I'm guessing the High Performance Computing guys might be interested as well.
I'd imagine that it depends on how heavily current GPU/CPU compute systems lean on the 'CPU' side of the arrangement:
If the CPU actually keeps reasonably busy(either with aspects of the problem that aren't amenable to GPU work, or with assorted housekeeping tasks required to keep the GPUs fed and coordinated across the cluster), Intel or AMD offer pretty good prices for chips that provide a lot of PCIe lanes, support tons of RAM, and are supported by most of the world's horrid legacy software. Plus, motherboards and other supporting gear are brutally commodified, which is always nice.
If the CPU is mostly idle, and mostly gets included because it's the cheapest way to get a bunch of PCIe lanes and boot an OS that can run CUDA drivers and a NIC, then a Tesla-like card that includes a weedy little ARM core and can run on a simple backplane, without any PC server components, would seem like a logical thing to produce.