**Strong sustained demand for AI training and inference compute, combined with ongoing supply constraints and recent provider price hikes, underpins the market's heavy weighting toward B200 rental rates at or above $7.00 per GPU-hour by end-2026.** As of mid-to-late September 2026, on-demand B200 rates span roughly $3.50–$16.11 across providers, with medians near $6.69–$6.79 and spot/interruptible tiers often $3.20–$4.95. Hyperscalers sit at the high end while neoclouds and marketplaces offer lower figures. Recent moves include Nebius raising B200 on-demand pricing ~19% to $8.50 effective October 1, alongside earlier reports of near-doubling quotes tied to allocation tightness. Blackwell production (launched Q1 2025) remains constrained by TSMC CoWoS capacity and hyperscaler absorption, with lead times for new capacity stretching months and most 2025–mid-2026 output already committed. These dynamics—persistent inference and training demand, limited near-term supply additions, and upward price momentum—support trader consensus around elevated year-end levels. The market-implied odds reflect skin-in-the-game assessment of these short-term pressures persisting through December, while acknowledging that any acceleration in new deployments or demand softening could moderate outcomes.