Tesla's AP is seemingly built ontop of Nvidia hardware, which is programmed using CUDA. CUDA is built ontop of LLVM, which Chris Lattner created. Seems like an obvious fit if you look at it from a purely technology perspective.
No way is it a leap. Lots of people are doing stuff to llvm to do gpu work. Take mapd for instance. There is a LOT of untapped llvm -> gpu goodness to be had.
