Traders see closely matched probabilities across the $4–$6 range for H200 rental rates at end-September because persistent HBM3e and CoWoS bottlenecks continue to constrain Hopper supply while Blackwell B200 ramp-up introduces displacement risk. Current marketplace rates cluster around $3.80–$4.60 per GPU-hour on providers such as Jarvislabs and Vast.ai, with dedicated capacity holding firmer premiums amid steady AI inference and training demand. Key swing factors include any acceleration in B200 availability that could free H200 inventory, shifts in Chinese export licensing, or utilization signals from hyperscalers. Near-term catalysts such as quarterly capacity reports or major model training announcements could quickly tilt sentiment toward either modest softening or sustained $4.50–$5.50 levels.