That's definitely true. It still makes sense to run your own AI/ML servers on-prem if you're saturating some $20k/each GPU boxes, for example.

But ... most web applications are still IO bound, and labour is surprisingly expensive.