Sustained AI inference and training demand, combined with persistent HBM memory shortages and NVIDIA's production shift toward Blackwell, underpins the tight clustering of trader odds around $2–3 per hour for H100 rentals by year-end 2026. Current median on-demand rates hover near $3.30 across providers, with neocloud marketplaces offering lower tiers than hyperscalers, while contract pricing has risen roughly 40% since late 2025 amid sold-out capacity. Blackwell ramp-up has displaced some frontier workloads but has not eased Hopper supply enough to trigger broad declines, as power constraints and booked deployments through September limit new H100 availability. Key swing factors include the pace of GB300 deployments and any softening in open-weight model inference loads before December.