**Current on-demand H100 rental rates cluster primarily in the $2–$3.50 per GPU-hour range across neoclouds and marketplaces, with medians near $3.00–$3.37 and transacted neocloud indices around $2.60–$2.80 as of mid-September 2026.** This positioning explains the market’s heavy weighting toward the $2.50–$2.75 bucket (47.5%) and adjacent $2.75–$3.25 outcomes (combined ~36%). Key drivers include sustained AI inference and training demand that has supported prices after earlier 2026 increases tied to HBM memory shortages and capacity constraints, offset by growing supply from specialized providers like Voltage Park, RunPod, Vast.ai, and Lambda. These neoclouds and spot marketplaces consistently undercut hyperscalers (often $6–$12/hr) by 50% or more, compressing premiums and anchoring rates in the observed band. Blackwell ramp-up adds some competitive pressure but has not yet displaced H100 demand significantly. With only two weeks until end-September, volatility is limited; recent modest upward drift (several percent over 90 days in trackers) and stable availability suggest the $2.50–$3.25 cluster will likely hold unless a sudden supply surge or demand spike emerges. Spot and interruptible tiers remain cheaper but carry availability risk, while reserved contracts trade at modest discounts. Traders appear to view the current equilibrium—driven by neocloud competition amid persistent AI workloads—as the most probable near-term outcome.