Maybe Google will let you buy some nodes. Yes, I’m kidding, but it is an interesting example of a distributed process across thousands of cheap hardware boards.
I wonder whether BOINC would let you use some of their distributed volunteer crunchpower?
I get you on the data transfer, network speed is almost always going to be a bottleneck for this kind of CPU speed.
Power consumption and PSU cost is going to be ugly on those multi chip boards anyway. The power supplies alone might cost as much as another dual core box for the cluster.
Blades are great for space consious but can be pricey. Basic rackmount stuff might be a possibility for you too. You might even be able to pick up some used stuff that will get you some good bang for your buck while reducing footprint