How One AWS Bill Went From Almost $8,000 a Month Toward $6,000

Two real scans, eight weeks apart, on the same account

Cloud bills rarely jump overnight. They drift.

A launch adds an oversized Fargate service. A prototype gets a managed database. An experiment ends, but the resources it created stay running. Buckets accumulate without a lifecycle policy because no one owns the cleanup decision. By the time the monthly total becomes uncomfortable, the waste is already spread across a dozen services.

This post is not a hypothetical. It's two greenops-scan reports run against the same AWS account, eight weeks apart, plus that account's own AWS Cost Explorer graph. Here is exactly what the scans found, what changed between them, and what AWS's own forecast says next.

The account

The account runs a mix of ECS/Fargate services, a DocumentDB (MongoDB-compatible) cluster, an Application Load Balancer, ElastiCache, a smaller RDS instance, AWS Transfer Family, and the usual security/compliance tooling (GuardDuty, CloudTrail, Config, KMS, Secrets Manager). By the most recent Cost Explorer data pulled through the scanner, the account was billing $7,726.20 for the month, with the two largest line items being ECS ($3,235.93) and DocumentDB ($1,506.78).

The account's real Cost Explorer history: a near-continuous climb from about $2,000/month in late 2023 to nearly $7,800/month by mid-2026. The last bar is AWS's own forecast (marked with `` and an 80% prediction interval) for the following month — not a confirmed result, but a trend AWS's own model expects given recent changes to the account.

What the first scan found

Running npx greenops-scan against a read-only profile for this account returned 53 findings worth an estimated $1,176.83 a month ($14,121.96/year) in identifiable waste. Nothing in the list was exotic — it was ordinary drift, spread across three modules.

ECS: 15 over-provisioned Fargate tasks — $756.83/month

The scanner flagged 15 Fargate task definitions running well above any resource profile the workload needed, ranging from 2 vCPU/4 GB up to a single task provisioned at 16 vCPU / 32 GB. It also flagged one ECS cluster with no running tasks or services at all.

Task size	Findings	Est. savings/mo
16 vCPU / 32 GB	1	$172.99
8 vCPU / 16 GB	4	$86.50 each
4 vCPU / 8 GB	2	$43.25 each
2 vCPU / 4 GB	7	$21.62 each
Empty cluster	1	$0 (still worth cleaning up)

S3: 37 buckets with no lifecycle policy — $415/month

Thirty-seven S3 buckets — mostly CodeBuild/CodePipeline artifact buckets — had no lifecycle rules, meaning every object stayed in the Standard storage tier indefinitely instead of transitioning to Glacier or being expired. Individually cheap ($10–15/month per bucket), collectively the third-largest line item in the report.

EC2: 1 stopped instance still billing storage — $5/month

A stopped t2.small instance was still incurring EBS storage costs. Small on its own, but the kind of thing that's invisible in Cost Explorer's per-service view and only shows up when something checks per-resource state.

What changed by the second scan

Eight weeks later, the same account, scanned again: 1 finding, worth $105.12 a month.

The one remaining item was a DocumentDB cluster (kaeinstancedb) still running on an Intel-based db.r5.xlarge instance class, where switching to a Graviton-based class would cut roughly 30% off that instance's cost.

That means between the two scans, 91% of the originally flagged waste was addressed — the 15 over-provisioned Fargate tasks, the 37 S3 buckets missing lifecycle rules, and the idle EC2 instance were no longer showing up as findings, while a new opportunity (the DocumentDB instance family) surfaced on the second pass.

Update: that last finding has since been resolved too. The kaeinstancedb cluster was migrated to db.r6g.large — Graviton, as the scan recommended, but one size class smaller than the db.r7g.xlarge the report suggested. That's a bigger cut than the flagged $105.12/month, since it downsizes the instance itself in addition to switching CPU architecture. With that change, every finding from the first scan has now been addressed.

Reading the Cost Explorer graph honestly

The graph above is the account's real multi-year cost history, not an illustration. Two things are true about it, and it's worth being precise about which is which:

Confirmed: the bill climbed steadily from roughly $2,000/month in late 2023 to nearly $7,800/month by mid-2026, and the scan-over-scan comparison above confirms that a meaningful chunk of flagged waste was cleaned up in that window.
Not yet confirmed: the final bar on the graph is AWS Cost Explorer's own forecast for the following month, explicitly marked with a double asterisk and an 80% prediction interval, projecting the bill down toward roughly $5,000–$6,000. That's AWS's own model reacting to the account's recent trend — a forecast, not a closed month.

We're reporting both numbers because a scanner report is only useful if it's honest about what it can and can't prove. The $1,176.83 → $105.12 change in flagged waste is a fact from two independent scans. The next month's total landing near $6,000 is AWS's own projection, and worth watching, not asserting as done.

What made the difference

Three things, consistently:

A clear list. The scanner turned an abstract "the bill keeps climbing" concern into 53 specific resources with issue descriptions and dollar estimates.
A safe first step. Because the scan is read-only and runs locally, nothing needed elevated write access to produce the list — the actual changes (right-sizing Fargate tasks, adding lifecycle rules, terminating an idle instance) were made deliberately, by hand.
A second (and third) scan to check the work. Re-running the same read-only scan two months later is what actually confirmed 91% of the flagged items were resolved, instead of assuming the fixes stuck — and the follow-up on the last item confirmed the rest.

A realistic takeaway

Not every account will show a full cleanup between two scans, and not every Cost Explorer forecast will play out exactly as projected. But the pattern here — over-provisioned compute, storage without lifecycle rules, and a stale instance family — is common enough that it's worth checking for on any account whose bill has been drifting upward for a while.