Capacity, interconnect, and committed-use discounts matter more than the logo. Spot GPUs are cheap until they vanish mid-run; reserved clusters are the opposite. Power and networking, not just chip count, decide whether a 'cloud GPU' can actually train at scale.