GPU capacity now lives wherever you can get it: hyperscalers, neoclouds, specialty providers, on-prem clusters. Your training data mostly stayed where it was. Every epoch, fine-tune, experiment sweep, and synthetic data run pays a hidden tax to close that distance, and most AI budgets don’t have a line for it. The tax shows up three ways. The toll: egress, cross-region sync, duplicate copies, and request fees on billions of small files. The wait: GPUs sitting idle because storage was sized for terabytes, not throughput, and on a synchronous training job one slow data loader stalls the whole cluster. The lock: regulation that makes leaving a cloud cheaper, while multi-cloud, in-parallel use can still be charged. Drawing on field work with AI-native companies and GPU cloud providers, this session puts real numbers on each tax, shows how to size storage for GB/s per GPU, and walks through the patterns that keep compute portable: neutral storage, caching next to the GPUs, throughput-friendly formats, and contract terms that hold up at scale. It also covers when the right answer is to move nothing at all.