AWS and Qumulo have demonstrated a training setup in which GPUs use a dataset stored in another cloud region without permanently replicating it. In their test, SageMaker HyperPod compute ran in Oregon while the source data remained in Ohio across a link with 60 milliseconds of latency.
A local Qumulo storage instance exposed the same file path to the training cluster and fetched blocks from the remote hub over private network peering. Its NeuralCache component predicted the next 4 KB blocks requested by the data loader and staged them on local NVMe storage. After 100 to 150 batches, the remote run reached 115–116 samples per second, compared with 116–117 for compute next to the data.
The test used a 1.02-billion-parameter Llama model on two p5.48xlarge instances, totaling 16 Nvidia H100 GPUs per cluster. Initial GPU use was 80–90 percent and then converged to 98–100 percent.
The result applies to this validated access pattern and predictable training workload, not every cross-region job. Teams still pay network-transfer and storage costs and depend on the remote link during cache misses. The approach is most useful when scarce GPU capacity is available away from a very large dataset that would be slow or expensive to duplicate.