The real magic, though, is flexibility: you don't have to bunch all of your server purchases into an annual (or worse) exercise, execute a buildout at a data center, and then somehow allocate them to internal teams consuming them prior to the next capacity planning exercise. You just fire API calls, and a bill arrives at the end of the month. The bill may be a lot of money, but the amount you were spending on planning, building, operating, fixing, and decommissioning your infra was a lot of money, too, split between big spiky equipment purchases and employee salaries with a dusting of leases and electricity on top.