Tight supply amid surging AI training and inference demand remains the core driver behind the dispersed odds on H100 rental prices by year-end 2026. Current medians hover near $3.40–$3.80 per GPU-hour across providers, with spot rates as low as $1.80 and hyperscaler list prices exceeding $7, reflecting bifurcated markets. NVIDIA’s Blackwell ramp and memory constraints have not eased bottlenecks as expected—executives project shortages through early 2028—while sustained large-model development keeps utilization high. Traders weigh neo-cloud competition and efficiency gains against enterprise commitments and hyperscaler pricing power, creating the close contest between $3+ and sub-$3 bands without a dominant consensus.