I think serverless currently means " I don't have to think about servers, provisioning or scaling". It's === to function-as-a-service. S3 is the best example so far since people see it as no servers at all, just static file serving (there USA server running S3, but as a user we don't really care)
With traditional infrastructure you have to think about capacity in discrete units of hardware. Should I provision one or two servers? Even if you can start and stop servers with an api call, you always have to keep these servers running, because rampup time is in the order of minutes.
At Graphcool we have thousands of serverless functions. Most of the time they are not running, but they are all ready-to-go wit delay measured in milliseconds, and they scale individually unbounded of the number of servers you have provisioned up front.
It's quite remarkable to think about the flexibility this characteristic afford you when architecting your applications.
