Why give any additional money to Nvidia when you can announce more profits (or get more compute if you're a government agency) by hiring more engineers to enable AMD hardware for less than a few million per year? It's not like Microsoft loves the idea of handing over money to Nvidia if there is a cheaper alternative that can make $MSFT go up.
Say your success rate for replicating CUDA+Nvidia hardware on AMD is 60%. But it will take 2 years. That's not going to be compelling for any large org, especially when the MI300x is cheaper, but not crazy cheaper than an h100.
Especially since CUDA is still rolling out new functionality and optimizations, so the goal posts will keep moving.
> Say your success rate for replicating CUDA+Nvidia[...]
Rational hyperscalers would just stop as soon as their tooling/workloads/models are functional on AMD hardware within an acceptable perf envelope - just like they already do with their custom silicon. Replicating CUDA is just unnecessary, expensive and time-consuming completionism; if some workloads require CUDA, they will be executed on Nvidia clusters that are part of the fleet.
It depends on how cheaper the total solution is and how available the hardware is. If I can't get Nvidia hardware less than six months after I get AMD hardware, I have a couple months to port my software to AMD and still beat my competitor that's waiting for Nvidia. It's always a matter of how many problems can you solve for a given amount of money x time.
Sure, but the "it depends" is carrying a lot of weight. NVIDIA's moat will get you testable software straight out the gate; any other stack currently is a game of "how long can we take to get this going".
Corporations simply arn't interested in long term gains unless there's a straightforward path.
It depends on the problems you have. If you need CUDA, then you married yourself to Nvidia. If you can use libraries that work equally well on both, then you would benefit.
When you are a government agency, it’s more palatable to spend the budget in a way it results in employment of nationals and development of indigenous technologies.
Because if you don't join NVIDA your likely hood of success goes down. So the "more profits" you speak of is gambling money. Most corporations arn't going to gamble.
