Persistent supply constraints on high-bandwidth memory and advanced packaging, combined with robust AI training and inference demand, currently support H100 on-demand rental medians near $3.30–$3.50 per GPU-hour across providers. Blackwell GPUs have not displaced Hopper capacity as quickly as anticipated due to their own allocation bottlenecks and slower-than-expected volume ramps, limiting downward pressure on legacy hardware. Neocloud competition and spot-market depth introduce variability, yet hyperscaler premiums and recent provider price hikes reflect sustained scarcity. Over the next quarter, the pace of new data-center capacity, HBM production increases, and enterprise budget cycles will determine whether rates compress toward the lower bands or hold in the $2.50–$3.50 range. Trader consensus across closely matched outcomes underscores uncertainty around these supply-demand inflection points.