This is hard because you need high bandwidth between the GPUs in your cluster, bandwidth far higher than broadband could provide. I'm not even sure whether the time spend synchronizing between far-away machines would offset the increase in computational power.