Comment by m_mueller

I’d say it mainly needs persistence and good execution (library support). NVIDIA has co-developed CUDA with their hardware, and largely stayed compatible with it, since around 2009, and around 2012 it first started taking off in the HPC space. Years later this enabled first their boom in crypto and then an even bigger one in AI. I don’t think this amount of R&D would be out of reach of today’s AMD (as NVIDIA wasn’t any bigger back then), but the backing of it needs to come from the very top.