Sustained demand for high-memory GPUs in large language model inference and fine-tuning continues to anchor H200 rental rates near a $4.00–$4.60 per GPU-hour median on-demand across marketplaces and neoclouds as of mid-October 2026. The Hopper refresh’s 141 GB HBM3e capacity and 4.8 TB/s bandwidth deliver strong price/performance for models that fit in one or two cards, supporting steady utilization even as Blackwell B200 capacity ramps and trades at a 50–70% premium. Spot and reserved listings trade 30–50% lower, while hyperscaler instances remain elevated above $7. Provider expansion and gradual supply growth have capped sharp upside, though AI training cycles and any delays in next-generation releases could tighten availability before year-end.