Cloud costs: the five items that inflate the bill

The cloud promises that you only pay for what you use. That is true, and that is exactly the problem: everything that runs gets paid for, including what nobody uses any more. A cloud bill almost never goes off the rails at once. It grows month after month, through items nobody looks at because none of them is dramatic on its own.

This article from our Infrastructure and cloud section goes through the five items that weigh the most, what makes them invisible, and the routine that keeps them in check. The price examples come from AWS because its rates are public and detailed; the other big providers bill equivalent items, under their own rules.

Why does a cloud bill drift?

Because creating a resource takes a minute and deleting it is nobody’s job. Flexera’s annual State of the Cloud report, published in March 2026, gives the measure of it:

The waste figure is a self-reported estimate, not a measurement of invoices, and Flexera attributes its rise to AI workloads. It still says the essential thing: even companies with a team for it let a large share of their budget slip. For a team of 20 to 500 people, with no dedicated team, the question is not whether there is waste, but where it hides.

Item 1: what stays switched on

The first item is the most mundane: machines, databases or test environments still running after the project ended, or sized for a peak that never came. A virtual machine costs the same whether it is working or waiting.

Three cases keep coming back:

• the forgotten test environment, created for a demo or an acceptance test and never deleted;
• the oversized machine, chosen large “to be safe” and never scaled back down;
• the service running at night and at weekends when nobody uses it outside office hours.

The fix is an organisational rule more than a tool: every resource carries the name of the project and of the person who answers for it, and a resource with no owner gets stopped. That is the job of the infrastructure foundation shared by all your projects: one account or project per use, and one budget per project.

Item 2: outbound traffic

Bringing data into a cloud is generally included. Getting it out, to your users, to another provider or to your offices, is paid by volume. Amazon EC2’s pricing page states 100 GB of outbound transfer to the internet included each month, across all services; beyond that, each gigabyte is billed on a sliding scale.

This item grows quietly when a website serves videos or heavy files, when a backup goes every night to another provider, or when an application exchanges a lot with a service hosted elsewhere. It also weighs on the day you want to leave. The European Data Act regulates these switching charges: according to Arcep, they are capped at the costs directly linked to the switch until 12 January 2027, then prohibited. AWS, for its part, announces outbound transfer at no charge for eligible customers who move all their data off its platform. Everyday traffic, on the other hand, is still billed.

Item 3: storage that piles up

Storage costs little per gigabyte, and that is why nobody watches it. It accumulates in layers: disk snapshots kept forever, backups with no retention period, application logs kept “just in case”, database copies made before a migration and never deleted.

The remedy is a written retention period for each type of data: so many days for daily backups, so many months for monthly ones, so many weeks for logs. It also serves the GDPR, which requires not keeping personal data longer than necessary. Our guide to backing up your own server details a reasonable retention rhythm.

Item 4: small services billed by the hour

These are the lines nobody chose consciously: they come with an architecture. One by one they look negligible; added up and multiplied by environments, they no longer are.

The last example is telling: a Kubernetes version you do not upgrade costs six times more per hour, without a single line of code changing. Overdue maintenance becomes a line on the bill. We cover what Kubernetes demands, and when a mid-sized company can do without it, in Kubernetes: when a mid-sized company does not need it.

Item 5: usage meters

The last item is the newest and the most variable: everything billed by the minute, by the call or by the token. Continuous integration minutes, function executions, calls to an AI model. Here the bill no longer depends on what is switched on, but on what your programs do, and a program that loops or starts from scratch after each failure can multiply consumption overnight.

AI calls follow the same logic, with one more lever: choosing the model for each task. Our article on the AI bill, tokens and model routing covers it in detail.

How do you keep control without slowing projects down?

The five items have one thing in common: none of them shows on the bill total. They show per project, with someone looking. The routine that is enough for a mid-sized company:

A budget alert stops nothing. At Microsoft Azure, the documentation says so explicitly, as we detailed in the foundation guide: consumption continues after the threshold is crossed. At other providers, read what an alert actually does before relying on it. The monthly review remains the only control that catches all five items at once.

This work does not replace that of the people who administer the infrastructure: it gives them the figures to make trade-offs, and gives management the means to understand their choices.

When a fixed price becomes simpler

For a stable workload, running day and night without big peaks, a server rented at a fixed price makes items 1 and 4 predictable: you pay for the machine, whatever its use, and item 2 depends on the traffic included in the offer. It adds others, mostly in operating hours. The choice is made use by use, and we detail it in dedicated server, VPS or cloud: choosing by use. For your software subscriptions, the same three-year comparison is in the real cost of leaving SaaS.

Where to start

Take the latest bill and sort each line into one of the five items. For each, find the project and the person who answer for it: the lines with no answer are your first source of savings. Then set one budget per project, followed by the monthly review.

The rest of the section covers the other operating decisions: what monitoring should watch, the disaster recovery plan to write, which tools to move out of SaaS first, self-hosted n8n in production, cloud migration without breaking things, Europe’s dependence on the American cloud and what NIS2 will require of French companies.

If your business tools can run on a fixed-cost server, our Leaving SaaS offer installs them there, with tested backups and documentation included. If your bill comes from a big provider, tell us about your projects: we will look with you at where the money goes.

Frequently asked questions

Sources

1. AWS, Amazon EC2 on-demand pricing (new tab)aws.amazon.com
2. Arcep, press release of 16 February 2026 on cloud switching charges (new tab)en.arcep.fr
3. AWS News Blog, announcement of the public IPv4 address charge (new tab)aws.amazon.com
4. AWS, Amazon VPC pricing (new tab)aws.amazon.com
5. AWS, Amazon EKS pricing (new tab)aws.amazon.com