A Correction in Compute: Tracing the Decline in H100 Rental Prices (medium.com)

Silicon Data’s new H100 Rental Index (SDH100RT), launched in May 2025, shows a material correction in GPU rental costs: the index fell to $2.36/hour on June 19, 2025, down 23% from a $3.06 peak on Sept. 1, 2024. That decline reflects easing H100 supply constraints as hyperscalers and regional data centers add capacity, a shift in buyer behavior (inference workloads now outpacing training, prompting more cost- and carbon-aware procurement), and intensifying competitive pressure from GPU-hosting providers and resale/marketplace models that have driven aggressive price cuts and volatility. For the AI/ML community this signals a normalization of the compute market and a need to move beyond raw $/hour quotes toward normalized, comparable metrics that capture real-world performance. Silicon Data stresses that the same H100 can deliver very different effective performance depending on node architecture, cooling, and interconnect—hence its companion benchmark, SiliconMark, which validates cost against compute quality. Looking ahead, the incoming Blackwell generation is expected to reshape demand and pricing dynamics, and Silicon Data plans an A100 Rental Index to extend transparent, historical and real-time pricing insight across configurations and geographies.