I suspect your traffic will grow exponentially for a year or two and every couple months or so there will be a mad scramble for increased capacity. If affordable Amazon EC2 would be ideal as you can add capacity as needed. Alternately this may be a perfect opportunity to experiment with an elephant in a bramble: a hadoop map/reduce cluster! Map tasks run on individual 'Pis in the cluster, Reduce task(s) on the master node, which can be a hefty x86 server. Pi->Pi comm. will be much less than Pi<->master node so this should be a fairly scalable solution. 100 'Pis will give you 400MB/s to 1GB/s peak flash disk bandwidth, 10Gbps Pis<->master node bandwidth, 140,000 bogoMIPS processing BW (more, if GPUs can be deployed) etc. Probably $5K to $6K H/W cost, $500/year electricity + Internet pipe cost + management costs. Hm... not ideal and it will take a few months but 'Eating your own dog food' is also the best way to shake out bugs!