**Persistent AI demand for training and inference workloads, combined with constrained H100 supply amid the shift to Blackwell production, keeps competitive on-demand rental rates clustered near $2.50–$3.00 per GPU-hour on neocloud and marketplace platforms as of late September 2026.** Median tracked prices sit around $3.00–$3.36 across dozens of providers, with floors near $1.30–$1.99 on spot or community tiers and premiums well above $6 on hyperscalers. Recent 90-day trends show modest upward pressure (+7% in some indexes) from HBM memory shortages, CoWoS packaging limits, and power constraints, though broader 2025–2026 declines from earlier peaks have stabilized the range. With resolution imminent, trader sentiment splits between the leading $2.50–$2.75 and $2.75–$3.00 buckets due to variability in form factor (PCIe vs. SXM), region, billing model, and short-term capacity fluctuations—factors that could shift the effective benchmark by a few percent before month-end.