esköre™

InsightsCloud Setup & Migration16 min readNovember 15, 2026

Why Is My AWS Bill So High? A Triage Guide to Finding the Cause Without Breaking Production

Your AWS bill has jumped and nobody can explain it. This triage guide shows how to find the line item, check the usual suspects and clean up without taking production down.

Cloud Setup & Migration at Eskore
Photo by Unsplash on Unsplash

Why is my AWS bill so high? It is the right first question, and the answer can usually be traced on the invoice. Every charge has a service, a usage type, a region and an account attached, and the cause sits somewhere in that list. The difficulty is that the AWS console is built for engineers. Most owners who open it are looking at a page of service names with nobody to ask.

This guide works through the bill in a fixed order. First, decide whether the bill jumped or has always been high, because those two situations have different causes. Then find the exact line item, check the usual suspects one at a time, and clean up in an order that does not take production down. The method applies to Azure and Google Cloud accounts too. If you have traced the cause and need someone to carry out the fix, we cover that near the end.

Why is my AWS bill so high? Check whether it jumped or stayed high

Start with the invoice currency. AWS prices in US dollars. If your bill is in USD and your bank or accountant converts it to Canadian dollars, the CAD figure moves with the exchange rate even when your usage is flat. Check the currency on the invoice and in the Billing console before you assume anything has changed in your architecture.

Next, pull daily cost for the last three months in Cost Explorer and look at the shape of the line:

  • A step. Cost sits at one level, then rises and stays there. Something changed on that day. Ask what was deployed, launched or switched on.
  • A spike. A few days well above normal, then back to baseline. This is usually a one-off: a load test, a bulk data job, a script in a loop, or a backup that copied far more than expected.
  • A slope. Cost climbs a little every week. This usually means logs, snapshots or stored data are growing and nobody prunes them.
  • A flat high line. Cost has been high since the start and never moves. This points to architecture and sizing, not a recent event.

Steps usually have a trigger you can find in a change log: a deployment, a traffic peak, a new region, an auto-scaling event, a forgotten test environment, or credits running out. Flat-high bills point to design: instances sized for a peak that never comes, a multi-AZ database nobody needed, or logs kept forever. Both are worth finding, but they need different questions.

Find the line item in Cost Explorer

Cost Explorer sits in the Billing and Cost Management console. You need an account with billing permissions. Work through these steps in order:

  1. Set the date range to the last three months and the granularity to daily.
  2. Group by Service. Note the top five. In most small accounts, two or three services account for nearly all of the bill.
  3. Change the grouping to Usage type and filter to the top service. This is usually where the real cause appears.
  4. Group by Region. A region you did not choose, or have never used, is a warning sign.
  5. Group by Linked account if you use AWS Organizations. This shows which account or team is running the spend.
  6. Group by a tag such as owner or product, if tags exist. If they do not, that is the first thing to fix. See the tagging section below.

Two things catch people out. First, Cost Explorer shows unblended cost by default. Reserved Instances and Savings Plans can make a single month look lumpy. An amortized view spreads those commitment charges across the period, which makes trend comparisons fairer. Second, history is limited. By default you get up to 12 months of daily data, while hourly and resource-level data have shorter windows, and some settings carry their own charges. Check AWS's Cost Explorer documentation before you rely on a comparison from several months back.

Drill into usage type and region first

Usage type names look cryptic, but the pattern is consistent. Names that include an instance type such as t3.large, or end in Hours, are compute time. Names containing DataTransfer describe data moving between places. Names containing NatGateway describe the hourly charge for a NAT Gateway and the data it processes. Names containing TimedStorage or Snapshot describe stored data. The exact names vary by service and region, so read them rather than guess.

A region or linked account that nobody on your team recognizes is the most useful signal you will get. It often means a forgotten test environment, a contractor's project, or something left over from an earlier product. Write down the resource IDs and, if you can find one, the owner. Do not touch anything yet. The sections below show what to check, and the cleanup section shows how to act.

The usual suspects and how to check them

Six causes show up again and again in AWS accounts of this size. Work through them one at a time. Each check is quick on its own, but changing several things at once makes it impossible to tell which change did what.

Suspect What to check Typical fix
Idle or oversized compute CPU and memory over several weeks; instances with little recent activity Right-size; schedule stops for non-production
Unattached EBS volumes and old snapshots Volumes in the "available" state; snapshots with no matching AMI or backup policy Confirm, back up, then delete
NAT Gateway and data transfer NatGateway usage types; S3 or DynamoDB traffic going through NAT Add gateway endpoints where they fit
Cross-AZ traffic Regional data transfer lines; chatty services split across zones Review the design; keep resilience where it matters
CloudWatch Logs without retention Log groups sorted by stored size; retention set to never expire Set retention to match need; reduce verbose logging
Auto Scaling and Lambda running away Group maximums and scaling history; invocation counts and retry settings Set limits; fix loops and retries

Idle or oversized compute

Over-provisioned instances are among the most common causes, and they are easy to get wrong in the other direction. A server that looks idle on a Tuesday may be the home of the nightly batch job. Look at CPU and memory utilization over at least a few weeks, including the busiest day of the month. EC2 does not report memory by default. It needs the CloudWatch agent, so check whether that is installed before you decide a server has spare capacity.

Right-sizing means moving to a smaller instance type, which usually needs a restart and a test. Schedule non-production environments to stop outside working hours. Get the owner to approve both changes. Stopping an EC2 instance halts its compute charges, but attached storage still bills. Check databases too. A database sized for launch week often runs at a fraction of its capacity for years.

Unattached volumes and old snapshots

When an instance is terminated, its root volume is usually deleted, but separately created data volumes often are not. Those volumes sit in the available state and keep billing. Snapshots pile up from backup jobs and from people who took one "just in case" and never came back.

Before you delete a snapshot, confirm it is not needed for recovery. Check whether it backs an AMI you still use, whether a backup policy created it and what retention that policy expects, and whether the volume or database it came from still exists. A snapshot that is the only copy of last quarter's data is not a cost problem. Snapshot storage is billed per gigabyte-month, and the rate varies by region, so check the current figure on AWS's pricing page before you estimate anything.

NAT Gateway and data transfer costs

NAT Gateways are the surprise line item for many small teams. They bill for every hour they run and for every gigabyte they process, on top of normal data transfer. A private subnet that pulls from S3 or DynamoDB through a NAT Gateway pays the processing charge for traffic that never needed to leave the AWS network.

Consider an illustrative case. A six-person software company in Austin moves its application into private subnets. Nothing else changes, but its NAT Gateway processing line climbs week after week, because the application pulls large objects from S3 through the gateway. The fix is a gateway endpoint for S3, which routes that traffic privately. The same logic works in Raleigh, Denver or New York, and in Toronto.

Gateway endpoints for S3 and DynamoDB are built for this case, and AWS does not charge an hourly fee for them, though you should confirm that on the VPC pricing page. They do not help with traffic to other internet services, such as a third-party payment API. Interface endpoints for other services do carry hourly and per-gigabyte charges, so compare before you add endpoints everywhere. Also check your public IPv4 addresses. AWS charges for them, so unused ones are worth releasing.

Rates vary by region, and third-party guides often disagree with each other. Before you put a figure in front of an owner, use AWS's pricing page for your region.

Cross-AZ traffic charges

Traffic between availability zones in the same region is billed per gigabyte in each direction. It is easy to miss because it appears under regional data transfer rather than under the service that caused it. Multi-AZ designs, load balancers that spread requests across zones, and chatty services split across zones all generate it.

Treat this as a trade-off, not as waste. Spreading across zones is what keeps an application running when one zone has a problem, and removing it to save money can turn a short outage into a long one. The question is whether each cross-zone path needs to exist. A service that calls a database in another zone on every request is a candidate for a design review. A database that must fail over to another zone is not.

CloudWatch Logs without retention

Log groups keep logs indefinitely unless you set a retention period. Ingestion and storage both bill, and the storage charge grows every month while nobody looks. In the CloudWatch console, sort log groups by stored bytes to find the largest ones. Then ask two questions: what does the business need to keep, and for how long?

A support team may need 30 days of application logs. A compliance requirement may call for longer, and for that, archiving to S3 is often more sensible than keeping everything in CloudWatch. Set retention on each log group to match the answer. Then check the source. Debug logging left on in production is a common reason ingestion is high, and turning it down cuts the cost where it is created. Verify current ingestion and storage rates on the CloudWatch pricing page before you quote a figure.

Runaway Auto Scaling and Lambda

An Auto Scaling group with no maximum size, or with scale-in policies that never trigger, can grow to whatever a traffic spike or a fault asks for and then stay there. Check the group's maximum, its scaling activity history, and whether instances are being replaced repeatedly. Repeated replacement often means a failing health check, which is a bug rather than growth. The same logic applies to Kubernetes node groups: a cluster autoscaler with a high ceiling and workloads that never shrink produces the same bill.

Lambda storms are usually loops. A function triggered by an S3 upload that writes back to the same bucket will call itself until someone stops it. A queue whose consumer fails and retries indefinitely does the same. Check invocation counts over time, the error rate, and the retry configuration on each trigger. If invocations rise in step with a source you can name, you have found the loop. Fix the loop first, then set concurrency limits so a fault cannot run without bound.

When credits, free tier or commitments change the bill

Sometimes nothing has changed in the architecture and the bill still rises. That is usually a billing change, not a usage change.

New AWS accounts can receive free-tier usage and promotional credits. Both have limits and end dates, and AWS has changed its free plan structure in recent years. When credits run out or the plan period ends, the same usage bills at standard rates. If your account is new, or the bill jumped after a first year of low costs, check this first. Look at credits and plan details in the Billing console, and read the terms on AWS's Free Tier page, because the rules depend on when your account was opened.

Commitments work the other way. Reserved Instances and Savings Plans give a discount in return for a term. When one expires, or the usage it covered changes, the discount falls away and net cost rises even though the hours have not changed. Put commitment end dates on a calendar. Renewal is a decision for the owner, not something to leave to an engineer who may have moved on.

Two more items to check. Stopped database instances still bill for storage and backups, and AWS restarts stopped RDS instances automatically after a set period, so stopping a database is not a long-term fix. Support plan charges also appear as their own line, which surprises people who have never looked at that part of the invoice.

Clean up safely: owner sign-off and a rollback plan

The usual way cost cleanup causes an outage is that someone deletes something they did not understand. Use this order for every change, however small:

  1. Confirm ownership. Find the owner from tags, from the account, or from CloudTrail event history, which shows who created a resource. If no owner can be found, the resource goes on a list for the business to decide on. It does not go on a delete list.
  2. Check the data path and dependencies. Which load balancers, DNS records, cron jobs, backups and services refer to this resource? Ask the people who run those systems, not only the person who read the bill.
  3. Stop before you delete. Stop the instance, detach the volume or disable the function. Give problems time to surface. A full business cycle, such as a month-end run, is a sensible minimum for anything that might be needed occasionally.
  4. Take a snapshot or backup. Make restoration possible before removal, not after.
  5. Wait for the agreed period. Agree the length of the wait with the owner in advance. Make it longer for anything that handles customer or financial data.
  6. Remove it. Only now.
  7. Record the change. Write down the resource ID, the owner, who approved it, the date and the exact restore steps. Store the record where the next person will look.

Get the owner's sign-off in writing for each change, even if it is a short email. A decision buried in a chat thread is hard to find six months later. Do not give an owner a saving figure until this process is complete. A cleanup that trims the bill and takes a customer-facing service down costs more than it saves.

Stop the next surprise with budgets and alerts

Budgets and alerts will not stop a bill from rising, but they will tell you within days rather than at month end. Set a budget for each account or product, with alerts at several thresholds, and send them to a named person. A budget alert is not a hard cap. It tells someone that spend has crossed a line.

Budget Actions can apply a policy or stop instances when a threshold is crossed. That is closer to a control, but it still reacts after spend has started, and it needs careful setup for production systems. Budget Actions carry charges beyond an included allowance, so check the current figures on AWS's Budgets pricing page before you rely on them.

Cost Anomaly Detection watches for unusual spending patterns and alerts you when something deviates from the norm. AWS announced it as generally available in 2020 and has documented it at no additional charge. Terms change, so confirm the current position on AWS's announcement and the pricing page. Create monitors for your major services and send alerts to a shared inbox that someone reads every day.

Build the rhythm around two checks:

  • Weekly, about five minutes. Review anomaly and budget alerts and assign each one an owner.
  • Monthly, about an hour. Reconcile the full bill against the review template below.

Tagging and ownership

Every billable resource should carry three tags: owner (a person or team), product (what it serves) and environment (production, staging or test). Those three are enough to trace most of a bill to someone who can answer for it. Require the tags on new resources through your deployment process, and tag existing resources in one pass from a short list. Then activate them as cost allocation tags in the Billing console. Tags only appear in cost reports after activation, so do this early.

Monthly review template

Use this checklist at each monthly reconciliation:

  • Top five cost drivers this month, by service and usage type, with the change from last month
  • Changes since the last review: deployments, new regions, new accounts, new environments
  • Anomalies and budget alerts, each with an owner and a status
  • Commitments and credits expiring in the next 90 days
  • Resources flagged for review, with the owner and a decision date
  • Actions agreed, with the approver's name and the date

Containers and multi-cloud accounts

Kubernetes moves the cost to different places, but the method stays the same. On EKS, the bill shows up as EC2 instances in node groups, so start with the size and count of each node group, then check how much of each node is actually requested. Pods that request far more CPU or memory than they use force extra nodes, and the cluster cannot shrink while those requests remain. Forgotten namespaces, idle workloads and load balancers left behind by retired services add cost that a service-level view will not label clearly. Tag by namespace or team where your tooling allows.

Eskore works across AWS, Azure and Google Cloud, and the triage method applies to all three: decide whether the bill jumped or is flat-high, drill down to the resource, check the usual suspects, then clean up with sign-off. The service names differ, but the questions do not. Where infrastructure is defined in Terraform, changes go through code review and can be reverted by applying the previous version, which makes both cleanup and rollback safer. If the environment itself needs rebuilding, Cloud Setup & Migration covers AWS, Azure and Google Cloud architecture designed for scale and cost.

When to bring in engineers

Many bills can be triaged in-house. A stopped instance or a forgotten snapshot is a fair job for an operations manager with the right access. Bring in engineers when:

  • No one owns the account. If the person who built it has left and nobody knows who can approve a change, the first job is ownership, not cost.
  • Spend is growing faster than traffic or revenue. That usually points to architecture rather than a single resource.
  • Resources are unsafe to delete. Shared databases, legacy systems and anything touching customer or financial data need engineers who can test a removal properly. If an old application is the root of the problem, Custom Software Development covers rebuilding it around how your business works now. For a broader view of replacing old systems, see How to Modernize Legacy Software for a Small Business Without a Risky Rewrite.
  • The architecture is the cost. Multi-AZ designs, heavy data movement between services and always-on capacity for peaks that rarely arrive are design decisions. The lowest-cost option on paper is not always the one that keeps the business running.
  • The bill keeps returning after cleanup. That signals a process gap, such as a deployment that recreates resources or an autoscaler with the wrong limits.

A scoping call is the right first step. We look at your bill, identify the likely causes and agree what we would check and change. Before any work starts, we agree a written scope, a fixed price and a launch date. Our case studies show the kind of work we take on. We have offices in Toronto, Thornhill, London and New York, and we work with clients across Canada and the United States. You own the code and the accounts throughout.

Why is my AWS bill so high? Talk to our cloud team

If your AWS bill has climbed and nobody can explain why it is so high, the fastest route is a conversation with people who have seen this pattern before. Bring your last three invoices and, if you can, read-only access to Cost Explorer. We will map the likely causes and agree a written scope before any work starts.

If the cost sits in the infrastructure itself, Cloud Setup & Migration covers AWS, Azure and Google Cloud environments designed for scale and cost. If a self-managed server with no monitoring or backups is driving the spend, VPS Setup & Hardening covers that. If a manual reporting process is pulling data out of your cloud every morning and adding cost, AI Automation can replace it.

Book a free 30-minute scoping call and we will start with your bill.

Questions

Need help with this?

Talk to our cloud setup & migration team.