**Strong trader consensus on B200 on-demand rental rates remaining at or above $7 per GPU-hour by end-2026 stems from sustained supply constraints and robust AI demand.** Current medians sit near $6.25–$6.84 across providers, with spot and marketplace floors as low as $3.20–$3.75, yet forward-booked capacity, TSMC CoWoS-L packaging limits, and HBM3e shortages continue to restrict new supply through late 2026. Hyperscalers and neoclouds have locked in most Blackwell output, while inference workloads (large language models and multi-agent systems) drive token consumption higher than training alone. Recent deployments, such as QumulusAI’s completed B200 handoff in September 2026, illustrate that available capacity clears quickly rather than flooding the market. Prices have remained flat over the past 90 days amid these bottlenecks, with no major relief expected before year-end. Key upcoming catalysts include additional GB300/B300 ramps and any narrowing of the CoWoS gap (projected toward 10% by December), though power constraints and hyperscaler pre-commitments limit downside risk for rates. Market-implied odds above $7 therefore reflect skin-in-the-game expectations that demand growth will continue to outpace incremental supply additions in the final months of 2026.