plans
  4. Leave 20-30% as On-Demand for burst and flexibility

Example: if your steady-state EC2 compute costs $10/hour and peaks to $15/hour during business hours, commit to $7-8/hour in Savings Plans. The remaining $2-3/hour base plus the $5 peak stays On-Demand.

Reserved Instances (RIs)

RIs still make sense for specific use cases:

  • RDS Reserved Instances: database instance types rarely change. A 1-year RI saves 30-40%; 3-year saves 55-65%.
  • ElastiCache Reserved Nodes: same logic as RDS.
  • OpenSearch Reserved Instances: if your cluster is stable.

For EC2, Savings Plans have largely replaced RIs because they are more flexible. Use RIs only when the deeper discount justifies the reduced flexibility.

4. Leverage Spot instances

Spot instances use AWS spare capacity at up to 90% discount. They can be interrupted with 2 minutes notice, but for the right workloads, the savings are transformative.

Good candidates

  • CI/CD build agents: Jenkins, GitHub Actions self-hosted runners, GitLab runners. A build that fails due to Spot interruption just retries.
  • Batch processing: EMR, AWS Batch, data pipelines. Use checkpointing so interrupted jobs resume rather than restart.
  • EKS/Kubernetes worker nodes: run non-critical workloads on Spot node groups with Karpenter or Cluster Autoscaler handling interruptions.
  • Dev/test environments: combined with auto-shutdown for maximum savings.
  • Training workloads: ML training with checkpoint-resume capability.

Spot best practices

  • Diversify instance types: request capacity across multiple instance families and sizes. Karpenter does this automatically for EKS.
  • Use capacity-optimised allocation: this strategy picks the Spot pool with the most available capacity, reducing interruption frequency.
  • Set up interruption handling: use the EC2 metadata service to detect the 2-minute warning and drain gracefully.
  • Mix with On-Demand: for EKS, run your control plane and critical pods on On-Demand, everything else on Spot.

A typical EKS cluster running 70% Spot and 30% On-Demand saves 50-60% on compute versus fully On-Demand. See our EKS cost optimisation guide for detailed techniques.

5. Optimise S3 storage costs

S3 is cheap per gigabyte, but at scale, the costs add up. Common waste patterns:

Lifecycle policies

Move data to cheaper tiers automatically based on access patterns:

TransitionUse caseSaving vs Standard
Standard to Infrequent Access (30 days)Logs, old uploads~45%
Infrequent Access to Glacier Instant Retrieval (90 days)Compliance archives~68%
Glacier Instant to Glacier Flexible (180 days)Rarely accessed backups~78%
Glacier Flexible to Glacier Deep Archive (365 days)Legal holds, long-term archives~95%

S3 Intelligent-Tiering

For data with unpredictable access patterns, S3 Intelligent-Tiering automatically moves objects between tiers based on usage. There is a small monitoring fee per object, but no retrieval charges. It is the right default for most data where you cannot predict access patterns.

Incomplete multipart uploads

Multipart uploads that fail leave orphaned parts that accumulate charges. Add a lifecycle rule to abort incomplete uploads after 7 days:

{
  "Rules": [
    {
      "ID": "abort-incomplete-uploads",
      "Status": "Enabled",
      "Filter": {},
      "AbortIncompleteMultipartUpload": {
        "DaysAfterInitiation": 7
      }
    }
  ]
}

Versioning costs

If bucket versioning is enabled (often required for compliance), old versions accumulate silently. Add lifecycle rules to expire non-current versions after a retention period (30-90 days depending on your requirements).

Storage Lens

Enable S3 Storage Lens (free tier available) for organisation-wide visibility into storage usage, activity, and cost efficiency. It identifies buckets with no lifecycle policies, high request costs relative to storage, and opportunities for tier migration.

6. Reduce data transfer costs

Data transfer is the hidden cost on many AWS bills. AWS does not charge for data coming in, but charges for data going out and between availability zones.

Cross-AZ transfer

Every cross-AZ request costs $0.01/GB in each direction ($0.02 round trip). For a busy microservices architecture making millions of cross-AZ calls, this adds up to thousands per month.

Mitigations:

  • Use topology-aware routing in Kubernetes (topology spread constraints, service topology)
  • Place tightly coupled services in the same AZ when the resilience trade-off is acceptable
  • Use VPC endpoints for AWS service traffic (S3, DynamoDB, SQS) to avoid NAT Gateway data processing charges

NAT Gateway costs

NAT Gateways charge $0.045/hour ($32/month) plus $0.045/GB processed. A busy cluster pulling container images, sending logs, and calling external APIs can easily push $500+/month through NAT.

Mitigations:

  • Use VPC endpoints for S3 and DynamoDB (free, eliminates NAT charges for those services)
  • Pull container images from ECR via a VPC endpoint
  • Cache container images with a pull-through cache or local registry mirror
  • Consider NAT instances for dev/test environments where high availability is not critical

CloudFront for egress

If you serve content to the internet, CloudFront data transfer is cheaper than direct S3 or EC2 egress in most regions. Route static assets and API responses through CloudFront even for non-cacheable content: the per-GB price is lower.

Inter-region transfer

Replicating data between regions costs $0.02/GB. Review whether cross-region replication is necessary for all buckets, or whether it can be limited to critical data. For disaster recovery, consider whether the RTO/RPO justifies continuous replication versus periodic backup.

7. Optimise RDS and database costs

Right-size instances

RDS instances follow the same pattern as EC2: teams provision for peak and never revisit. Check CloudWatch metrics:

  • CPUUtilization: if the average is below 20% and P99 below 60%, downsize
  • FreeableMemory: if consistently above 50% of total, the instance has more memory than needed
  • DatabaseConnections: if consistently below 25% of the max for that instance class

Use Aurora Serverless v2 for variable workloads

Aurora Serverless v2 scales compute in half-ACU increments based on demand. Minimum capacity can be set as low as 0.5 ACU. For databases with variable traffic (internal tools, staging environments, batch-heavy workloads), Serverless v2 eliminates paying for idle capacity.

Reserved Instances for stable databases

Production databases rarely change instance types. A db.r6g.xlarge On-Demand costs roughly $620/month. A 1-year RI reduces that to $390. A 3-year RI brings it to $250.

Read replicas for read-heavy workloads

Instead of scaling up a single instance, add read replicas and route read queries to them. This is often cheaper than a larger primary instance and improves availability.

Consider Graviton-based instances

RDS Graviton instances (db.r7g, db.m7g) are 10-20% cheaper than their x86 equivalents with comparable or better performance. The switch requires a maintenance window but no application changes.

8. Optimise Lambda and serverless costs

Lambda pricing is based on the number of invocations, execution duration, and allocated memory. Common waste:

Right-size memory

Lambda CPU scales linearly with memory allocation. A function allocated 1,024 MB that uses only 200 MB of memory is not just wasting memory; it is paying for CPU it does not use. Use AWS Lambda Power Tuning (an open-source tool) to find the optimal memory setting that minimises cost for each function’s execution profile.

Reduce execution time

  • Minimise cold starts by keeping functions warm (Provisioned Concurrency for critical paths, or architectural patterns that reuse execution environments)
  • Move initialisation code outside the handler (SDK clients, database connections)
  • Use connection pooling with RDS Proxy instead of opening a new database connection per invocation

Graviton for Lambda

Lambda functions on arm64 (Graviton2) are 20% cheaper per GB-second. Most runtimes (Python, Node.js, Java, .NET, Go) support arm64. Add "architectures": ["arm64"] to your function configuration.

Step Functions vs Lambda chaining

For workflows that chain multiple Lambda invocations, evaluate whether AWS Step Functions Express Workflows reduce cost. Express Workflows charge per state transition rather than per invocation and duration, which is cheaper for short, high-frequency workflows.

9. Automate waste removal

Manual cost reviews find savings. Automated cleanup prevents waste from accumulating between reviews.

Automated checks to implement

Waste typeDetectionAutomation
Unattached EBS volumesCloudWatch event on DetachVolumeLambda: snapshot and delete after 7 days unattached
Unused Elastic IPsTrusted Advisor checkLambda: release if unassociated for 48 hours
Old snapshotsAge-based queryLambda: delete snapshots older than retention period
Stopped instancesCloudWatch StatusCheckFailedLambda: terminate instances stopped for 30+ days
Unused load balancersALB with zero healthy targets for 7 daysLambda: alert and auto-delete after confirmation
Orphaned Lambda versionsVersion count exceeding thresholdLambda: retain only the 3 most recent versions

Implementation pattern

  1. Create an EventBridge scheduled rule (daily or weekly)
  2. Trigger a Lambda function that queries for the waste condition
  3. For low-risk actions (delete old snapshots, release unused EIPs): auto-remediate and log
  4. For higher-risk actions (terminate instances, delete volumes): create a notification and require manual approval

AWS Nuke for non-production

For sandbox and development accounts, consider aws-nuke on a schedule. It deletes all resources in an account except those you explicitly allowlist. Run it nightly on sandbox accounts to prevent weekend resource accumulation.

10. Optimise container costs on ECS and EKS

Fargate right-sizing

Fargate tasks are billed per vCPU and GB of memory per second. Over-provisioned task definitions waste money every second:

  1. Enable Container Insights to see actual CPU and memory utilisation per task
  2. Reduce task CPU/memory to match P95 utilisation plus 20% headroom
  3. Use Fargate Spot for fault-tolerant tasks (up to 70% discount)

EKS node optimisation

  • Karpenter: replaces Cluster Autoscaler with faster, more intelligent node provisioning. It selects the cheapest instance type that fits pending pods, consolidates underutilised nodes, and handles Spot interruptions.
  • Right-size pods: set resource requests to match actual usage (P95 + 10% buffer). Oversized requests waste node capacity.
  • Namespace quotas: prevent teams from requesting excessive resources in shared clusters.

See our EKS cost optimisation guide for all 12 techniques.

11. Enforce tagging and governance

Cost optimisation without governance erodes over time. Every resource should carry tags that enable cost allocation:

Required tags

  • team or owner: who is responsible
  • environment: production, staging, dev, sandbox
  • service or application: which product or service
  • cost-centre: for finance allocation

Enforcement

Use AWS Service Control Policies (SCPs) in AWS Organizations to deny resource creation without required tags. Alternatively, use AWS Config rules to detect and flag untagged resources.

Tag-based cost allocation

Activate cost allocation tags in the Billing console. Once activated, tags appear in Cost Explorer and the CUR, enabling per-team, per-service, and per-environment cost views.

Without tags, your cost data is a flat list of AWS service charges. With tags, it becomes a map of where money goes and who controls it.

Building a cost-aware culture

Tools and automation handle the mechanics. The lasting change comes from culture:

  • Share cost dashboards with engineering teams, not just finance. Engineers cannot optimise what they cannot see.
  • Include cost in architecture reviews. Every design document should have a “cost implications” section.
  • Set team-level budgets and review them in sprint retrospectives.
  • Celebrate savings. When a team reduces their bill by 30%, recognise it the same way you recognise a shipped feature.

For teams that want help building this practice from the ground up, our cloud cost optimisation service combines technical implementation with organisational change.

Key takeaways

  • Start with visibility: Cost Explorer, CUR, and anomaly detection are free and essential
  • EC2 right-sizing and Graviton migration typically deliver the largest single savings
  • Savings Plans at 70-80% of steady-state spend save 30-60% on committed workloads
  • Spot instances save up to 90% for fault-tolerant workloads; diversify instance types and handle interruptions
  • S3 lifecycle policies and Intelligent-Tiering prevent storage costs from growing unchecked
  • Data transfer is the hidden cost: VPC endpoints, topology-aware routing, and CloudFront reduce it significantly
  • Automated waste removal prevents savings from eroding between manual reviews
  • Tagging and governance make cost optimisation sustainable, not a one-off project

Spending more than you should on AWS? Book a free consultation and we will identify your top savings opportunities together.

#aws #cost-optimization #savings-plans #spot-instances #finops

Related service

Need hands-on help? Our certified engineers deliver this as a managed service.

Cloud Cost Optimisation
Found this useful? Share it with your team or link to it from your documentation.

Need help implementing this?

Our certified engineers can handle the setup, migration, or ongoing operations for you.