**Current B200 (NVIDIA Blackwell) on-demand rental rates cluster tightly around $6.50–$7.50 per GPU-hour as of August 23, 2026, aligning with the market’s two leading outcomes (37% and 34.7%).** Real-time listings show a broad spread—lowest reliable on-demand offers near $3.49–$3.50 (Prime Intellect, Vultr) and spot/interruptible as low as $3.20, while median on-demand sits at roughly $6.69–$6.79 across neocloud and marketplace providers. Higher-tier or hyperscaler quotes reach $8–$16+, creating the observed price dispersion. Supply remains constrained by lingering HBM3e memory and TSMC CoWoS packaging limits, with a multi-million-unit backlog and lead times of 8–16 weeks for non-priority buyers. This scarcity supports elevated pricing relative to H100/H200 GPUs, even as more providers (14–27 tracked) bring B200 capacity online. Demand stays robust for large-model training and inference workloads that leverage the 180 GB HBM3e VRAM and ~8 TB/s bandwidth, though the newer B300 is beginning to compete on certain FP4-heavy tasks. Key differentiating factors include provider tier and commitment type: neoclouds and marketplaces offer the bulk of mid-$6 to low-$7 rates with greater availability, while hyperscalers command premiums for guaranteed enterprise-grade access. Spot and reserved options pull averages lower, and competitive entry by additional vendors has compressed some quotes over the past year even as the overall median has risen modestly. With end-of-August resolution imminent and no major new supply shocks reported, trader sentiment reflects this steady-state balance between constrained high-end capacity and growing lower-cost options.