Sustained AI training and inference demand continues to anchor H200 rental rates near $4.50–$5.30 per GPU-hour on-demand across major providers as of late October 2026, with spot rates often $2–$3.50 lower amid tight HBM3e memory supply. NVIDIA’s ongoing production constraints, flagged through 2028, and hyperscaler capex surges have offset some relief from Blackwell B200/B300 deployments, which command 30–60% premiums due to higher bandwidth and FP4 performance. Recent indices show H200 medians rising modestly year-over-year while Blackwell availability expands, creating substitution pressure for memory-bandwidth-bound workloads. Power grid bottlenecks and data-center build delays further limit effective supply, supporting trader focus on whether H200 utilization stays elevated through year-end earnings and next-gen rack rollouts.