Personal anecdote on ROI - I was at an early stage startup earlier this year where we had some burstable long-running GPU tasks (<100 VMs). Accross GCP and OCI we couldn't get our hands on L40S on-demand, and had to resort to T4s (released 2018). Sometimes even these were unavailable, and we would have a P4 (2016!) fallback. AWS sells A100s (2020) at $4/hr except they don't even have capacity for x1 versions, you have to rent x8.

AWS runs a hell of a lot of old junk. I was surprised at how we managed to save a lot of money not using it as well.