[roenxi]

I actually find that a really interesting question with a really interesting answer - the scaling properties of large groups of people are unintuitive. In this case, my guess would be high market complexity, and the entire userbase to ignore that complexity in favour of 1-2 vendors with simple and cheap options. So the market overall will just settle on de-facto standards.

Of course, based on what we see right now that standard would be Nvidia's CUDA; but while CUDA is impressive I don't think running neural nets requires that level of complexity. We're not talking about GUIs which are one of the stickiest and most complicated blocks of software we know about, or complex platform-specific operations. I'd expect that the need for specialist libraries to do inference to go away in time and CUDA to be mainly useful for researching GPU applications to new problems. Training will likely just come down to raw ops/second in hardware rather than software.

It isn't like this stuff can't already run on other cards. AMD cards can run stable diffusion or LLMs. The issue is just that AMD drivers tend to crash. That is simultaneously a huge and a tiny problem - if they focus on it it won't be around for long. CUDA is an advantage, but not a moat.