Recent modest upward pressure on H100 on-demand rates, with medians near $3.32 per GPU-hour across dozens of providers as of mid-September 2026, anchors trader expectations around the $2.75–$3.25 range. Sustained large-language-model training and inference workloads continue to support demand even as newer Hopper and Blackwell variants enter the mix, while neocloud platforms such as RunPod, Vast.ai, and Lium undercut hyperscalers by 50–100 percent through aggressive capacity additions. Spot and reserved pricing remain softer near $2.20–$2.50, creating resolution uncertainty depending on whether the market references blended on-demand listings or lower-tier availability. Short-term volatility from regional supply shifts and contract roll-offs keeps the three leading buckets closely matched.