**H200 rental prices sit in a tight band around the $3.50–$5.00 per GPU-hour median across marketplaces like Vast.ai, FluidStack, and Jarvislabs as of late July 2026, with on-demand listings commonly ranging $2.50–$6.00 depending on region, commitment length, and whether capacity is spot or reserved.** Strong sustained demand for memory-intensive inference and long-context workloads—driven by large language model serving—supports the floor, while HBM3e supply constraints and 30+ week lead times for H200 NVL systems limit rapid expansion. At the same time, Blackwell B200 ramp-up is introducing higher-memory (192 GB) alternatives at $5–$10+ per hour where available, creating substitution pressure that could soften H200 rates if B200 allocations increase meaningfully before September. Recent cloud-provider price adjustments and modest median increases on peer-to-peer platforms reflect this balance between persistent AI compute hunger and gradual generational displacement. Traders see roughly equal odds across the listed brackets because near-term supply tightness competes with credible downside risks from B200 availability, seasonal demand shifts, and any acceleration in CoWoS/HBM capacity. Key swing factors include B200 shipment volumes, hyperscaler spot discounts, and any major model-training announcements that could spike short-term H200 utilization.