# Find unattached EBS volumes
aws ec2 describe-volumes \
  --filters Name=status,Values=available \
  --query 'Volumes[*].[VolumeId,Size,CreateTime,Tags]' \
  --output table

Idle Elastic IPs: Unattached Elastic IPs cost $3.65/month each (AWS changed pricing in 2024 to charge for all public IPv4 addresses). Across a large account, dozens of orphaned EIPs add up.

Unused load balancers: ALBs cost $16.20/month minimum (in us-east-1) even with zero traffic. NLBs cost $6.75/month per AZ. Audit load balancers with zero healthy targets or zero requests over the past 30 days.

Unattached Azure Managed Disks: Same problem as EBS — disks persist after VM deletion.

# Find unattached Azure disks
az disk list --query "[?managedBy==null].[name,diskSizeGb,sku.name]" -o table

Old snapshots: EBS snapshots, Azure snapshots, and GCP disk snapshots accumulate over time. A snapshot retention policy (keep last 7 daily, 4 weekly, 12 monthly, delete the rest) prevents unbounded growth. For 10 TB of snapshots, you save roughly $500/month on AWS.

Shut Down Non-Production After Hours

Development and staging environments typically run 24/7 despite being used only during business hours (roughly 10 hours/day, 5 days/week). Scheduling these environments to run only during work hours reduces their compute cost by approximately 70%.

AWS: Instance Scheduler solution (CloudFormation template from AWS) or a Lambda function triggered by EventBridge on a cron schedule.

Azure: Auto-shutdown feature on VMs (built-in), or Azure Automation runbooks for complex scheduling.

GCP: Cloud Scheduler triggering Cloud Functions that start/stop instances.

For a team with 20 development VMs averaging $200/month each, after-hours shutdown saves approximately $2,800/month ($33,600/year).

Delete Unused Environments

Old feature branch environments, abandoned proof-of-concept deployments, and "temporary" infrastructure that became permanent. Run a quarterly audit: if a resource has had no meaningful activity (network traffic, API calls, logins) in 30 days, flag it for deletion. If no one claims it within 14 days, terminate it.

Phase 3: Right-Sizing (Weeks 3-5)

Right-sizing delivers 15-25% savings by matching resource allocation to actual workload requirements.

Compute Right-Sizing

Most organizations overprovision by 40-60%. The pattern is predictable: a developer requests an m5.2xlarge because the workload "might need it," and no one revisits the sizing after deployment.

AWS Compute Optimizer analyzes CloudWatch metrics (CPU, memory, network) and recommends optimal instance types. It is free and often recommends downsizing by 1-2 instance sizes.

# Get Compute Optimizer recommendations
aws compute-optimizer get-ec2-instance-recommendations \
  --query 'instanceRecommendations[*].[instanceArn,currentInstanceType,recommendationOptions[0].instanceType,recommendationOptions[0].projectedUtilizationMetrics]'

Graviton migration: ARM-based Graviton instances (AWS) provide approximately 40% better price-performance than equivalent x86 instances for many workloads. If your application runs on Linux and does not depend on x86-specific libraries, Graviton migration is the single highest-impact right-sizing action.

Azure Advisor provides VM right-sizing recommendations based on CPU and memory utilization. Look for VMs with average CPU under 20% — these are candidates for downsizing.

GCP recommender provides machine type recommendations. GCP's custom machine types allow precise sizing — instead of choosing between n2-standard-4 (4 vCPU, 16 GB) and n2-standard-8 (8 vCPU, 32 GB), create a custom machine with exactly 6 vCPU and 24 GB.

Database Right-Sizing

Databases are frequently overprovisioned because downtime risk makes teams conservative. However, most RDS instances run at 15-25% CPU utilization with memory to spare.

Steps: 1. Enable Performance Insights (AWS) or Query Performance Insight (Azure) to understand actual workload 2. Identify read-heavy workloads and add read replicas instead of scaling up the primary 3. Move from provisioned to serverless where available (Aurora Serverless v2, Azure SQL Serverless, Cloud SQL on GCP) 4. Evaluate whether a managed database is necessary at all — some workloads running on RDS PostgreSQL could run on DynamoDB or S3 + Athena at a fraction of the cost

Storage Right-Sizing

S3 storage analysis frequently reveals that 60-70% of stored data has not been accessed in 90+ days. S3 Intelligent-Tiering automatically moves objects between access tiers based on usage patterns, saving 40-68% on infrequently accessed data.

For manually managed storage:

Implement S3 Lifecycle policies to transition objects automatically. A 100 TB dataset where 70% is archival saves approximately $1,200/month by moving cold data to Glacier Instant Retrieval.

Phase 4: Rate Optimization (Weeks 5-8)

Once you have right-sized resources, lock in lower rates through commitment programs and alternative pricing models.

Reserved Instances and Savings Plans

AWS Savings Plans: Commit to a consistent amount of compute spend (measured in $/hour) for 1 or 3 years. Compute Savings Plans provide up to 66% discount and apply to EC2, Lambda, and Fargate across any instance family, region, or OS. EC2 Instance Savings Plans provide up to 72% discount but are locked to a specific instance family and region.

Strategy: analyze your steady-state compute usage over the past 3 months. Purchase Savings Plans covering 60-70% of that baseline (to account for optimization headroom). Keep the remaining 30-40% as on-demand for flexibility.

Azure Reserved Instances: 1-year or 3-year commitments for VMs, databases, and other services. 3-year reservations save up to 72%. Combine with Azure Hybrid Benefit for Windows workloads to reach 80%+ savings.

GCP Committed Use Discounts (CUDs): 1-year or 3-year commitments for compute and memory resources. Spend-based CUDs (similar to AWS Savings Plans) offer up to 55% for 3 years and apply to Compute Engine and GKE across any machine type.

Spot Instances

Spot instances (AWS), Spot VMs (Azure), and Preemptible/Spot VMs (GCP) use spare cloud capacity at 60-90% discount. The trade-off: they can be reclaimed with short notice (2 minutes on AWS, 30 seconds on GCP).

Workloads suitable for spot: - Batch processing (data pipelines, video encoding, ML training) - CI/CD runners (build and test jobs) - Stateless web servers behind auto-scaling groups (with on-demand fallback) - Big data analytics (Spark, EMR, Dataproc clusters)

Workloads not suitable for spot: - Databases and stateful services - Single-instance applications with no redundancy - Latency-sensitive services without graceful degradation

AWS strategy: Use mixed instance policies in Auto Scaling Groups — specify multiple instance types across multiple availability zones to maximize spot availability. Set a capacity-optimized allocation strategy instead of lowest-price.

Negotiate Enterprise Agreements

At $50,000+/month cloud spend, you qualify for custom pricing. At $100,000+/month, you get meaningful discounts. At $500,000+/month, you negotiate EDPs with committed spend in exchange for 10-25% overall discounts.

Negotiation leverage: demonstrate multi-cloud capability, commit to growth targets, consolidate spend across business units onto a single agreement, and time negotiations near the end of the cloud provider's fiscal quarter.

Phase 5: Architecture Optimization (Ongoing)

The deepest savings come from architectural changes that fundamentally reduce resource consumption.

Serverless Migration

Serverless services (Lambda, Azure Functions, Cloud Functions, Fargate, Cloud Run) charge only for actual usage. A Lambda function that processes 1 million requests/month with an average duration of 200ms costs approximately $3.00. An equivalent EC2 t3.medium running 24/7 costs $30.37.

The breakeven point is roughly 30% utilization. Below 30% average utilization, serverless is cheaper. Above 60%, provisioned compute is cheaper. Most event-driven, API-based, and scheduled workloads fall well below 30% utilization.

Data Architecture Optimization

Replace expensive real-time queries with pre-computed views. Instead of running a $50/day Athena query every hour, materialize the results into a DynamoDB table that costs $5/day to serve. Instead of keeping 2 years of log data in Elasticsearch at $3,000/month, archive logs older than 30 days to S3 and query with Athena on demand for $200/month.

Reserved Capacity Pooling

For organizations with multiple AWS accounts, use AWS Organizations to share Reserved Instances and Savings Plans across accounts. A reservation purchased in the production account automatically applies to matching usage in any account within the organization, maximizing utilization.

Building a FinOps Practice

Cost optimization is not a one-time project — it is an ongoing practice. The FinOps Foundation defines three phases: Inform (visibility), Optimize (action), and Operate (continuous governance).

Key organizational practices: - Monthly cost review with engineering leads and finance - Cost optimization targets in engineering OKRs (e.g., "reduce per-transaction infrastructure cost by 15%") - Automated anomaly detection with rapid response playbook - Cost attribution to teams with accountability for budget adherence - Right-sizing review triggered by any new deployment or significant architecture change

Citadel Cloud Management's cloud courses cover cost optimization strategies for AWS, Azure, and GCP with hands-on exercises in rightsizing, commitment planning, and architectural optimization. The Cloud Toolkits collection includes Terraform modules with cost-optimized defaults, FinOps dashboard configurations, and automated cleanup scripts.

For enterprise teams building a FinOps practice, the Enterprise Bundles provide comprehensive cost management frameworks, tagging governance templates, and executive reporting dashboards.

Ready to cut your cloud bill? Start with Citadel's free cloud courses to learn cost optimization fundamentals, then implement with production-ready toolkits. Browse all resources for FinOps frameworks, automation scripts, and enterprise cost management solutions.