**Current A100 rental rates show wide dispersion, with spot and marketplace on-demand listings clustered between roughly $0.27–$0.90 per GPU-hour and median on-demand nearer $1.5–1.9 across providers as of late September 2026.** Newer Hopper and Blackwell accelerators are absorbing frontier training workloads, releasing older A100 capacity into the on-demand and neo-cloud segments and exerting downward pressure, while established inference and fine-tuning pipelines continue to anchor prices through favorable unit economics. Recent indices reflect modest declines of 5–10% over the prior month amid this substitution effect. Over the next eight weeks, additional Blackwell availability, hyperscaler spot dynamics, and any surge in cost-sensitive inference demand will determine whether rates compress further toward the lower ranges or stabilize near current medians. The evenly matched market-implied odds across buckets underscore the sensitivity to these near-term supply and workload shifts.