On your own hardware, cost is a one-time capital decision; in the cloud it's a continuous *capacity variable* you must measure like any other. Metered "use-it-or-lose-it" pricing means running instances below peak is pure waste, and costs "rise surprisingly fast" once you can launch 100 instances at once. Two consequences specific to cloud capacity: - **The instance menu is coarse.** Providers offer fixed tiers (small → large multi-core), not the off-the-shelf customization you'd spec yourself. You must fit your needs to the menu — and many small instances often beat a few large ones, but it's per-application (the [[Capacity Test Differs by Instance Class]] caution applies). A menu gap can even force the wrong size until the provider adds a tier. - **The approval gap.** Buying 10 servers needed management sign-off; launching instances that sum to the same cost over time often doesn't. Track cloud usage *and spend* in the capacity process to close that gap. --- *Source: [[The Art of Capacity Planning]] (John Allspaw, O'Reilly 2008) — App A — Virtualization and Cloud Computing*