Amazon has been selling storage by the gigabyte-month since early last year and virtual servers by the hour since the summer after. The servers are still officially in beta, which has not stopped a surprising number of people from putting real things on them.
Most coverage treats this as a bookshop getting into hosting, which makes it sound like a diversification stunt. It is not. It is a bookshop noticing that it already owns the thing.
Think about what Amazon has to build to survive December. Retail traffic in the last six weeks of the year is a multiple of what it is in February, and you cannot buy capacity in November for a peak you hit in December. So you build for the peak, you run it all year, and for ten months of the twelve a large fraction of that hardware is sitting there drawing power and depreciating while it waits for Christmas.
Selling the idle months is not a new business. It is a way of making an existing cost less painful.
That is why the pricing looks the way it does. Ten cents an hour for a small instance is not a number you arrive at by costing out a hosting product and adding margin. It is a number you arrive at when the alternative is the machine earning nothing at all. Traditional hosts have to price to cover the whole asset. Amazon only has to price above the marginal cost of the hours nobody else wanted, and no dedicated host can meet that without losing money on purpose.
For a small operation the effect is genuinely new. Spinning up ten machines for an afternoon and then destroying them used to require a purchase order, a data centre contract, and a conversation about a twelve month minimum. Now it requires a credit card and an API call. That is a real change in what a two-person company can attempt.
The part I keep circling back to is what happens when it stops.
Your own server failing is a bad afternoon and it is yours to fix. A shared platform failing is a bad afternoon for everyone on it simultaneously, and none of you can do anything except read the status page and wait. The terms of service are, as far as I can tell, a fairly thin document about best effort. There is no meaningful compensation and no escalation path that a small customer can actually use.
Nobody has properly tested this yet, because nothing large has broken for long enough. The service has had wobbles, and they have been short, and everyone has shrugged and moved on.
My guess is that the day this gets taken seriously as a risk is not the day someone writes a thoughtful piece about single points of failure. It is the day a few thousand sites go dark at once during working hours, and a lot of people discover what they agreed to when they clicked through.
Until then the price is very hard to argue with, and I am using it too.