**Current H200 rental rates cluster in the $4–$6 per GPU-hour range on major platforms, reflecting balanced but uncertain supply-demand dynamics into year-end.** NVIDIA’s ongoing HBM3e memory constraints, flagged by the company as extending through 2028, limit rapid capacity expansion even as production ramps. This keeps premium Hopper GPUs scarce relative to exploding inference and fine-tuning workloads, while newer Blackwell parts command higher pricing and draw some demand away. Spot and marketplace listings have softened modestly with added neocloud and community supply, yet on-demand medians remain anchored near $4.50–$5.80 across providers. Key swing factors over the next weeks include further Blackwell availability, hyperscaler utilization rates, and any last-minute memory allocation shifts that could push realized prices toward either the lower or upper end of the tightly bunched market buckets.