Current on-demand H200 rental rates sit at a median near $4.50–$4.90 per GPU-hour across dozens of providers as of late October 2026, with spot listings as low as $1.33 and hyperscaler instances reaching $7–$10. The H200’s 141 GB HBM3e memory and 4.8 TB/s bandwidth continue to command a modest premium over H100 for memory-intensive LLM training and inference, while neocloud marketplaces keep the floor lower through flexible supply. Prices have edged up roughly 4% over the past 90 days amid sustained AI demand, though expanding Blackwell capacity and reserved-term discounts are capping sharper gains. Key near-term variables include further B200 ramp-up, enterprise budget cycles, and any new large-model releases that could tighten or loosen utilization.