For an enterprise: autoscaling is either not a value add ("You can cut $40 a month off our $300 hosting bill? Absence of 'yay'.") or Really Freaking Hard And App Specific. SmugMug, as I recall, did a lot of work to dynamically spin up and shut down instances to minimize their bill (not a Heroku app, obviously), because their usage patterns consistently predicted high load for workers at certain times and low loads at other times.

For a ramen-munching startup: shaving $40 off our bill is very motivating. However, while we have dreams of scaling up, if you actually do it we're going to wig out when we see a four figure hosting bill. By the way, we're pathological customers who understand just enough of this to be a danger to ourselves, and we will blame you on the Internets when your system does exactly what we tell it to do.

I see your point, but why would a startup wig out when they scale? Presumably, revenues would scale with demand, so I would be quite grateful that my provider allowed me to satisfy the demand quickly and reliably...

Real-life example of what Patrick is talking about: http://news.ycombinator.com/item?id=1688904

Oh, now I see. It's not so much about scaling up, then, as it is about forgetting to scale back down after a spike.

Because its profit margins depend on clients provisioning at least a little more capacity than they need.

Because it has many other things to do.