**Sustained AI training and inference demand continues to support H100 rental rates near recent medians of $3.39–$3.57 per GPU-hour across tracked providers, while expanding neocloud and marketplace capacity—offering on-demand floors as low as $1.99 and spot rates below $1—creates downward pressure.** Hyperscalers maintain premiums at $6.88–$12.29 due to bundled networking and reliability, widening the spread versus specialist clouds. Blackwell B200 and H200 ramp-up is absorbing some high-end demand, easing supply constraints on Hopper hardware and contributing to modest recent gains of about 6% over four weeks alongside a longer-term 13% annual rise. With end-of-October resolution imminent, trader consensus around the $2.75–$3.00 band reflects expectations of stable effective pricing amid these competitive dynamics rather than sharp moves.