I think it's more likely that the model gets pulled from a SAN into NVIDIA pods, and agents/harnesses would run on a separate random Xeon box or something on the same subnet, using the pod through OAI v1 API. That's easier to maintain overall.