Which, if you notice how AMD plays in the AI space is exactly their plan: cost advantage and being “open” so that the communities take up some of the slack of HIP competing against CUDA. It certainly isn’t perfect but AMD doesn’t have the head count to do anything about it really. I don’t know if there is a better path for them to take, I at least don’t see it.
And remember: all of AMD’s R&D is between CPU, semi-custom, GPU, FPGA, memory control, etc. nVidia technically has a decent chunk of non-GPU-specific (tegra/SoC stuff like for the Switch, or high-speed networking interfaces) but still their GPU teams outnumber all of AMD’s R&D by almost 10x alone.
Intel had a similar scale advantage but they for some reason stagnated for (depending on how you count) 5-10 years which allowed AMD to dig themselves out of their hole. nVidia at least hasn’t been that incompetent, though they have stumbled quite a bit recently on their total inertia.
With what money should they hire more GPU team with?