Ongoing AI demand for large language model training and inference continues to outpace H200 supply, keeping median on-demand rental rates near $4–$4.70 per GPU-hour across providers as of late October 2026. Traders see roughly even odds between stabilization in the $4.50–$5.00 range and further upside above $7.50, reflecting uncertainty over Blackwell B200 availability, which commands $8–$10 but offers higher performance, alongside AMD MI300X competition at lower rates. Recent provider price hikes, rising token consumption, and limited near-term capacity additions from NVIDIA’s roadmap sustain upward pressure, while spot markets and reserved deals provide downside buffers. Resolution hinges on December demand trends and any last-minute supply shifts.