**Sustained demand for AI training and inference has driven H100 rental rates higher in 2026 after 2025 declines.** Median on-demand prices currently sit between roughly $3.30 and $4.20 per GPU-hour across specialist clouds and marketplaces, with recent weekly gains around 9% and year-over-year increases of 13-21% as hyperscalers and neoclouds absorb new supply through long-term contracts. Power constraints, high-bandwidth memory allocation, and forward purchases of Blackwell capacity have kept H100 availability tighter than anticipated, supporting firmer pricing even as newer accelerators enter the mix. Spot and reserved rates remain lower, while hyperscalers continue charging substantial premiums. Key near-term catalysts include further Blackwell ramp timing, major model training runs, and any shifts in data-center energy or export policies that could alter effective supply.