On the hyperscaler platforms, an identical virtual machine can carry three or four different prices. The difference is not the hardware. It is what you have promised the provider in exchange.
On-demand
The default. You pay by the second or the hour, you can stop at any time, and you are guaranteed capacity while it runs. It is the most expensive per hour and the only model with no strings attached.
Right for: anything you cannot predict, development environments, the first months of a new workload, and any service where an interruption is unacceptable and the volume is small.
Reserved and committed use
You commit to a quantity of capacity — or, on some platforms, to a level of spend — for a fixed term, typically one or three years, in exchange for a substantial discount. Paying more of the term upfront usually deepens the discount.
The mechanics differ by provider and matter more than the headline percentage:
- What is committed. A specific instance type in a specific region is less flexible, and usually cheaper, than a commitment to an amount of compute you can spread around.
- Whether it can be changed. Some commitments can be exchanged for a different type; some cannot.
- Whether it applies automatically. Most apply as a billing discount to matching usage rather than reserving a particular machine.
Right for: steady baseline load you are confident will still exist in a year. The mistake is committing to peak rather than baseline, which locks in capacity you use for two hours a day.
Spot and preemptible
Spare capacity, offered at a steep discount, reclaimed by the provider — usually with a short warning — whenever it is needed elsewhere. The saving is large. The condition is that your workload must tolerate being terminated mid-run.
Right for: batch processing, CI builds, video encoding, model training with checkpointing, and any queue-driven work that can restart. Wrong for: databases, anything holding session state, anything a user is waiting on.
Making spot work is mostly an engineering question — checkpoint often, make jobs idempotent, handle the termination signal, and spread across instance types so a single reclamation does not take everything.
The flat-rate alternative
Several providers in this catalogue do not sell commitment discounts at all: the plan has a published monthly price and that is the price. That is not a worse deal — it is a different one. The comparison that matters is the flat monthly figure against the discounted hourly figure you would realistically qualify for, not against the on-demand rate you would be trying to avoid paying.
AWS vs Hetzner and Azure vs DigitalOcean are both, at heart, this comparison.
A sensible sequence
- Run on-demand or flat-rate while you learn the real shape of the workload — a month at minimum.
- Identify the baseline that is genuinely always running.
- Commit to that baseline only, on the shortest term that still gives a worthwhile discount.
- Move interruptible work to spot, with checkpointing.
- Leave the variable peak on-demand, or handle it with autoscaling.
Committing before step two is how organisations end up paying for three years of capacity they stopped using in month four.
Where to go next
- Cutting cloud costs without downtime
- Estimating cloud costs before you build
- How cloud billing works
Questions
Frequently asked
Are reserved instances always cheaper?
Only if the usage actually persists for the term. A commitment sized to peak rather than baseline frequently costs more than on-demand, because you pay for capacity that sits idle most of the day.
What happens when a spot instance is reclaimed?
The provider sends a short termination warning and then stops the instance. Workloads must checkpoint and be safe to restart; anything holding state a user depends on is unsuitable.
Do all cloud providers offer commitment discounts?
No. Several providers use flat published monthly pricing instead. Compare the flat figure against the discounted rate you would realistically qualify for, not against the on-demand rate.