Friend of doppenhe here.

Amazon, heroku etc aren't fast enough for us due to spin-up time. Basically when we get a particular command from a user, we need to do a LOT computation on many parallel machines, but for a short burst of time.

Keeping a whole cluster of machines reserved would be prohibitively expensive at this stage. We would prefer to keep a small base-load capacity reserved, and then spin up a large number of instance for a short time. Response time is key here.

It would be best if they were already up and running, waiting for a command, but idle most of the time. Fine if it's shared with others. Kind of like a computation insurance for short bursts of high-demand.

Anyone know any services like that?