A team buys a multi-year spend commitment expecting machines to be waiting in its standby zone — what has it actually bought?

## Two promises that sound alike and are not Both instruments are bought in advance, both attract the word `reserve` in ordinary speech, and both appear on the same bill. They promise different things. A **term commitment** is a promise about **spend**. You agree to consume a stated amount — a quantity of compute, or an hourly spend — over a stated period, and the provider discounts what you consume. It is a financial arrangement layered over usage. If your launch is refused, the discount has nothing to apply to. A **capacity reservation** is a promise about **allocation**. Capacity matching a specific shape, in a specific zone, is set aside for your account. No other tenant may be placed on it. You pay for it from the moment it exists, used or not, and a launch that matches it is drawn from stock that was never in the shared pool. ## The comparison in one table | | Term commitment | Capacity reservation | |---|---|---| | What it promises | a price | an allocation | | What you are exposed to | consuming less than you committed to | reserving the wrong shape, zone or amount | | When billing starts | as usage occurs, at the discounted rate | when the reservation is created, idle or not | | Does it help a refused launch | no | yes, where the launch matches it | | Typical scope | an account or organisation, often flexible across shapes | a named shape in a named zone, matched narrowly | The two are **orthogonal, not alternatives**. On many platforms the usage of a reservation is still eligible for the committed rate, so buying both is normal and sensible: the commitment answers `what will this cost`, the reservation answers `will it be there`. Designs differ in the details — how narrowly a reservation matches a launch, whether it can be shared with other accounts in the same organisation, whether it can be time-bounded — so confirm the matching rules on the platform in front of you rather than assuming the ones you learned first. ## Why an idle reservation is not cheap The intuition that an unused reservation should cost little is the common mistake here, and the reason it is wrong is physical. A held allocation is hardware the provider may not sell to anybody else. Somebody is paying for a machine that exists and is idle, and it is going to be you. Providers therefore price a held allocation close to the price of running the machine. That has a design consequence worth stating plainly: if holding the allocation costs about what running the machine costs, then for many workloads the better arrangement is simply to **run the machines**. Warm machines take the launch off the critical path entirely, they are continuously proven to work, and they absorb load immediately. A reservation still wins where running is not free of other costs — per-machine licensing, data that must not be live in two places, a workload whose mere presence would cause problems — or where the allocation is held for a dated event rather than for an unpredictable failure. ## How reservations fail - **Wrong zone.** The reservation sits in the zone the failover leaves, not the one it enters. - **Wrong shape.** Matching is narrow. A launcher walking a preference list draws from the reservation only on the entry that matches it; every other entry is an ordinary contended launch. - **Wrong amount.** Held to the peak fleet it is unaffordable; held as a token it serves nothing. - **Orphaned.** The architecture moved to a different family a year ago and the reservation is still billing. - **Mistaken for a discount.** A team buys reservations for the steady baseline expecting a lower unit price, and gets certainty it did not need at a price it did not intend. ## What each instrument is actually for Buy a **commitment** for the part of the fleet whose consumption you are confident about over the committed period — steady baseline capacity that will exist whatever happens. Buy a **reservation** for the small part of the fleet whose *launch must not be refused*: the minimum serving set in a standby zone, the capacity behind a dated event you cannot move, the machines a regulated workload must be able to produce on demand. Everything else is better served by shape flexibility and a degradation plan, which cost engineering effort rather than money every month. The test that separates them, in an interview and in a review, is one sentence: **a discount changes what capacity costs; a reservation changes whether it is there.** A plan that needed the second and bought the first discovers the difference in the one hour it cannot afford to.