Current on-demand rental rates for NVIDIA B200 GPUs, built on the Blackwell architecture with 180 GB HBM3e memory and 8 TB/s bandwidth, range from roughly $5 to $10 per GPU-hour across providers like RunPod, Lambda, CoreWeave, and Vast.ai, with spot and marketplace options often lower at $4–$6. Sustained demand for large language model training and inference has supported these levels despite broader availability since the 2025 ramp-up, as hyperscalers and AI labs compete for capacity. Recent 30-day price increases of 15%+ reflect tight supply relative to workloads, though growing listings from new entrants and multi-GPU clusters introduce downward pressure. Key watchpoints include further production scaling, potential B300 introductions, and seasonal shifts in enterprise commitments that could alter trader-implied odds on 2026 price thresholds.