Also, at web scale encoding time isn't nearly as important as the ability to go parallel which is significantly cheaper and easier to scale using a bunch of EC2 instances.
We process 2-300 videos a month, though during competitions we process 10x that in a month. All of the bandwidth, storage, processing is less than $10K a year with an average encoding FPS of 33fps.
I would love pointers to sources if this isn't the case and I can do this for massify cheaper. Would be appreciated.
EDIT: We also encode 2-3 different versions of the video at varying sizes and bitrates.
Have you tried Fluendo? I think it has GPU support to a certain extent. Not sure if this is a help, but here you go:
For example http://www.nvidia.com/object/cuda_home.html#state=detailsOpe...
The irony is that GPU is a more powerful processor than CPU, just not as versatile and easy to use (only a small subset of programming tools can use it yet) but it changes.