[loudmax]

Beefy VRAM is a start, but Nvidia's real moat is CUDA. If PyTorch runs on AMD's ROCm, or the Intel equivalent, as well as it runs on CUDA, then we'll see some real competition here.

Chris Lattner's Mojo programming language may present an alternative here, but it's still closed source.

[3abiton]

How will Mojo solve what is essentially a GPU driver issue (ROCm vs cuda)? AMD's track at supporting GPU computing is very poor to be trusted so far.