[antiherovelo]

For a DC operator, GPUs have high runtime costs: think about all amps drawn and heat spewed by a box full of GPUs.

The fraud thing has been a factor, for sure. You need some guarantees that your client is legit before you spend money on them.

Running big clusters of GPUs is a more challenging technical exercise than ordinary Hypervisors/VMs/K8s/whatever. Doubly so if you are doing multi-tenancy and containerization. It takes a lot of effort to rig that stuff up properly and integrate it with the rest of you management stack, and folks tend to be protective of it once it’s fired up.