My 'back of the napkin' calculations indicate it would take very little time for it to pay for itself by purchasing your own NVIDIA GPU(s) and running them from a colo.
Especially if your requirements are for the instance(s) to run 24x7.
In the many years I've used cloud I've never requested support, so in my experience I don't need to calculate that in.
And even when you consider machine costs, as long as you can get at least 1 yr (not an unreasonable assumption) out of your equipment you're more than making up for what it would've cost in the cloud.
The demand right now has also fallen dramatically even from regular customers.
https://www.tomshardware.com/news/coronavirus-motherboard-gr...
You can even buy Tesla V100s that power AWS accelerated instances.
Over the last decade there was often a major problem in GPU supply. Huge users like Amazon were presumably buying in bulk direct from manufacturers as custom runs so weren’t trying to buy them on the open market like everyone else.
Even other clouds had major problems getting them at some points.
If anything, it generated a lot of cheap secondhand cards for sale in some places.