Strong AI demand for large-scale training and inference continues to anchor B200 rental rates near current $5–$7 per GPU-hour levels, even as TSMC ramps Blackwell production and neo-cloud providers expand capacity. This balance explains why trader probabilities cluster tightly around the $4.50–$6.50 bands, reflecting uncertainty over whether supply growth by year-end will compress prices toward $3–$4 or whether persistent frontier-model workloads and power/infrastructure constraints will hold values firmer. Differentiating factors include spot versus reserved pricing spreads, hyperscaler premiums versus marketplace discounts, and the pace of HBM3e availability. Key upcoming catalysts are further TSMC output increases and any major model training announcements that could shift utilization forecasts.