For inference you could use a maxed-out Mac Ultra; the RAM is shared between the CPU and GPU.