Articles

Cloud Cost Savings: Beyond the Basics

March 24, 2026 7 min read

The most surprising thing about cloud cost optimization is that the most effective strategies often involve spending more money upfront.

Let’s look at a typical scenario: a web application running on EC2 instances behind an Application Load Balancer (ALB). Imagine you’ve got a fleet of m5.large instances, always running, and you’re paying on-demand rates. Your monthly bill is $5,000.

Here’s how we might optimize:

1. Right-Sizing Instances

  • Diagnosis: Use CloudWatch metrics for CPU utilization, memory utilization (if you have the CloudWatch agent installed), network in/out, and disk I/O. Look for instances that are consistently below 50% CPU utilization or have high memory usage but low CPU.
  • Command:
    aws ec2 describe-instances --query 'Reservations[*].Instances[*].{InstanceId:InstanceId,InstanceType:InstanceType,CPU:State.Name,CPUutil:Metrics[?MetricName==`CPUUtilization`].Average[]|[0].Value|[0],NetworkIn:Metrics[?MetricName==`NetworkIn`].Average[]|[0].Value|[0],NetworkOut:Metrics[?MetricName==`NetworkOut`].Average[]|[0].Value|[0]}' --filters "Name=instance-state-name,Values=running"
    (Note: This command is a simplified example; actual metric retrieval often involves aws cloudwatch get-metric-statistics with specific StartTime, EndTime, and Period.)
  • Fix: If you find m5.large instances (2 vCPU, 8 GiB RAM) that are only using 1 vCPU and 4 GiB RAM, consider changing them to t3.medium (2 vCPU, 4 GiB RAM) or t3.small (2 vCPU, 2 GiB RAM) if memory is also underutilized.
  • Why it works: You’re paying for compute and memory resources you aren’t using. Downsizing to a smaller instance type with the same or fewer resources directly reduces your hourly cost. A t3.medium might cost $0.0416/hour compared to an m5.large at $0.096/hour, saving $1,300/month for a single instance running 24/7.

2. Reserved Instances (RIs) / Savings Plans

  • Diagnosis: Identify stable, predictable workloads. Look at your EC2 usage over the last 30-90 days. If you consistently use a certain number of instances of a specific type (e.g., 10 m5.large instances 24/7), you have a prime candidate.
  • Fix (RIs): Purchase a 1-year or 3-year EC2 Reserved Instance for m5.large instances in your region. For example, a 1-year m5.large Standard RI in us-east-1 might cost around $0.045/hour on-demand equivalent, a 53% discount from the $0.096/hour on-demand rate.
  • Fix (Savings Plans): Commit to a certain hourly spend for compute usage. A 1-year Compute Savings Plan might offer a 50% discount, meaning a $1/hour commitment effectively covers $2/hour of eligible on-demand usage. For our $5,000/month bill, we might commit to ~$2,000/month in Savings Plans.
  • Why it works: You’re pre-paying for capacity or committing to a long-term usage level. Cloud providers offer significant discounts (up to 70%+) in exchange for this commitment, as it guarantees them revenue and helps them manage their infrastructure utilization.

3. Spot Instances

  • Diagnosis: Identify fault-tolerant, stateless workloads that can handle interruptions. This includes batch processing, CI/CD build jobs, rendering, or certain big data analytics tasks.
  • Fix: Configure your application or auto-scaling group to launch EC2 instances from the Spot market. For example, instead of launching m5.large on-demand, launch them as Spot instances. Spot prices fluctuate but can be as low as 90% off on-demand rates. A Spot m5.large might cost $0.02/hour.
  • Why it works: You’re bidding on spare compute capacity. When the capacity is needed by on-demand users, your Spot instance can be terminated with a 2-minute warning. This is extremely cost-effective for workloads that can tolerate such interruptions.

4. Auto Scaling and Elasticity

  • Diagnosis: Examine your instance usage patterns over time. Are instances consistently running at low utilization during off-peak hours (nights, weekends)?
  • Fix: Implement an Auto Scaling Group with a scaling policy based on metrics like CPU utilization or request count. For example, set a policy to scale down to 2 instances when average CPU is below 30% and scale up to 10 instances when it’s above 70%. Schedule scaling to reduce instances during predictable low-traffic periods.
  • Why it works: You only pay for the compute resources you need when you need them. Instead of paying for 10 instances 24/7, you might pay for 10 instances 8 hours a day and 2 instances 16 hours a day, dramatically reducing costs.

5. Storage Optimization

  • Diagnosis: Analyze your EBS volume usage. Are there volumes attached to terminated instances? Are you using gp2 volumes when gp3 would be more cost-effective or performant? Are you archiving data that doesn’t need immediate access?
  • Fix (EBS): For gp2 volumes, performance scales with size. For gp3, you can provision IOPS and throughput independently from size. A gp3 volume might offer better performance at a lower cost than a gp2 volume of equivalent size. Identify and delete unattached EBS volumes.
  • Fix (S3): Move infrequently accessed data to S3 Glacier or Glacier Deep Archive. For example, moving 1TB of data from S3 Standard ($0.023/GB/month) to S3 Glacier Flexible Retrieval ($0.004/GB/month) saves $19/month per TB.
  • Why it works: You’re paying for storage tiers and configurations that match your access patterns and performance needs. Using the cheapest viable storage for data you rarely touch, and ensuring you aren’t paying for orphaned disks, directly reduces storage bills.

6. Data Transfer Costs

  • Diagnosis: Examine your VPC Flow Logs or network traffic monitoring tools for significant data egress to the internet or other AWS regions.
  • Fix: Use VPC Endpoints for services like S3 and DynamoDB. This allows your EC2 instances to access these services privately within the AWS network, avoiding NAT Gateway charges and internet data transfer costs. If transferring data between regions, consider using AWS Direct Connect or transferring during off-peak hours if applicable, though inter-region transfer is often unavoidable for disaster recovery.
  • Why it works: Data transfer out of AWS or between regions incurs costs. By keeping traffic within the AWS network where possible, you eliminate these charges.

The "Spend More to Save More" Paradox

The most common "aha!" moment in cost optimization is realizing that committing to Reserved Instances or Savings Plans often requires an upfront or long-term financial commitment that looks like increased spending. However, the massive discounts (often 50-70%) applied to your baseline, predictable usage far outweigh the upfront cost, leading to significant net savings over time. For our $5,000/month bill, a $1,500/month commitment to Savings Plans could easily bring the total down to $3,500-$4,000/month, saving $1,000-$1,500 monthly.

The next challenge is managing the dynamic aspects of cost, like understanding how ephemeral workloads and fluctuating traffic patterns impact your bill and how to optimize for them.

Want structured learning?

Take the full Cloud Computing course →