How do Serverless GPU Cloud Providers deal with that then? Do they go down the Infiniband-to-NAS rabbit hole to build all of their infrastructure? Or do they just setup an NVME RAID cache to hold the models locally? (Maybe an LRU? With the system memory also being used?)

I imagine in the real world that model usage follows a zipfian distribution, ie, a small number of models (<10) represent 95% of the machines for inference workloads. And, for those machines, you can just load the weights off of your ~40gbit ethernet connection since they're never cycling.

But for that last 5%, I feel like that's where it becomes important. If I'm running a weird, custom model and I want Lambda-like billing... what's the stack? Is the market big enough that people care? (And do most people just use LORAs which are much easier to hot swap?)

Training I imagine is a totally different ballpark because you're constantly checkpointing, transferring data at each step, etc, versus inference. That's a world I know a lot less about though!