Cloud bills grow silently. A fast-scaling SaaS client approached RapidStackLab with a monthly AWS invoice that had swollen to over $34,200/month across two Kubernetes clusters, multiple RDS instances, and unmonitored data transfer. In just 30 days of targeted SRE intervention, we reduced their monthly spend to $19,800/month—unlocking $172,800 in annualized savings with zero downtime.
Here is the exact technical blueprint we used, the gotchas we navigated, and how you can implement the same optimizations in your infrastructure.
Phase 1: Eliminating the NAT Gateway Tax (Saved $3,800/mo)
AWS charges $0.045 per GB for data processed through a NAT Gateway in addition to the hourly gateway fee. In this client's architecture, backend worker pods were continuously pulling gigabytes of machine learning models and assets from Amazon S3 and querying DynamoDB via public endpoints, traversing the NAT Gateway.
The Fix: We provisioned AWS VPC Gateway Endpoints for S3 and DynamoDB. Gateway endpoints are completely free of charge and route traffic directly over the internal AWS network, entirely bypassing the NAT Gateway. This single configuration reduced NAT Gateway data processing fees by over 80%.
Phase 2: Replacing Cluster Autoscaler with Karpenter (Saved $5,400/mo)
The client was using standard Kubernetes Cluster Autoscaler with rigid EC2 Auto Scaling Groups (ASGs). When a pod requested 2.5 CPU cores, the autoscaler would spin up an entire m5.2xlarge instance, leaving 5.5 cores idle.
The Fix: We deployed Karpenter, AWS's open-source node autoscaler. Karpenter groups pending pods and dynamically provisions the exact instance type needed (or combines multiple pods on a single machine). We configured Karpenter NodePools with:
- Spot instance interruption handling with automated fallback to on-demand instances.
- Consolidation policies that actively bin-pack pods and terminate underutilized nodes.
- Sub-minute node launch times compared to 4–6 minutes under traditional ASGs.
Phase 3: Graviton3 (ARM64) Migration (Saved $3,600/mo)
AWS Graviton3 instances (c7g, m7g) offer up to 25% better compute performance and 20% lower hourly cost compared to equivalent x86 (Intel/AMD) instances.
Because the client's services were containerized in Docker, we updated their GitHub Actions CI/CD pipelines to build multi-architecture Docker images (linux/amd64 and linux/arm64) using Docker Buildx. After smoke-testing on staging, we shifted 70% of production workloads to Graviton3 nodes.
Phase 4: EBS GP2 to GP3 & Snapshot Automation (Saved $1,600/mo)
Legacy EBS gp2 volumes require provisioning excess gigabytes just to achieve higher IOPS baseline performance. In contrast, gp3 volumes decouple IOPS and throughput from volume size and are 20% cheaper per GB.
We executed an automated Terraform migration to convert all persistent volumes from gp2 to gp3 with zero unmounting or downtime. Furthermore, we audited AWS Lifecycle Manager policies to prune unattached EBS volumes and snapshots older than 30 days.
Want a comprehensive FinOps & Cloud Audit?
RapidStackLab provides 24/7 Managed Cloud SRE, infrastructure audits, and continuous FinOps optimization. We typically uncover 30–50% in immediate cloud waste.
Schedule Infrastructure Audit →Summary & Key Takeaways
Effective cloud cost reduction is not about downgrading server specs or risking production stability. It is an engineering discipline: optimizing network topologies, using intelligent node allocators, migrating to modern processor architectures, and enforcing automated storage lifecycles.
Explore our Cloud Engineering and DevOps & CI/CD Pipelines services to learn how we keep enterprise infrastructure fast, resilient, and lean.