Remote AI training matched local-data speed after a short cache warmup
An AWS and Qumulo test kept training data in Ohio while GPUs ran in Oregon with 60 ms network latency. Predictive caching reached 115–116 samples per second without copying the full dataset.