Nvidia should have a far more well oiled machine in designing/validating new silicon. Software stack? Nvidia already has that. All models are optimized for CUDA - not Grog.

At a very high level, Nvidia just needs to break out the Tensor cores into its own chips, add loads of SRAM, and run CUDA. They already have all the other datacenter stuff figured out such as interconnects.