Earlier quoted context omitted.
Actually self-hosting without using the cloud at all involves building a multi-gpu rig. NIMs are just to try out the concept. Honestly, doing it this way is still a bit more secure than using OpenAI or Anthropic directly because its a slight guard against your code being used in future training sets for the next iteration of their AI.
And that's bad because? You want to use a product without improving it?
There's two categories of folks that won't want to send their data out to "improve the product"
The first is the companies that are concerned about their IP (but comfortable enough using a cloud provider).
Then there's a smaller group of folks who don't want to train the next iteration of the product.