Pricing Models & Commitments

## The three axes AWS compute pricing is not a ladder of tiers; it is three independent ways of buying the *same* underlying capacity. Each buys a discount by taking a specific risk off Amazon's hands and putting it on yours. ## On-Demand: the list rate On-Demand is what you pay if you do nothing. You launch an instance, you are billed for the time it runs (per second, with a one-minute minimum, for most Linux instances), and you stop being billed when you terminate it. Nothing is promised in either direction: you can walk away at any moment, and AWS charges the published rate for the region, instance type and operating system. On-Demand is the correct answer more often than cost-conscious engineers admit. Workloads that run for a few hours a week, environments that may be deleted next quarter, and anything whose shape you cannot yet predict all belong here. Buying a discount on capacity you later stop using is a loss, not a saving. ## Commitments: you take the demand risk A commitment is a promise about the future. Two products implement it: - A **Savings Plan** commits you to a fixed dollar amount of compute usage per hour — say $12/hour — for one or three years. Any eligible usage up to that amount is billed at discounted rates; usage above it falls back to On-Demand. - A **Reserved Instance** commits you to a specific instance configuration (family, region, platform) for one or three years, and discounts matching usage. Both come with payment options — All Upfront, Partial Upfront, No Upfront — where paying more of it in advance buys a slightly deeper discount. The essential property is that the commitment is a *floor on your bill*: if your usage falls below the committed amount, you still pay it. That is exactly why the discount exists. You have taken the demand risk away from AWS, and AWS pays you for it. Commitments also do not, by themselves, guarantee that capacity will be available when you ask for it. A Savings Plan is purely a billing instrument. ## Spot: you take the interruption risk Spot sells capacity AWS is not currently selling at On-Demand rates, at a discount that is typically the deepest of the three. The price of that discount is that AWS can reclaim the instance when it needs the capacity back. Spot therefore belongs to workloads that can be interrupted and retried — batch processing, CI runners, stateless scale-out fleets behind a queue — and never to a single stateful instance you cannot afford to lose. (The mechanics of interruption handling belong with EC2 purchase options; here Spot matters as the third pricing axis.) ## How they layer The experienced answer is that these are not alternatives to choose between once, but layers stacked over a usage graph: 1. Plot hourly compute usage over a representative period. 2. The flat *trough* — the amount you are certain to consume every hour of every day — is the commitment target. 3. Interruptible work above the trough goes to Spot. 4. What is left, the spiky and unpredictable part, stays On-Demand. That ordering also explains a common mistake in reverse: commitments are a discount on the architecture you have, so if a workload is oversized, idle overnight, or about to be rewritten, fixing that first lowers the baseline you are about to commit to for three years. ## Order of application A useful detail: discounts are applied automatically at billing time, not selected at launch. You do not launch a "Savings Plan instance" — you launch an instance, and AWS matches your commitments and reservations against the usage on the bill. That is why a commitment purchased in one account can benefit usage elsewhere in the same billing family, and why a badly targeted commitment silently attaches to whatever usage it matches rather than failing loudly. ## What interviewers are listening for They want to hear that the discount is compensation for a transferred risk, and that you can name which risk each option transfers. A candidate who says "Reserved Instances are cheaper, so use those" has skipped the entire question. A candidate who says "commit to the baseline, Spot the interruptible tier, On-Demand the rest, and rightsize before committing" has answered it.