On a related-different axis, I've consistently seen on-prem GPUs running identical workloads ~35% faster than the same workloads on the same cloud hardware, regardless of intermediate infra stack layering/versioning choices. Weird but I'm not complaining!