Skip to content

Independent marketplace for verified cloud accounts

Reserved, Spot and On-Demand Pricing Compared

The same machine sells three ways at three prices. What changes is the commitment you make and the interruption risk you accept. How to match model to workload.

The same virtual machine, with the same specification, in the same region, can be bought three different ways at three very different prices. Nothing about the hardware changes. What changes is the commitment you make and the interruption risk you accept, and the provider prices those exactly as an options market would.

Understanding the three models is probably the best return on an hour of cost reading, because it applies to capacity you have already decided to buy. This is not about using less. It is about paying less for the same thing.

Three ways to buy the same machine

On demand means you pay a published rate per unit of time, you can start and stop whenever you like, and the provider guarantees capacity for as long as you keep paying. Maximum flexibility, highest rate.

Commitment pricing, sold under names like reserved instances, savings plans and committed use discounts, means you promise to spend at a certain level or use a certain amount of capacity for a fixed term, typically one or three years. In exchange you get a materially lower effective rate. Less flexibility, lower rate.

Spot or preemptible capacity means you bid on the provider’s unused inventory, you get a steep discount, and the provider can reclaim the machine at short notice when a paying customer wants it. Lowest rate, no guarantee of continuity.

That is the whole framework. Everything else is detail about how each provider implements it, and the implementations differ enough that the names are not portable between platforms.

On demand: paying for optionality

On demand is the default and it is the correct choice more often than cost optimisation enthusiasts admit. You are buying the right to change your mind, and early in a project that right is worth a great deal, because you do not yet know your instance shape, your region, your scaling pattern or whether the project will exist in six months.

It is also the right choice for genuinely intermittent work. A machine that runs four hours a day cannot benefit from a commitment that bills continuously, and a workload that doubles and halves with the season is better served by paying the premium during the peak than by committing to the peak all year.

Where on demand becomes a mistake is the steady state that nobody revisited. A production fleet that has run the same shape for eighteen months at on demand rates is leaving a substantial saving on the table for no reason other than inattention. That is not a pricing problem, it is a review problem.

Commitments: trading flexibility for a lower rate

Commitment products vary along several axes, and knowing which axes your provider offers is most of the skill.

Term length is the obvious one: longer terms give deeper discounts. Payment structure is next, with full upfront, partial upfront and no upfront options trading cash flow against rate. Then scope: some commitments are tied to a specific instance family, size and region, which gives the best rate and the least flexibility, while others commit only to a level of spend and apply automatically to whatever you run, which is more forgiving and discounts less. Some products allow you to exchange or modify a commitment mid term, some do not.

The decision rule is straightforward once you have measured. Find your baseline, meaning the capacity level you have been at or above for essentially the entire measurement window, and commit to a comfortable fraction of it. Leave the variable portion on demand. Teams that commit to their average rather than their floor end up paying for capacity they are not using, which eats the discount they were chasing. Getting that floor right depends on knowing your actual utilisation, which is what our guide to right sizing cloud instances is for.

One structural note: not every provider sells commitment products at all. Operators who price on simple fixed monthly plans effectively give you something close to the committed rate with no commitment, which is a real advantage that never shows up in a discount comparison. Accounts such as Buy Hetzner Account and Buy Linode Account work this way, and the AWS and Linode comparison shows how a flat plan stacks up against a discounted hyperscaler rate.

Spot and preemptible capacity: renting the gaps

Spot capacity is the provider selling idle inventory. The discount is the largest of the three models by a wide margin, and the condition is that you can be evicted, usually with a short warning measured in seconds to a couple of minutes, sometimes with no warning at all depending on the platform.

That condition is fatal for some workloads and almost irrelevant for others. Batch processing, media encoding, test suites, data pipeline stages, model training with checkpointing and any stateless worker pulling from a queue can all tolerate eviction, provided the work is idempotent and progress is saved outside the instance. Anything holding state in memory for a long time, anything with a hard deadline, and anything where a partial result is harmful should stay off spot.

Pricing behaviour differs between platforms in a way that matters for planning. On some, the spot rate floats with supply and demand, so your cost varies through the day and a capacity crunch can push it close to the on demand rate. On others the discount is fixed and only availability varies. The first kind needs a maximum price set deliberately, since leaving it at the default means accepting whatever the market does. The second kind is simpler to budget for and gives you no control over whether you get a machine at all.

Three practices make spot workable. Checkpoint frequently enough that an eviction costs you minutes rather than hours. Spread requests across several instance types and zones, because availability varies independently between pools and a diversified request is far less likely to be fully evicted at once. And handle the termination signal properly: drain connections, finish or abandon the current unit of work cleanly, and let the orchestrator reschedule.

Matching purchase model to workload shape

Classify every workload you run into one of four shapes and the purchasing decision mostly makes itself.

  • Always on, stable size. This is your commitment base. Databases, core API tiers, anything that has not been off in a year.
  • Always on, variable size. Commit to the floor, run the variable portion on demand, and consider spot for the top slice if the workload is stateless.
  • Scheduled and interruptible. Spot, with checkpointing. Nightly batch, reporting, build farms, test runners.
  • Unpredictable and short lived. On demand, or a serverless model where you pay per execution and the idle cost is zero.

Development and staging environments deserve special mention: they are frequently on demand, always on, and used eight hours a day by a handful of people. Scheduling them off outside working hours saves more than any pricing model, and costs nothing but a scheduled task.

The commitment mistakes that hurt

Committing before you have measured is the most common. You need a minimum of a month of real utilisation data, and three months is better, because the pattern you see in week one is usually the pattern of a system still being built.

Committing to a narrow scope right before an architectural change is the second. If a migration to containers, a different instance family or a different region is on the roadmap within the term, buy the flexible commitment product even at a shallower discount, or wait.

Third, committing in a currency or a region you are about to leave. Fourth, stacking commitments without tracking coverage, so that overlapping purchases exceed your actual baseline and the surplus is simply wasted. Fifth, and most avoidable, letting a commitment lapse silently at the end of its term and reverting to full rate without noticing for two months.

A mix that holds up over a year

A sensible portfolio for a team with a year of operating history looks something like this: commitment coverage for the stable floor, deliberately set below the measured minimum so that it is always fully used. On demand for the band between the floor and the normal peak. Spot for everything batch shaped, interruptible and idempotent. Scheduled shutdown for all non production environments. And a calendar reminder a month before every commitment expires.

Review the mix quarterly rather than continuously. Purchase model optimisation has real savings in it, and it also has a strong pull toward spending engineering time on arithmetic instead of on the product. Set the mix, write down why, and revisit it when the architecture changes. For the wider question of how these choices fold into a forecast, read it with our guide to predictable pricing versus pay as you go, and browse the account catalogue if a flat rate provider would simplify the whole exercise.

Related comparisons

Hetzner vs Contabo
vs

Hetzner vs Contabo

Both are German hosts competing on low prices, and Contabo is usually the cheaper number on the page. The useful…

Hetzner vs OVH
vs

Hetzner vs OVH

Two large European providers that own their data centres but sell very different things. This compares product breadth, bare metal…

Hetzner vs Digital Ocean
vs

Hetzner vs Digital Ocean

Hetzner sells hardware value from a mostly European footprint. DigitalOcean sells a developer platform with managed add-ons and deep documentation.…

Hetzner vs Linode
vs

Hetzner vs Linode

Both platforms were built by engineers for engineers, but they answer to different owners now. This compares operating philosophy and…

Accounts discussed in this guide

Buy Hetzner Account
Buy Hetzner Account
Hetzner

Buy Hetzner Account

Hetzner has earned a reputation among developers as one of the most powerful and affordable cloud platforms in the world.

From$50 3 options
Buy Linode Account
Buy Linode Account
Linode

Buy Linode Account

Linode has long been regarded as one of the most dependable and developer-friendly cloud hosting platforms in the industry, and its acquisition by Akamai…

From$25 2 options

Keep reading

Reading Cloud Storage Performance Specifications
Infrastructure 7 min read

Reading Cloud Storage Performance Specifications

IOPS, throughput and latency are three promises, not one. How to read a storage spec sheet so the performance you get matches…

Need the account, not just the theory?

Browse documented configurations across every provider in the catalog, or ask us which one fits your workload.

Telegram