Cloud Cost Optimization: Practical Strategies That Actually Work

The $180K Data Transfer Bill Nobody Saw Coming

A Series B startup I was consulting for got an AWS bill for $412,000 in March when the previous monthly average had been $230,000. Cost Explorer showed the usual compute and RDS lines, unchanged. The culprit, two days into the investigation, was a single EventBridge rule shipping every production log line to a newly launched observability SaaS in a different region. $180,000 of cross-region data transfer in thirty days, charged at $0.02 per GB, because someone had shipped a "fire-and-forget" log exporter on a Friday afternoon.

That is the shape of cloud waste in 2026. It is not the instance that was slightly too big -- it is the line item in the bill that nobody reads because the category looks boring. Across the dozen or so cost audits I have run, the pattern holds: the first 20 percent of savings comes from turning off things people forgot they provisioned. The next 20 percent comes from rightsizing after you measure. The last 10 percent comes from commitment discounts, and only after the first two rounds because committing to oversized capacity is how you buy three years of waste.

This guide is the playbook in that order -- quick wins first, then rightsizing, then Reserved Instances and Savings Plans, then the boring line items (storage tiering, data transfer) that quietly eat mid-size budgets. Each section has the actual numbers so you can estimate what your bill should drop to.

Quick Wins: First Week Savings

Step 1: Find and Kill Idle Resources

Run through this checklist. Every team has at least three of these:

• Unattached EBS volumes -- left behind after terminated instances. They cost money until you delete them.
• Idle Elastic IPs -- AWS charges for EIPs not attached to running instances.
• Old snapshots -- EBS snapshots accumulate silently. Set lifecycle policies.
• Dev/staging environments running 24/7 -- schedule them to shut down after hours. That alone saves 65% on non-production compute.
• Unused NAT Gateways -- $32/month each, plus data processing fees.
• Overprovisioned RDS instances -- that db.r5.2xlarge running at 5% CPU should be a db.r5.large.

Step 2: Enable Cost Allocation Tags

You can't optimize what you can't attribute. Tag every resource with at minimum:

• Environment (production, staging, dev)
• Team or Owner
• Service or Application

Enable these as cost allocation tags in AWS Billing. Now your Cost Explorer reports show spend per team, per environment, per service. Teams that see their own costs make different decisions.

Step 3: Set Up Billing Alerts

Create AWS Budget alerts at 50%, 80%, and 100% of your expected monthly spend. Also create anomaly detection alerts in Cost Explorer -- they catch sudden spikes from misconfigured resources or runaway processes before they become $10,000 surprises.

Rightsizing: Pay for What You Use

Rightsizing means matching instance sizes to actual utilization. It's the highest-impact optimization for most teams.

1. Collect metrics -- use CloudWatch (CPU, memory, network) or a third-party tool for at least two weeks of data.
2. Identify candidates -- any instance consistently below 40% CPU and 60% memory is likely oversized.
3. Downsize one step -- move from xlarge to large, or from large to medium. Monitor for a week.
4. Consider Graviton -- ARM-based instances (c7g, m7g, r7g) offer 20-40% better price-performance than x86 equivalents. Most Linux workloads run on Graviton without code changes.

Pro tip: AWS Compute Optimizer analyzes your CloudWatch data and recommends specific instance types and sizes. It's free and surprisingly accurate. Enable it account-wide and review recommendations monthly.

Commitment Discounts: Reserved Instances and Savings Plans

Option	Discount	Flexibility	Commitment
On-Demand	0%	Full	None
Savings Plans (Compute)	Up to 66%	Any instance family, region, OS, tenancy	1 or 3 years
Savings Plans (EC2 Instance)	Up to 72%	Specific instance family in a region	1 or 3 years
Reserved Instances (Standard)	Up to 72%	Specific instance type, AZ, tenancy	1 or 3 years
Reserved Instances (Convertible)	Up to 66%	Can change instance type within family	1 or 3 years
Spot Instances	Up to 90%	Full (but interruptible)	None

For most teams, Compute Savings Plans are the right default. They apply automatically to EC2, Fargate, and Lambda usage across any region. Start by committing to your baseline -- the minimum compute you know you'll use for the next year. Leave headroom for growth and use on-demand for the variable portion.

Watch out: Don't buy Reserved Instances or Savings Plans before rightsizing. Committing to oversized instances locks in waste for 1-3 years. Rightsize first, let the new baseline stabilize for a month, then commit.

Storage Optimization

S3 Lifecycle Policies

S3 offers multiple storage classes at different price points. A lifecycle policy automatically transitions objects through them:

Storage Class	Cost (per GB/month)	Use Case
S3 Standard	$0.023	Frequently accessed data
S3 Infrequent Access	$0.0125	Data accessed less than once per month
S3 Glacier Instant	$0.004	Archive with millisecond retrieval
S3 Glacier Flexible	$0.0036	Archive with minutes-to-hours retrieval
S3 Glacier Deep Archive	$0.00099	Long-term archive, 12-hour retrieval

{
  "Rules": [
    {
      "ID": "ArchiveOldLogs",
      "Status": "Enabled",
      "Transitions": [
        { "Days": 30, "StorageClass": "STANDARD_IA" },
        { "Days": 90, "StorageClass": "GLACIER" },
        { "Days": 365, "StorageClass": "DEEP_ARCHIVE" }
      ],
      "Expiration": { "Days": 730 }
    }
  ]
}

S3 Intelligent-Tiering automates transitions based on access patterns for $0.0025/1,000 objects monitored. It's worth it for unpredictable access patterns.

For reference: cloud cost optimization is the continuous process of reducing cloud spend while maintaining or improving performance and reliability. It is not a quarterly project -- it is rightsizing, commitment coverage, and waste elimination built into the engineering culture so the bill tracks intentional usage, not default sprawl.

Data Transfer: The Hidden Cost

Data transfer is the most overlooked cost on cloud bills. The pricing is asymmetric and confusing:

• Data in -- free (AWS wants your data)
• Data out to internet -- $0.09/GB (first 10 TB/month), dropping with volume
• Cross-AZ -- $0.01/GB each direction ($0.02 round trip)
• Cross-region -- $0.02/GB
• To CloudFront -- free (but CloudFront charges for distribution)

Strategies to reduce data transfer costs:

1. Use VPC endpoints for S3 and DynamoDB -- they're free and avoid NAT Gateway data processing charges
2. Keep compute and storage in the same AZ when possible
3. Use CloudFront for frequently accessed content -- it's cheaper per GB than direct S3 egress
4. Compress data before transfer -- gzip or zstd can reduce transfer volumes by 60-80%

Tools for Cost Management

Tool	Type	Starting Price	Best For
AWS Cost Explorer	Native	Free	Basic cost analysis and forecasting
AWS Compute Optimizer	Native	Free	Rightsizing recommendations
Infracost	Open source	Free / $50+/mo	Cost estimates in Terraform PRs
CAST AI	SaaS	Free tier / usage-based	Kubernetes cost optimization and autoscaling
Vantage	SaaS	Free tier / $50+/mo	Multi-cloud cost reporting and recommendations
Kubecost	Open source	Free / Enterprise	Kubernetes cost allocation per namespace/pod

Kubernetes Cost Optimization: The Separate Battle

A Kubernetes cluster hides cost from the bill. You see one EKS line item and a pile of EC2 instances, but not which namespace or team is using them. Three optimisations specific to Kubernetes move the needle more than generic rightsizing.

1. Node-level autoscaling: Cluster Autoscaler polls ASGs; Karpenter calls the EC2 Fleet API directly and bin-packs across instance types. On the three EKS clusters I migrated, Karpenter cut compute costs 28-35 percent purely through better bin-packing and ARM Graviton selection.
2. Pod-level rightsizing: every pod has a requests.cpu and requests.memory. Most teams set these once at project kickoff and never revisit. Use Vertical Pod Autoscaler in Off mode to generate recommendations, then commit them as PR-reviewed changes. Typical saving: 20-40 percent of node count.
3. Namespace-level cost attribution: install Kubecost or OpenCost, tag every namespace with team and product, and show each team their monthly spend in a Slack digest. Behaviour changes within two billing cycles -- more than any tool alone does.

Migration Walkthrough: Reserved Instance Refresh

A team finishing a three-year RI commitment in a quarter needs a migration plan, not a last-minute panic. This is the sequence I have used twice without a single coverage gap.

1. 90 days before expiry: pull the RI Utilization and Coverage reports. Identify families that are over-covered (idle RIs) vs under-covered (on-demand spend).
2. 75 days before: run a fresh rightsizing pass. Never re-up an RI for an instance type you no longer use. Compute Optimizer recommendations feed directly into this.
3. 60 days before: decide Savings Plans vs RIs. Compute Savings Plans are the default for most workloads. EC2 Instance SPs only win if you have inflexible instance family requirements.
4. 45 days before: buy the new commitment in two tranches -- 60 percent of the forecast in tranche one, 40 percent thirty days later. This hedges forecast error.
5. Day of expiry: let the old RIs expire cleanly. Monitor the On-Demand spend line for 48 hours to catch any commitment gap, and adjust tranche two if needed.

Watch out: never stack 3-year RIs for workloads you are not certain will still exist in 3 years. The discount is larger but the lock-in is real -- I have seen teams paying for Reserved Instances of a workload they decommissioned 18 months prior because the RIs were non-convertible. Compute Savings Plans are a better hedge for anything that might change.

Frequently Asked Questions

Make Cost a First-Class Metric

Cost optimization isn't a one-time project. Build it into your engineering culture. Show teams their cloud spend on dashboards next to performance metrics. Include cost estimates in pull requests with Infracost. Review the bill monthly as a team. The goal isn't minimal spending -- it's intentional spending, where every dollar maps to a business outcome. Start with the quick wins, rightsize before committing, and measure everything.