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

Paying for compute: commitments, autoscaling and interruptible instances

By updated

Improve this page

The displayed price of an instance is the highest of the three you can pay. This lesson presents the three ways to buy compute and their trade-off between flexibility and price, the autoscaling that adjusts capacity to the load, the instances sold on spare capacity, and the rare situations that justify leaving the shared model.

What you will learn

  • Choose between pay-as-you-go, commitment and spare capacity.
  • Tune the metric and the parameters of an autoscaling setup.
  • Identify what can run on interruptible capacity, and what cannot.
  • Recognise the three cases that justify a dedicated host or bare metal (a physical server rented whole, with no virtualisation).

What you pay for, and the three ways to buy

The same compute is sold at three prices, depending on the commitment you accept. The gap between the most expensive and the cheapest is considerable.

ModePrincipleFor what
On-demandfull price, no commitmentunpredictable workloads, tests, peaks
Reserved or commitmentdiscount in exchange for a one-to-three-year commitmentthe baseline that runs permanently
Interruptible (spot)large discount, the machine can be reclaimed at short noticebatch compute, tolerant of interruption

The classic mistake is to buy reserved capacity too early, before knowing your real baseline. A rigid reservation, tied to a precise machine type, that you resize six months later, makes you pay for capacity you no longer use, and can cost more than the full price. Some commitment offers are more flexible: they cover an hourly spend amount rather than a given machine. They are called Savings Plans at AWS, savings plan for compute at Azure and flexible CUDs at Google Cloud, for one or three years. They reduce the risk without removing it. Measure first, commit afterwards, and only on what no longer moves.

Switching a machine off does not stop the whole bill. At AWS, a stopped instance no longer bills the compute, but its EBS volumes (the disks attached to the machine) stay billed as long as they exist, as does a public IPv4 address: 0.005 USD per hour, whether it is used or idle, according to the VPC pricing page retrieved on 9 October 2026.

At Scaleway, a stopped instance likewise keeps its storage and its IP: only the compute is no longer billed, and the volumes and flexible IPs keep being billed until they are deleted. Standby mode is not a shutdown: it is billed like a running instance (Scaleway documentation).

A commitment, for its part, is paid whether the machine runs or not: it is the price of the discount. To really stop paying, you have to delete the resources (after a snapshot if needed), not just switch them off.

The details of the pricing models are covered in Pay-as-you-go and pricing models.

Varying capacity automatically

Autoscaling adds and removes machines according to a metric, and its trap is not the threshold setting: it is the delay. An instance that takes four minutes to start and become useful arrives after a two-minute peak. You pay for the machines added, the service saturates anyway during the peak, and you are then left with machines you do not need.

Three settings decide the outcome, and the last one is the one people forget:

SettingWhat it doesThe value that works
The triggering metricchooses what you react torarely the processor alone: queue length or response time reflect the real pain better
The cooldownprevents chaining decisionslong enough for an added machine to have taken effect before the next decision
The preparation timetime before a new machine servesto be reduced first, by preparing a complete image rather than installing at startup

Reducing the preparation time is worth more than fine-tuning the thresholds. An image that already contains the application can start in one minute where an installation at startup takes five, and that difference alone turns a decorative mechanism into a useful one.

Interruptible instances: up to 90% cheaper

Providers resell their unused capacity at a knock-down price, on one condition: they can take it back from you. AWS advertises up to a 90% discount compared with the on-demand price on its Spot Instances page, read on 14 November 2026; the real discount depends on the provider, the machine type and the moment. It is a very powerful savings lever, often poorly exploited.

The notice before reclaim is short, and it differs from one provider to another:

ProviderCommercial nameNotice before reclaim
AWSSpot Instancestwo minutes
Google CloudSpot VMsshutdown signal followed by thirty seconds, best effort, to stop (120 seconds of notice optionally before that signal)
AzureSpot virtual machinesthirty seconds

This notice strongly limits the possible uses. Between thirty seconds and two minutes, you can finish a short task and hand over cleanly; it is not enough to drain a session server. The question to ask is therefore not "can I save money", but "what happens if this machine disappears right now, in the middle of its work?".

The workloads that suit it all share the same property, that of being resumable:

  • Batch processing split into short tasks, where losing a task means rerunning it elsewhere.
  • Tests and continuous integration, where a failure on reclaim results in a new run, not an incident.
  • The extra nodes of a compute cluster, provided the minimal baseline stays on standard machines.
  • Model training with regular checkpoints, where you restart from the last checkpoint rather than from the beginning.

What does not suit it is everything that holds state that nobody else holds: a database, an in-memory session server, the single instance that serves your site. The reasoning is the same as for the maintenance imposed by the provider, described in operating a cloud environment: an architecture that survives the abrupt disappearance of a machine opens up savings that a fragile architecture cannot take.

When should you leave the shared model?

By default, your virtual machine shares a physical server with those of other customers. That is what makes the cloud affordable, and it is acceptable in most cases. Three reasons justify leaving it, each towards a different and clearly more expensive offer.

What you are looking forThe matching offerWhat it really changes
Not sharing the hardware with other customersdedicated host or dedicated instanceisolation becomes physical, which some audits require
Counting licences per physical processordedicated host, with visibility on the socketsa software licence billed per socket, that is, per physical processor, becomes computable, and sometimes cheaper
Removing the virtualisation layer entirelybare metal serveryou recover raw performance and hardware features, you lose the snapshot and the recreation in one minute

Bare metal is paid for mostly in flexibility. The machine is not recreated in a few seconds, its provisioning takes minutes or even hours, and the snapshot mechanisms you have grown used to disappear. Consider it only if you have measured the gain, not if you assume it.

Graphics cards deserve a separate warning. An instance fitted with an accelerator costs far more per hour than an ordinary machine. At Scaleway, the smallest L4 instance, L4-1-24G, was listed at 0.79 EUR excluding tax per hour in zone fr-par-1 on the public price list of 12 October 2026. This instance is subject to tight quotas, and its availability is not guaranteed in every zone.

A banal drift costs a lot: a machine switched on for a trial and forgotten on Friday evening runs all weekend, about 62 hours. That makes about fifty euros for this smallest L4, and hundreds of euros or more for a machine with several cards. The exact amount depends on the model and the provider: calculate it with the price page before starting. Set the budget and the alert before starting the first machine, not after.

Key points

  • The displayed price is the highest of the three purchase modes, but a commitment is only taken on the measured baseline, not on a size that is still moving.
  • Switching off stops the compute, not the volumes or the IP addresses. Delete what you no longer use.
  • Autoscaling is decided on the preparation delay, not on the threshold: prepare a complete image.
  • An interruptible instance suits what resumes without state, and not what holds a unique state.
  • The shared model is enough in most cases; a dedicated host or bare metal is justified by a requirement, not by a reflex.

Knowledge check

Check that you have grasped the essentials of this guide. The questions cover only 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