I'm guessing what they meant is that they use toolchains that are retargetable to other GPUs (and typically compile down to PTX (nVidia assembly language) on nVidia GPUs rather than go through CUDA source -- GCC and clang can both target PTX). For example XLA and most SYSCL toolchains support much more than nVidia.