Current H100 on-demand rental rates cluster around $3.30–$3.60 per GPU-hour across specialist clouds and hyperscalers, with recent modest gains reversing 2025 softness amid sustained AI training and inference demand. Trader sentiment remains split because new Blackwell capacity (B200/B300) is ramping unevenly while Hopper production has largely ceased, creating uncertainty over whether supply will ease or tighten further by December. Hyperscaler premiums, spot-market volatility, and forward curves showing potential near-term firmness versus later softening add to the dispersion across price bins. Key swing factors include Blackwell yield and deployment speed, hyperscale capex pacing, and any major open-weight model releases that could sustain or shift utilization.