Back of the napkin math:

Googling the TFlops as an estimate of power shows a roughly 12x improvement on the H100 over the 4090. A single 4090 takes 160 hours so a single H100 should take about 13 hours.

AWS will rent a p5.48xlarge instance of 8xH100 for $31.464/hr. That will take roughly two hours and cost around $60 bucks.

Assume I'm off by an order of magnitude, this is still a reasonable cost to recover key infrastructure. If the $60/endpoint stands then it would be reasonable to recover workstations this way