**Persistent supply constraints and robust AI demand continue to support elevated B200 rental rates into late 2026.** Current on-demand pricing spans roughly $3.50–$16 per GPU-hour across providers, with medians near $6.50–$6.80 and many listings clustered in the $5–$7 range; hyperscalers command premiums while specialist clouds and marketplaces sit lower. Recent signals reinforce tightness, including Nebius’s announced 17–21% increases effective October 1 (lifting B200 to $8.50/hour) and reports of strong year-over-year revenue growth at GPU cloud operators driven by inference and training workloads. Blackwell B200 availability remains allocation-constrained, with hyperscalers absorbing priority output and enterprise lead times still measured in weeks for non-preferred buyers. Production yields have improved from earlier dual-die challenges, yet overall supply has not outpaced demand growth, keeping utilization high and limiting downward pressure. With only three months until year-end and no major capacity wave expected before then, traders see limited scope for sharp declines, positioning $7.00+ as the consensus outcome while lower bands reflect scenarios of faster-than-expected ramp or seasonal softening.