I hadn't appreciated before that Gentry's work was both general (any computation can be performed as a boolean circuit of addition and multiplication) but also impractical : a trillion fold increase is computational burden for an encrypted web search? That is going to kill the advantages of distributed computation.
Schneier also opened my eyes to fact that previous existing encryption schemes are homomorphic under certain operations, such as RSA under multiplication.
I wonder if there would be any gain to distributing parts of a computation and then reassembling it rather than having a fully remote computation done under homomorphic, but very costly, encryption. Of course the gains of parallel computation would have to more than offset the costs of distribution, but that is always the case for parallel applications.