Cloud Cost Optimization Tips
Cloud costs are often the second-largest expense after payroll for SaaS companies. Without active management, spending grows faster than revenue. This guide covers practical cost optimization strategies that reduce bills by 30-50% without sacrificing performance.
Right-Sizing Instances
The most common waste is over-provisioned resources. Use cloud provider tools to analyze utilization:
* **AWS Compute Optimizer**: Analyzes CPU, memory, and network utilization to recommend instance types.
* **GCP Rightsizing Recommendations**: Built into the Compute Engine console.
* **Azure Advisor**: Provides cost recommendations across all services.
Target utilization rules of thumb:
| Resource | Target Utilization |
|----------|-------------------|
| CPU | 40-70% average |
| Memory | 60-80% average |
| Disk IOPS | Below 80% of provisioned |
Downsize instances that consistently run below 20% utilization. For variable workloads, consider scaling horizontally rather than vertically.
Reserved Instances and Savings Plans
Commit to usage in exchange for discounts:
| Option | Discount | Commitment |
|--------|----------|------------|
| AWS Reserved Instances | 40-60% | 1 or 3 years |
| AWS Savings Plans | 40-60% | 1 or 3 years ($/hour) |
| GCP Committed Use | 40-57% | 1 or 3 years |
| Azure Reserved | 40-60% | 1 or 3 years |
Start with 1-year commitments for baseline workloads (30-50% of your total compute). Use 3-year commitments for stable, predictable workloads. Combine Savings Plans with Spot Instances for maximum flexibility.
Spot and Preemptible Instances
Use spot instances (AWS), preemptible VMs (GCP), or low-priority VMs (Azure) for fault-tolerant workloads:
# AWS: Request spot instances in Auto Scaling
aws autoscaling create-auto-scaling-group \
--mixed-instances-policy file://spot-policy.json
# GCP: Create preemptible VM
gcloud compute instances create worker \
--preemptible
Ideal workloads: batch processing, CI/CD runners, stateless web workers, data analytics, rendering.
Savings: 60-90% compared to on-demand pricing. Combine with Spot Instance interruption handling (checkpointing, graceful shutdown).
Storage Optimization
Storage costs accumulate silently. Audit your storage regularly:
* **Delete unused volumes**: Snapshots of deleted volumes and unattached EBS volumes.
* **Use lifecycle policies**: Move infrequently accessed data to colder tiers.
* **Object storage tiers**:
| Tier | Cost/GB/Month | Use Case |
|------|--------------|----------|
| S3 Standard | $0.023 | Active data |
| S3 Infrequent Access | $0.0125 | Accessed monthly |
| S3 Glacier | $0.0036 | Archived data |
| S3 Deep Archive | $0.001 | Regulatory retention |
Set S3 Lifecycle rules to transition objects automatically:
{
"Rules": [
{
"Id": "MoveToIA",
"Filter": {"Prefix": "logs/"},
"Status": "Enabled",
"Transitions": [
{"Days": 30, "StorageClass": "STANDARD_IA"},
{"Days": 90, "StorageClass": "GLACIER"}
],
"Expiration": {"Days": 365}
}
]
}
Network Egress Costs
Data transfer out of cloud providers is expensive. Minimize egress:
* **Use the same region**: Keep services that communicate frequently in the same region. Cross-region traffic is billed.
* **CloudFront/CDN**: Serve static assets through a CDN. CloudFront data transfer to the internet is cheaper than S3 direct access.
* **Leverage direct connect**: For large data transfers, use AWS Direct Connect or equivalent.
* **NAT Gateway costs**: Use NAT instances instead of NAT Gateway for high-volume traffic (cost savings of 70-80%).
Autoscaling
Scale resources to match demand:
# AWS Auto Scaling with target tracking
autoscaling:
target_tracking:
- predefined_metric_specification:
predefined_metric_type: ASGAverageCPUUtilization
target_value: 60
For containerized workloads, use Kubernetes Cluster Autoscaler:
# Karpenter for AWS EKS
kubectl scale deployment api-server --replicas=0 # Automated idle scaling
Karpenter and similar tools scale nodes based on actual pod resource requests, eliminating node-level waste.
Database Cost Optimization
Databases are often the most expensive service:
* **Serverless databases**: Use Aurora Serverless, Cloud SQL auto-scaling, or Azure SQL Serverless for variable workloads.
* **Read replicas**: Add replicas for read-heavy workloads instead of upscaling a single instance.
* **Connection pooling**: Use PgBouncer or RDS Proxy to handle thousands of connections without provisioning for peak.
* **Delete old data**: Archive historical data to object storage.
Reserved Capacity with Spot
Combine reserved capacity for baseline with spot for spikes:
# Terraform: Mix OD and Spot in ASG
resource "aws_autoscaling_group" "app" {
mixed_instances_policy {
launch_template {
launch_template_specification { ... }
override {
instance_type = "t3.medium"
}
}
instances_distribution {
on_demand_base_capacity = 2
on_demand_percentage_above_base_capacity = 50 # Rest from Spot
spot_allocation_strategy = "capacity-optimized"
}
}
}
Monitoring and Budgets
Set up cost monitoring to catch anomalies early:
# AWS Budget with action
aws budgets create-budget \
--budget-name "Monthly-Infra" \
--budget-file monthly-budget.json \
--notifications-with-subscribers file://alert-config.json
# GCP budget alert
gcloud billing budgets create \
--billing-account=XXXXXX \
--display-name="Monthly Budget" \
--budget-amount=5000USD \
--threshold-rules=percent=50,percent=90 \
--notifications-pubsub-topic=budget-alerts
Set up alerts at 50%, 80%, and 90% of budget. Tag resources by cost center and review spending weekly.
Summary
Cloud cost optimization is an ongoing process, not a one-time cleanup. Start with right-sizing (the quickest wins), add reserved capacity for baseline workloads, use spot instances for fault-tolerant work, and configure autoscaling to match demand. Monitor storage lifecycle, minimize data egress, and optimize database tiering. The most effective approach is a 20% time investment per quarter: review spending, tag resources, and implement the top three savings opportunities. Most organizations can reduce their cloud bill by 30-50% within three months by following these practices.