Hi Chris,
Our current setup utilizes two EC2 instances of the c1.Medium type which provide 5 EC2 Compute Units and 1.7GB of RAM each. Again these are 32-bit servers as once the tile cache of the server is populated with the most visited areas we are not seeing a large CPU or Memory utilization. For disk this is a bit of a judgement call as there are advantages and disadvanteges to utilizing the local 'instance storage' or the EBS volumes. For our needs we found that the EBS volumes were more fault tolerant and provided the ability to recover from failure much easier so we are hosting all the tile data from EBS volumes.
We setup four Elastic Load Balancers each accessing these two instances. We attempted several configurations but found that Internet Explorer was only able to handle the number of connections provided by the four load balancer setup. Any additional load balancers simply were not accessible by IE and thus did not increase performance.
With each Elastic Load Balancer you are provided with an external DNS name so that you can access the server attached to that load balancer via a URL, but these are not very 'friendly' DNS names so we utilzed our hosting providers DNS to map a friendly name to each of the load balancers, something like wmk1.thinkgeo.com to elb.us.amazonaws.com, wmk2.thinkgeo.com and so on. These urls were then added as wmsuris using the same type of code found here: gis.thinkgeo.com/Support/Dis...fault.aspx
AWS has worked very well for use as you can easily scale up your servers to meet your needs and the Elastic Load Balancers will automatically scale their internal resources to match the load.
Hope this helps.