Current H200 rental medians hover near $4.8–$5.7 per GPU-hour on-demand across neoclouds and marketplaces, with spot rates notably lower. The tightly clustered Polymarket odds reflect uncertainty over whether persistent demand for large language model inference—leveraging the Hopper GPU’s 141 GB HBM3e memory and 4.8 TB/s bandwidth for 70B+ models and long contexts—will hold prices near or above $6 through November, or whether rising H200 supply plus early Blackwell (B200/B300) deployments will cap or ease rates. China export approvals add a wildcard by potentially shifting global allocation. Key near-term catalysts include hyperscaler pricing moves, any frontier model training surges, and Blackwell availability reports that could redirect enterprise workloads.