nvidia is unique because of cuda and all ml software written/optimized for it. AMD can add cuda support too (and they started working in this direction) but they'll not be able to out-compete nvidia, with this strategy the'll stay at 2'nd place (by buying amd cards for inference you are by default less competitive as business doing ML compared to those that can buy nvidia) and will just strengthen nvidia's cuda moat. Imo it's less about money and more about a ton software that was optimized for nvidia cards and software that was built on top of that optimized software, it can't be easily ported to run as efficient on AMD, it's tons of code and it's unlikely devs would just port it + port all the new stuff that is researched at exponential scale.