Sustained demand for memory-intensive AI workloads, including large language model training and inference, continues to outpace H200 supply even as Blackwell B200 units reach broader availability at median rates near $7 per GPU-hour. Current on-demand medians for the 141 GB H200 hover around $4–$5 across marketplaces and providers like Vast.ai, RunPod, and CoreWeave, with spot options as low as $2–$3, though several platforms have announced increases effective December 1. NVIDIA’s recent decision to redirect TSMC capacity away from H200 production toward next-generation chips has tightened near-term availability, while token-volume growth since mid-2026 supports elevated pricing. These dynamics underpin trader consensus around the $6.50+ bracket at resolution, with limited downside catalysts visible before year-end.