Apparently they've built the NPU in such a way that it can use existing PCIe slots and they have built a translation layer to make it behave like any other GPU. Theoretically meaning that GPU-heavy workloads don't need to be completely rewritten for their NPU.