For enthusiasts, even renting from Runpod, Salad, Vast.ai which are an order of magnitude cheaper than established cloud providers is much more expensive than buying a Rtx 3090 or 4090. Which got me thinking why don't companies who need training don't put their money together, buy some H100 and share them?
Because that's a lot of capex to take on, not to mention accounting hassles, depreciation, and install/maintenance expenses. Your average bunch of startups does not want to tie up a giant chunk of money in hardware, and even most established companies don't want to either, hence the giant "move to the cloud" epidemic of the last years. On top of that, you need some serious expertise to properly wire up an AI training cluster so you can actually use the GPUs to their full potential, and such experts are rare.
The problem is that MBA beancounters don't have a way to put a number behind "what if the cloud provider suddenly decides to terminate us", "what if creds leak and someone mines shitcoins for a week or two", or (for everyone outside FVEY) "what if the NSA decides to compel the cloud provider to turn over our server data containing our trade secrets", and so this gets all ignored.
CSP's are taking on that KYC risk too.
OCI is a hyperscaler/established provider and offers GPUs for quite cheap compared to the others. They are competitive with Runpod, Vast, and Salad, especially if you negotiate (which everyone who knows this business does).