
The Hidden Traps in AWS Bills: 5 Architectural Mistakes That Quietly Drain Your Budget
Why does your AWS bill spike even when traffic stays flat? A breakdown of the subtle networking, storage, and retention defaults that quietly inflate cloud costs and how to fix them.
The End-of-the-Month Mystery
Almost every developer who builds on AWS has a moment like this:
You launch a project or prototype. You check the pricing for your EC2 instances, calculate database hours on RDS, make sure your Lambda functions fall well within the free tier, and expect a monthly bill of around $20 to $40.
Then the monthly invoice arrives, and it’s $140.
Your traffic didn’t spike. Your database didn’t melt down. Your instances didn’t randomly scale up. So where did the extra $100 go?
When we design architectures in the cloud, we spend 90% of our energy thinking about compute and databases. But in practice, the quietest budget killers on AWS aren't servers—they are idle gateways, default retention policies, unattached disks, and cross-zone networking.
Let’s look at five common architectural traps that quietly inflate AWS bills, why they happen, and how to design around them.
1. The $32/Month NAT Gateway Just Sitting There
Following security best practices means keeping your backend services (like Lambda functions or ECS containers) in private subnets without public IP addresses.
The catch? The moment your private service needs to reach out to the internet—whether to call a third-party API or download a package—it needs an AWS NAT Gateway placed in a public subnet.
The Catch
A single NAT Gateway costs $0.045 per hour just to exist.
That translates to roughly $32.40 per month, before you’ve even processed a single byte of traffic. If you set up a standard multi-AZ architecture with one NAT Gateway in each of three Availability Zones, you're paying nearly $100 every month just for idle routing.
What to Do Instead
- Use VPC Gateway Endpoints for S3 and DynamoDB: If your private backend is mostly talking to S3 or DynamoDB, you don’t need a NAT Gateway for that traffic at all. AWS provides VPC Gateway Endpoints for S3 and DynamoDB completely free of charge.
- Use a Lightweight NAT for Dev/Staging: For pet projects or staging environments where high throughput isn't required, you don't always need a managed NAT Gateway. A tiny
t4g.nanoNAT instance or open-source solutions likefck-natcan cut that cost from $32 down to under $3/month.
2. Inter-AZ Data Transfer: The Invisible Network Tax
Spreading services across multiple Availability Zones (AZs) within an AWS region is the standard way to build resilient, fault-tolerant applications.
The Catch
While data transfer within the same Availability Zone is free, moving data across Availability Zones costs $0.01 per GB in each direction ($0.02 per GB round-trip).
This doesn't seem like much, but it compounds fast in chatty microservice architectures:
- Your web service in
AZ-aqueries a database replica living inAZ-b.
- Your caching layer in
AZ-cconstantly streams data back to workers inAZ-a.
If your internal microservices exchange several terabytes of data across AZ boundaries each month, you can easily rack up hundreds of dollars in cross-zone networking fees for traffic that never even left your AWS account.
What to Do Instead
- AZ-Affinity Routing: Whenever possible, route internal traffic to services and replicas located in the same Availability Zone.
- Keep Chatty Workloads Co-located: Group tightly coupled services that exchange high data volumes in the same zone, while maintaining independent standby replicas across other zones for disaster recovery.
3. Zombie EBS Volumes and Forgotten Snapshots
When you terminate an EC2 instance, it's easy to assume everything attached to it gets cleaned up automatically.
The Catch
Unless you explicitly enabled the "Delete on Termination" setting for secondary Elastic Block Store (EBS) volumes, those disks will simply detach and sit in your account as "unattached" volumes.
A standard 100 GB
gp3 volume costs around $8.00 per month. If you frequently spin up and tear down instances for experiments or testing, orphaned volumes can quietly linger for months.The same goes for automated snapshot scripts: taking daily snapshots without an automatic deletion rule means you end up paying for hundreds of gigabytes of cold backup data you will never need.
What to Do Instead
- Audit Unattached Disks: Periodically search the EC2 console for volumes in the
available(unattached) status and clean up what is no longer in use.
- Automate Retention with AWS Backup: Instead of writing custom cron scripts to take snapshots, use AWS Backup with clear lifecycle rules that automatically delete snapshots after 7, 14, or 30 days.
4. CloudWatch Logs Set to "Never Expire"
Logging is essential for troubleshooting and monitoring system health. When you deploy a Lambda function, ECS task, or API Gateway, AWS automatically creates an Amazon CloudWatch Log Group.
The Catch
By default, every newly created CloudWatch Log Group is set to "Never Expire".
CloudWatch charges $0.50 per GB for log ingestion and $0.03 per GB per month for storage. In active systems with verbose logging, gigabytes of debug logs from months or years ago stay stored forever, steadily increasing your monthly bill.
What to Do Instead
- Set Explicit Expiration Periods: Change the default retention on log groups from "Never Expire" to 7, 14, or 30 days for standard application logs.
- Archive to S3 Glacier for Compliance: If you have compliance or auditing requirements that require keeping logs for years, export them to Amazon S3 and transition them to S3 Glacier, where storage costs are a fraction of CloudWatch storage (~$0.0036/GB vs $0.03/GB).
5. Over-Provisioned Capacity vs. Serverless Mismatches
Choosing how you pay for databases and compute usually comes down to two models: Provisioned Capacity (paying for a dedicated baseline) and On-Demand / Serverless (paying per request).
The Catch
Picking the wrong model for your traffic pattern will always waste money:
- Provisioned Capacity on Spiky/Low Traffic: If you configure an Amazon DynamoDB table or Aurora database with high provisioned capacity to handle occasional spikes, but the service is idle 80% of the day, you're paying for unused headroom 24/7.
- On-Demand on High, Steady Traffic: Conversely, running high-throughput, predictable workloads entirely on on-demand pricing can end up costing significantly more than reserving baseline capacity.
What to Do Instead
- Match the Pricing Model to Traffic: Use On-Demand / Serverless (like DynamoDB On-Demand or Aurora Serverless v2) for unpredictable or intermittent traffic so you pay $0 when traffic drops.
- Use Savings Plans for Stable Workloads: Once your workload becomes predictable and runs 24/7, adopting a 1-year Savings Plan can cut your compute costs by 30% to 50%.
Final Thoughts
Good cloud engineering isn't just about making things run—it's about understanding how architectural choices affect cost.
Before deploying any new workload to AWS, it pays to ask three simple questions:
- Is there an idle gateway or provisioned service running 24/7 that isn't strictly necessary?
- Where is data crossing network boundaries (AZs, VPCs, Internet)?
- What is the lifecycle and expiration plan for our storage and logs?
Catching these subtle details early keeps your cloud footprint clean, resilient, and cost-effective.
What’s the most surprising hidden charge you’ve ever encountered on an AWS bill? Share your experience in the comments below!
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article