I don't see why Intel couldn't make high end PCI cards that do super fast matrix multiplications and activation functions. Glue that together with a C-API and

I don't see why Intel couldn't make high end PCI cards that do super fast matrix multiplications and activation functions. Glue that together with a C-API and Python bindings. For all the interesting concepts in CUDA, it seems like gamers want pixels and polygons, and neural networks want gemm/relu. It doesn't seem like Nvidia really has a moat that others (Tenstorrent?) can't cross.