Spot pricing on H100 and Blackwell-class instances has dropped roughly 30% across the major clouds since May. None of the providers announced it — it shows up only in the hourly rate cards and the spot-market history.
Three things are driving it: hyperscalers front-loading Blackwell capacity ahead of demand, a wave of smaller neoclouds (RunPod, Lambda, Vast.ai) undercutting on spot inventory to build utilization, and enterprise reserved-instance commitments freeing up on-demand capacity as more workloads move to committed contracts.
The practical takeaway: if you've been avoiding spot instances for training runs because of interruption risk, it's worth re-checking the math now — the discount has widened enough that even a checkpoint-and-resume setup with occasional restarts often comes out cheaper than a reserved instance.