On demand, reserved, spot and committed use: cloud pricing models compared

The same virtual machine can cost four quite different amounts depending on how you buy it. Getting that choice right is usually worth more than any amount of code tuning, and it takes an afternoon rather than a quarter.

On demand

The list rate, no commitment, billed by the second or the hour depending on the provider and service. You can start and stop freely.

Correct for anything unpredictable, anything short lived, and anything you are still measuring. It is also the right default for the first three months of a new workload, because committing before you know the shape of your usage is how people end up paying for capacity they never use.

Commitments

All three major providers sell a discount in exchange for a one or three year promise, and the shapes differ enough to matter.

AWS offers savings plans, where you commit to an hourly spend and the discount applies across eligible compute, and reserved instances, where you commit to a specific instance family in a region for a deeper discount.

Azure offers reservations for a specific virtual machine series in a region, and a savings plan for compute that commits to an hourly amount and applies flexibly. Hybrid Benefit stacks on top for Windows and SQL Server workloads.

Google Cloud offers resource based committed use discounts, which lock vCPU and memory in a region for the largest reduction, and spend based commitments that apply more flexibly. It also applies sustained use discounts automatically with no commitment at all, which is unique among the three.

The pattern is consistent: more specificity buys a deeper discount, more flexibility costs some of it.

Spot and preemptible capacity

Providers sell spare capacity at a large discount and reclaim it when they need it, with a short warning. AWS calls it spot, Azure calls it spot virtual machines, Google Cloud calls it Spot VMs.

It works well for:

• Batch processing and data pipelines that checkpoint.
• CI and build fleets.
• Rendering and encoding.
• Stateless web tiers behind an autoscaling group with a small on demand base.

It works badly for databases, anything holding a long lived session, and anything where a two minute eviction notice is not enough time to move.

The practical pattern is a blended group: a small on demand or committed base for reliability, spot for the majority, and an autoscaler that replaces reclaimed instances. That captures most of the discount with very little fragility.

Flat rate providers

The smaller providers in our catalogue take a different approach entirely. DigitalOcean, Linode, Hetzner, UpCloud and others publish a flat monthly price with an hourly cap, and that is the price. No commitments, no discount tiers, no optimisation programme.

That is worth something real. A team without a dedicated cost function will frequently spend less on a flat rate provider at list than on a hyperscaler where nobody claimed the discounts. Predictability has value that does not appear in a rate comparison.

The costs that no pricing model reduces

Commitments cover compute. They do not cover the line items that most often surprise people:

• Data transfer out. Rarely discounted meaningfully and frequently the largest single item. See data transfer costs.
• Storage volumes billed on provisioned rather than used capacity.
• Managed service premiums, which are usually charged at list.
• Logging and monitoring ingestion, which scales with verbosity rather than with traffic.
• Idle resources. A load balancer with no traffic, a disk with no instance, an address with nothing attached.

Deleting an idle resource beats discounting it every time.

A sequence that works

1. Run on demand for a quarter and measure.
2. Delete what is idle. This is usually the largest single reduction and it costs nothing.
3. Right size what remains, using actual utilisation rather than the original guess.
4. Move interruptible work to spot.
5. Commit to the baseline that is left, starting with a one year term.
6. Review quarterly, because workloads move and commitments do not.

Provider specific mechanics are in how Azure pricing works and Google Cloud pricing explained. If credits are part of the picture, how cloud credits work covers how they interact with all of this.