[nemothekid]

>However it's worth noting that in the end, it's the consumer that loses out when these technical/capability moats maintain a de facto sort of monopoly on certain use cases.

I think in most cases you are right, but for CUDA, the consumer is winning. CUDA isn't some special secret complex algorithm just for nvidia GPUs. nvidia just spent a decade giving a fuck about the developer experience across a wide range of industries. CUDA isn't just AI, it's also physics, numerical modeling, cryptography, biology and more. They found a thousand use cases and spent time listening to customers and built a platform around it - it just turned out AI ended up being a huge money maker.

The problem is, and will continue to be, that Intel and AMD will only see the "AI" money bags and ignore every other part of platform, from debugging, compilers, language integration, GUIs, and even bug fixing. If Intel is saying here "we want to eliminate CUDA by investing billions into OpenCL and ensure OpenCL has a top of line developer experience and platform" than I'd be excited. But what I'm reading is "we will replace a few function calls for CUDA in Pytorch", which might be fun for a while up until you have to debug some performance issue and you realize you can get in touch with a CUDA engineer on github instead of emailing some dead Intel mailing list.

[dopeboy]

Well said. It sounds like NVIDIA went through the hard work of listening and building for years (decades!). I can imagine some top level execs looking at that success and looking to replicate the tech, studying the use cases that customers want.