I started on Amazon grew to something like 30 servers on there doing a variety of tasks in a few groupings(Server group A has 5 servers all the same + load balancer for example). Used scalr.net to manage it and thought it was great.

I've recently started moving everything over to GoGrid(20+ servers there now) since I needed to be on the west coast for latency issues. I'm not as big of a fan of their services but it's ok for what it is. I still have a number of complaints though not as many as last time I wrote about it.

I also have a hand full of servers with rackspace cloud for redundancy but they get very little use at the moment.

Also our whole external infrastructure is now in the cloud. We have one server with softlayer and a mail server that are separate. In house we have another 15 or so servers for log processing.

We use autoscaling or rather did when with amazon using scalr and once I get my puppet setup running fully on GoGrid I'll be developing one there too. The problems I face can entirely be solved just by adding more hardware since all my systems are modular so it works well for me but your mileage may very.

Scalr support autoscaling by things like load/ram/bandwidth as well which is nice since a lot of times you need to scale out for different reasons.