Skip to content
Français
Account: not signed in, sync available My learning
Select theme
Français

Pay-as-you-go: decoding cloud pricing models

By updated

Improve this page

When a beginner launches their first cloud VM, they think they are paying for "the VM". In reality, from the very first second, the meter is running on seven simultaneous axes: compute time, disk space, outbound traffic, I/O requests, licences, snapshots and backups, support. This page breaks down each axis, explains the billing granularities across providers, compares the three commitment models (on-demand, reserved, spot) and lists the five pitfalls that blow up bills without anyone noticing before the following month's bill. It ends with a method to estimate the monthly cost of an architecture before deploying it.

What you will learn

  • The 7 simultaneous billing axes you pay for every second
  • The billing granularity by provider and its real impact
  • The 3 commitment models (on-demand, reserved, spot) and their respective savings
  • The five billing pitfalls to anticipate
  • How to audit a cloud bill in four steps, without panicking
  • How to estimate the monthly cost of an architecture, transfers included, on a dated price grid

Prerequisites: having understood the accounting shift of the cloud. If needed, start with CAPEX vs OPEX (in French).

1. The 7 simultaneous billing axes

A common framing mistake is to reason at the VM level. The displayed price of an instance (for example "c5.xlarge at 0.2020 USD/h" in Paris, AWS price list of 14 November 2026, Linux, on-demand) covers only the compute. Alongside it, six other meters run in parallel, sometimes invisible on the product page, but very much present on the bill.

Compute time, the visible cost

Compute time is the most publicised axis. You pay for the VM as long as it is in the running state. The rate depends on the instance type (CPU, RAM, GPU), the region (for this instance, Paris costs ~19% more than Ohio: 0.2020 against 0.17 USD/h), and the OS (Windows adds the Microsoft licence: for this same instance in Paris, 0.386 USD/h against 0.2020 on Linux, about +91%). A c5.xlarge in eu-west-3 costs 0.2020 USD/h on-demand (AWS price list retrieved on 14 November 2026), or ~147 USD/month (730 h) if it runs 24/7.

Storage, the forgotten cost

Every disk attached to a VM is paid for as long as it exists, even if the VM is stopped. A 100 GB EBS gp3 volume (AWS's block disk) costs you ~9 USD/month in Paris (0.0928 USD/GB-month), whether it is used or not, attached or not. This mechanism explains why a test environment "switched off at night" keeps costing money: the monthly bill only drops to zero if you destroy the resources, not if you stop them.

Network egress, the expensive trap

Data going into the cloud is free, data coming out is billed. Rates vary drastically between providers: 0.09 USD/GB at AWS after 100 GB free per month (first tier of 10 TB, AWS Data Transfer pricing retrieved on 14 November 2026), 0.087 USD/GB at Azure (France Central, public pricing API, 14 November 2026), 0.12 USD/GiB at GCP (the GiB, or gibibyte, is about 1.074 GB; Premium tier up to 1,024 GiB per month, from the europe-west9 region in Paris to Europe, Google Cloud pricing page retrieved on 14 November 2026), but 0 at OUTSCALE, whose grid shows free outbound traffic outside the Asia region. An application that serves 10 TB/month to its users pays nearly 900 USD per month just in bandwidth at AWS. This trap deserves its own page: see Egress and data transfer economics.

I/O operations, the invisible cost

On managed services (S3, DynamoDB, RDS), every request is billed. S3 Standard bills, in Paris, 0.0053 USD per 1,000 PUT requests and 0.00042 USD per 1,000 GET requests (S3 price list of 9 October 2026). At small scale, this is negligible. At the scale of an application that makes 50 million GETs per day (1.5 billion over 30 days), we are talking about ~630 USD per month in requests alone, not counting object storage or egress.

OS and software licences

A Windows VM bills the licence on top of the compute, and SQL Server weighs far more. For the same c5.xlarge in Paris (AWS price list of 14 November 2026): 0.2020 USD/h on Linux, 0.386 USD/h on Windows (×1.9), 0.866 USD/h with Windows and SQL Server Standard (×4.3) and 1.886 USD/h with Windows and SQL Server Enterprise (×9.3). Always check whether the licence is included or billed separately before signing off a design.

Snapshots and backups

A snapshot is not a full copy of the volume, except the first one. This is a point that estimates often miss, and it changes everything: EBS snapshots are incremental, only the blocks modified since the previous snapshot are stored. The rate is 0.053 USD/GB-month in the Paris region.

The practical consequence is that a daily snapshot does not cost the size of the volume each day. On a 200 GB volume where 5 GB change per day, one year of daily snapshots amounts to:

first snapshot 200 GB
364 increments x 5 GB 1,820 GB
---------
total stored 2,020 GB, or about 2 TB

Multiplying 200 GB by 365 would give 73 TB, which is 36 times too much. The real bill therefore depends on your daily change rate, a figure you measure, not a rule of three on the volume size.

Support

Basic support is free at all hyperscalers, but it only gives access to the documentation and the forums: opening a ticket on a production incident requires a paid plan. At AWS, billing is tiered and decreasing with the monthly spend, which rules out using a single percentage.

AWS planMonthly minimumTiers on monthly spend
Business+29 USD9% up to 10,000 USD, 7% from 10,000 to 80,000, 5% from 80,000 to 250,000, 3% above
Enterprise5,000 USD10% up to 150,000 USD, 7% from 150,000 to 500,000, 5% from 500,000 to 1 M, 3% above

The calculation is done tier by tier, not by applying the rate of the last tier to the whole amount. On an account that spends 50,000 USD per month on Business+:

9% x 10,000 USD = 900 USD
7% x 40,000 USD = 2,800 USD
---------
3,700 USD per month

That is more than double what a flat 3% rate would give. It is an item regularly forgotten in initial budgets. Rates retrieved from the AWS support pricing page on 9 October 2026.

2. Billing granularity: second, minute, millisecond

Not all providers have the same billing precision. This difference looks technical, but it shows up on the bill: on a workload that starts and stops often, the granularity gap radically changes the amount.

Provider / serviceGranularityYear introduced
AWS EC2 Linuxper second (60 s minimum)October 2017
AWS EC2 Windows and SQL Serverper second (60 s minimum)June 2021
AWS EC2 Red Hat Enterprise Linuxper second (60 s minimum)April 2024
Azure Compute (VM)per minute, rounded down-
GCP Compute Engineper second (60 s minimum)since 2017
Outscale OAPI VMsper hour, with the started fraction of an hour billed per secondOutscale grid consulted on 9 October 2026
AWS Lambdamillisecondsince December 2020 (100 ms increments from 2014 to 2020)
Azure Functions Consumptionmillisecondsince 2017
GCP Cloud Run100 milliseconds (rounded up)-

Fine granularity (second, ms) favours CI/CD workloads, short batch jobs and event-driven functions. Coarse granularity (full hour) penalises short usage: at a provider that billed the whole hour, an instance launched for 5 minutes would cost as much as an instance launched for 60 minutes. This is no longer the case at OUTSCALE, whose grid specifies that each fraction of an hour consumed is billed per second.

Practical consequence on a CI (illustrative example, at €0.33 per machine hour): a job that starts 50 times per day for 6 minutes costs ~€50 per month with per-second billing and ~€500 per month if the whole hour were billed. The gap is a factor of 10, for exactly the same computation performed.

3. The 3 commitment models

The on-demand price is rarely the one mature companies pay. Three models overlap, each with a flexibility/cost trade-off.

On-demand: total flexibility, maximum price

You launch, you pay for what you use, you stop whenever you want. No commitment, no reservation. It is the default when you launch an instance for the first time, and also the most expensive. A perfect fit for MVPs, R&D, unpredictable spikes, workloads that will die in 2 months.

Reserved Instances and Savings Plans: up to -72%

By committing to 1 or 3 years of usage, you obtain massive discounts. A 3-year All Upfront Reserved Instance at AWS on a c5.xlarge goes from 0.2020 USD/h to the equivalent of 0.076 USD/h (1,996 USD paid upfront for three years, standard offer, Linux, eu-west-3, AWS price list retrieved on 14 November 2026), a -62% reduction. Savings Plans are more flexible (a commitment to an hourly amount, not to an instance type). For this same instance and the same term, a Compute Savings Plan All Upfront comes to 0.102 USD/h (-50%) and an EC2 Instance Savings Plan to 0.076 USD/h (-62%). AWS advertises up to 66% off for the former and 72% for the latter.

The catch: a 3-year commitment is as rigid as buying a server. A standard Reserved Instance and an EC2 Instance Savings Plan stay tied to their region (only a Compute Savings Plan applies in every region), you pay even if you stop the VM, and a Savings Plan cannot be cancelled during its term. Only a standard Reserved Instance can be resold, on the AWS Reserved Instance Marketplace and under conditions. This is the CAPEX disguised as OPEX that I discuss in CAPEX vs OPEX (in French).

Spot and Preemptible: up to -90%

Spot instances (AWS), Spot VMs (Azure, which replaced the former Low-Priority term) and GCP Preemptible VMs (now grouped under the Spot name) are VMs sold on the market for surplus capacity. Providers advertise discounts of up to 90% (AWS, Azure) or 91% (GCP) compared with the on-demand price, as noted in October 2026. These are ceilings: the real discount varies with the machine type, the region and the moment, and the price is variable, not fixed. In return, the provider can cut the machine when it needs the capacity elsewhere, with a notice that depends on the provider: 2 minutes at AWS, 30 seconds at best at Azure (scheduled events), and at GCP a shutdown signal followed by 30 seconds at best to stop, with an optional 120-second notice before that signal.

A perfect fit for interruption-tolerant workloads: ML training, batch processing, video encoding, parallel CI/CD, fuzzing. Unsuitable for stateful workloads (databases, live user sessions), unless you put a heavy fault-tolerance layer in place.

ModelReduction vs on-demandCommitmentRisk
On-demand0%nonenone
Reserved 1 year (standard)-37 to -41%1 yearlong commitment
Reserved 3 years All Upfront (standard)-62% (up to -72% advertised)3 years + upfront paymentvery long commitment
Savings Plan 3 years-45 to -62% depending on type (up to -66% Compute and -72% EC2 Instance advertised)3 years (hourly amount)long commitment, more flexible
Spot / Preemptibleup to -90% (variable)noneinterruption with a notice of 30 s to 2 min depending on provider and configuration (GCP: 30 s at best to stop, 120 s of optional notice before that stop)

The discounts on the Reserved and Savings Plan rows are those of a Linux c5.xlarge in Paris, retrieved on 14 November 2026 from the price list published by AWS. They vary with the instance family, the region and the payment option.

4. Five billing pitfalls to know

Five drifts come up often when a team discovers its cloud bill. Each card describes the mechanism, the figure when a dated price allows it, and the remedy.

Stopped instance, EBS still billed

You stop an instance on Friday evening to save money, but the EBS disk stays attached and billed around the clock. The saving is partial, not total.

Take a c5.xlarge in Paris: 147 USD per month of compute, and 18.56 USD for 200 GB of gp3. Switching the machine off removes the compute but not the disk, so you save about 89% of this machine's bill, not all of it. The gap becomes significant when the volume is large: on 2 TB of disk, storage weighs more than compute and stopping the machine barely lowers the bill anymore.

To really free the costs, you must destroy the VM (after a snapshot), not just stop it. An Auto Scaling Group that scales down to zero instances, or an IaC pipeline that destroys then recreates the environment, automates this gesture.

NAT Gateway billed 24/7

An AWS NAT Gateway is the gateway through which machines without a public address reach the Internet (the Network module presents it). It costs 0.05 USD/h, or ~36.5 USD/month, plus 0.05 USD per GB processed (eu-west-3 rates of 9 October 2026). Many architectures deploy one per AZ and per environment by default.

With 3 environments, 2 AZ and one NAT Gateway per AZ, you reach 6 NAT Gateways, or ~219 USD/month for the NAT alone, traffic excluded.

To lower the bill, keep a single NAT Gateway on the dev/test environment (not one per AZ), remove it for workloads that do not need to reach the Internet, and use a vpc-endpoint to reach AWS services.

Cumulative snapshots without lifecycle

A backup policy without retention makes the bill drift. Since snapshots are incremental, the bill does not follow the volume size but the accumulation of modified blocks, which grows as long as nothing expires.

It is not the first snapshot that drifts, it is the chain. If each machine produces a daily increment kept without limit, the snapshot billing line grows steadily, month after month, without any event flagging it. And deleting the oldest ones does not bring it back down as long as the following ones reference their blocks.

Decide the retention before the first backup, for example 7 daily, 4 weekly, 3 monthly and 1 yearly, and automate it with Data Lifecycle Manager on the AWS side. For long-term archiving, the archive tier drops to 0.0132 USD/GB-month, at the price of a delay and a restore cost.

Invisible cross-AZ traffic

Traffic between two AZs is billed 0.01 USD/GB in each direction at AWS (AWS Data Transfer pricing, Paris, 9 October 2026). A 3-tier application whose components are spread over several AZs pays this rate on every call that crosses a zone, and this line goes unnoticed as long as nobody looks for it.

The typical case is a Kubernetes cluster whose scheduler places pods without regard to zones: pods that talk a lot to each other can end up in different AZs, and every exchange is then billed.

To limit these exchanges, use topology aware routing, zone-aware load balancing, and placement rules that bring communicating pods closer together.

Logs accumulated indefinitely

CloudWatch Logs bills ingestion at ~0.60 USD/GB (0.5985 USD, custom logs in the Standard class, eu-west-3, 9 October 2026), then storage per GB-month (0.0315 USD, Standard class, eu-west-3, AWS price list retrieved on 14 November 2026). The log volume therefore decides the line: a service that writes 100 GB of logs per day pays ~60 USD of ingestion per day, before any storage.

Leaving the DEBUG level on in production multiplies the messages written. Each one is billed at ingestion, then kept as long as no retention expires it: this line can therefore weigh heavily on the bill, for a single service.

Keep the INFO level in production, set an automatic 30-day retention on application logs and archive beyond that, for example to S3 Glacier.

5. Three good practices to keep control of the bill

To avoid being a victim of the cloud bill, three practices have become fundamentals of cloud financial governance (FinOps). They are implemented in the following logical order.

Mandatory tagging from the start. Define a consistent tagging policy (CostCenter, Project, Environment, Owner, ManagedBy) and enforce it with a blocking policy at provisioning time. Without tags, you cannot allocate the bill to teams or projects, and you pay globally without fine-grained visibility. The cost of putting tagging in place late is much higher than doing it from the start, because you then have to retro-tag existing resources.

The monthly human audit of the bill. A self-refreshing dashboard that nobody looks at has no effect: you need a one-hour monthly ritual where a designated person opens the cost console, identifies the top items, explains the monthly change, and lists the month's priority corrective actions. To read the bill, four steps are enough: group by service (which product costs the most), then by usage type (compute hours, GB stored, GB outbound, requests), filter by tag (who consumes), and compare with the previous month to isolate the line that moved. An unknown or sharply rising line is the starting point of the investigation, with the five pitfalls of the previous section as the list of suspects.

Cloud budgets with alarms from the start. The three hyperscalers offer a budget service with alerts (AWS Budgets, Azure Cost Management, Google Cloud Billing); at the sovereign providers, the offering varies and may be missing, in which case you must export the consumption and set your own alert. An alert is not a ceiling: it warns, it stops nothing (unless you configure an automated action). Setting thresholds at 50%, 80% and 100% of the expected monthly budget, with email and team messaging notification, avoids bad surprises caused by a leak or abnormal usage.

6. How do you estimate the monthly cost of an architecture before deploying it?

A reliable estimate adds up the items one by one, with a quantity and a dated unit price for each, and does not forget the transfers. The method comes down to five steps. List the resources of the architecture. Work out for each one the monthly quantity (730 hours for a machine that is always on). Read the unit price off the provider's public price grid, noting the date and the region. Multiply, then add the transfers, which hide in the outbound traffic to the Internet and in the traffic processed by the egress gateway. Add it all up, keeping the grid in front of you so you can redo the calculation the day a price changes.

Here is an example with the prices of this page (AWS, Paris region, price lists of 9 October and 14 November 2026 depending on the item, in dollars, no conversion). The architecture: two c5.xlarge machines that are always on, two 200 GB gp3 disks, a NAT Gateway that processes 500 GB per month, and 1,000 GB of outbound traffic to the Internet.

ItemMonthly quantityUnit priceMonthly cost
Compute2 machines × 730 h0.2020 USD/h294.92 USD
gp3 disks400 GB0.0928 USD/GB-month37.12 USD
NAT Gateway, hours730 h0.05 USD/h36.50 USD
NAT Gateway, processed traffic500 GB0.05 USD/GB25.00 USD
Outbound traffic900 GB billed (100 GB free)0.09 USD/GB81.00 USD
Total474.54 USD

The total exceeds the compute alone (294.92 USD) by 61%: this is the order of magnitude of the 1.5 to 2 factor rule seen above, and it is why an estimate that stops at the machine is wrong by more than a third. The transfers weigh 106 USD out of 474 here, more than a fifth of the bill.

The same calculation puts a figure on a saving. Ten t3.medium test servers at 0.0472 USD/h (Linux, Paris region, price list of 9 October 2026) that are always on cost 10 × 0.0472 × 730 = 345 USD per month. Stopped 16 hours per day, so on for 243 hours per month, they cost 10 × 0.0472 × 243 = 115 USD, which is 67% less on compute. The disks stay billed when the instance is stopped: the saving only applies to the instance hours.

Redo this exercise with your provider's grid. The prices in the example only serve to show the mechanics: the method survives a change of rates, the figures do not.

Put it into practice

The lab provides the file of an architecture, a synthetic and dated price grid, and the usage of one month. You write a cost-estimate.json file with one line per compute class, then one line for block storage, object storage, snapshots, billable outbound traffic and API requests. The file also gives the monthly total, the total over the year, the budget and the remaining margin. The validation recomputes each line from the grid and the usage: it checks the quantity, the unit, the unit price and the cost, the free allowance deducted from outbound traffic, then the total, the currency, the date of the grid and compliance with the budget. The prices are those of the exercise, not those of a provider.

Key points

  • A cloud instance is billed on 7 simultaneous axes: compute, storage, egress, I/O, licence, snapshot and backup, support, not just the displayed VM.
  • Billing granularity varies by provider (second, minute, millisecond) and directly impacts short workloads by a factor of 10.
  • The 3 commitment models offer flexibility/cost trade-offs: on-demand for the variable part, Reserved/Savings Plans for the stable base, Spot for what tolerates interruption. A possible starting mix aims for 60 to 70% commitment, to adjust to your usage.
  • Five recurring pitfalls make bills drift: forgotten EBS volumes, unrationalised NAT Gateways, cumulative snapshots, invisible cross-AZ traffic, logs without retention.
  • Three structural reflexes from day 1: mandatory tagging by policy, monthly human audit, budgets with alerts (which warn without capping).
  • An estimate adds up the items one by one, each with a quantity and a dated unit price, transfers included: in the example of section 6, transfers weigh more than a fifth of the bill.

Knowledge check

Check that the essentials of this guide are in place. The questions only cover what has just been explained here.

Knowledge check

Check what you have learned with this interactive quiz

6 questions
6 min
70% to pass

How it works

  • The timer starts when you click Start
  • Multiple choice, true/false and short answer questions
  • You can move between questions
  • Detailed results are shown at the end

Starts the quiz and the timer