Strong AI demand for large language model training and inference continues to anchor H100 rental rates near a $3.00–$3.50 per GPU-hour median on-demand across specialist providers as of mid-October 2026, with hyperscalers listing substantially higher. Recent price increases at Nebius and AWS Capacity Blocks reflect sustained utilization, while Blackwell B200 rollout and expanded neo-cloud capacity introduce downward pressure on spot and reserved H100 availability. Marketplace floors near $1.30–$2.00 highlight segmentation between premium on-demand and interruptible supply. Forward curves on prediction platforms show modest easing into late 2026, driven by new cluster deployments and workload migration to newer accelerators. Key near-term catalysts include hyperscaler capacity block renewals and any acceleration in AMD MI300X or custom silicon adoption.