> Where's AMD in all of this? Nowhere.

They're around, selling AMD Strix Halo (up to 128GB unified memory) on the low, equivalent to DGX Spark, level; and MI* chips for enterprise AI inference.

The problem they have and they've always had is that everyone uses CUDA. ROCm works, but everyone's toolchain is CUDA exclusive or CUDA first, so it's much harder to use AMD hardware.

Using higher level tooling (vLLM, llama.cpp) I have DGX Spark comparable performance for 1/3 the price (but without the same software out of the box experience nor the clustering network options).