The simple answer is: no!

Think about OpenCL and in general stream computing - which I believe is a the future of computing through a slow convergence of multicore and manycore. Having a more tightly coupled CPU and GPU (i.e. no PCI express bottleneck) would have the huge advantage of being able to do more streaming tasks on the GPU while not having to worry that communication will kill the performance.

There are already quite many GPU accelerated applications in various HPC research and industrial applications as well as desktop applications (just a few examples: Adobe CS5, OS X Snow Leopard, Badaboom media converter,...)

Interestingly if you enter proper sentences, e.g. at least use punctuation marks, it doesn't go nuts anymore. Only one "." makes this much difference: http://translationparty.com/#6834172

I think this benchmark puts the bar a bit too high. First of all, a translator is not designed to produce invertible translations. Moreover, as the goal is to produce an understandable translation of a human-written text, the artifacts introduced by the machine-created translation are most probably magnified quite a bit with the second round of translation. Still, it's interesting to see that the Google algorithms actually do an OK job even in such an artificial benchmark.