Recent transaction-based indices place B200 on-demand rentals near $7.03 per GPU-hour as of mid-September 2026, with neocloud medians around $6.69 and wide dispersion from $3.49 on marketplaces to over $14 at hyperscalers. Persistent AI training and inference demand, especially for 180 GB HBM3e workloads, sustains premiums even a year after broader availability, while HBM3e and CoWoS capacity limits keep supply tight. Spot and interruptible tiers trade materially lower, widening the range traders must weigh. With only days left before month-end, incremental neocloud capacity additions and any last-minute hyperscaler allocations or demand surges from model releases represent the main swing factors that could shift the index within the closely contested $6.50–$7.50 band.