H100 one-year rates rise 38% to $2.35

SemiAnalysis published H100 cloud pricing data in March 2026 showing one-year contract rates at $2.35 per GPU-hour, up from a $1.70 range in October 2025. The market tracker also marked H100 on-demand capacity as sold out in February and March 2026, according to its public dashboard. The move was cited in social media posts on May 17, 2026, which pointed to tighter supply for Nvidia’s Hopper-generation chips even as newer Blackwell systems were entering the market. SemiAnalysis describes the data as part of its AI cloud total cost of ownership and GPU rental pricing products. ### How much did one-year H100 pricing actually move? SemiAnalysis listed H100 one-year pricing at $1.45-$1.95 per GPU-hour in October 2025 and $1.80-$2.35 in February 2026 before moving to $2.00-$2.70 in March 2026. Using the top end of those quoted one-year ranges, the March figure of $2.35 versus the October figure of $1.70 amounts to about a 38% increase. The dashboard says pricing reflects 25th-75th percentile rental ranges and assumes 25% prepayment for terms of three months and longer. (api.semianalysis.com) ### What does the dashboard show about availability? The SemiAnalysis dashboard showed H100 on-demand availability as “Sold Out” in February 2026 and again in March 2026. The same page listed H100 spot pricing at $2.83 in February and $2.82 in March, while the selected index sat at $2.75 for March 2026. The public preview does not name individual cloud providers tied to those sold-out readings, but it frames the figures as market pricing compiled from its broader AI cloud pricing model. (api.semianalysis.com) ### Which H100 market is this measuring? SemiAnalysis says the dashboard tracks H100 daily index data and neocloud contract pricing across spot, on-demand, one-month, three-month, six-month, one-year and longer-duration terms. That means the $2.35 figure refers to contracted rental pricing in cloud markets rather than Nvidia’s list price for physical chips or a single hyperscaler’s posted on-demand rate. (api.semianalysis.com) SemiAnalysis separately says its AI cloud total cost of ownership model incorporates historical and future rental price analysis and install-base projections through 2028. ### How does that compare with public cloud list prices? Amazon Web Services says its P5 instances are powered by Nvidia H100 Tensor Core GPUs, while Google Cloud’s GPU pricing page lists H100 offerings under Compute Engine and Microsoft documents H100-based NCads H100 v5 virtual machines. Third-party price trackers crawled on May 18, 2026 showed materially higher public-cloud equivalents in many cases, including about $6.98 per GPU-hour for Azure’s Standard_NC40ads_H100_v5 and roughly $11.06 for Google Cloud A3-high single-GPU configurations, while AWS P5 pricing is sold as an eight-GPU instance rather than a single card. (api.semianalysis.com) Those figures are not directly comparable to SemiAnalysis contract rates because they bundle different instance shapes, regions and service terms, but they show the gap between posted hyperscaler prices and brokered contract markets. ### Why are H100 contracts still being watched when Blackwell is shipping? Nvidia’s H100 remains a standard training and inference part in cloud fleets, and Nvidia’s product page still markets the chip for enterprise AI and high-performance computing workloads. SemiAnalysis’ public materials show it continues to track H100 alongside H200, B200, GB200 and GB300 products, suggesting Hopper remains an active commercial market rather than a legacy line that has disappeared from pricing screens. (aws.amazon.com) A March 2026 market write-up citing SemiAnalysis data said H100 rental prices were rising even as buyers expected newer systems to ease pressure on older hardware. ### What should readers watch next in this market? March 2026 is the latest month shown on SemiAnalysis’ public H100 contract dashboard, with contract terms extending from one month to five years. The next concrete signal will be a fresh dashboard update showing whether H100 on-demand capacity remains sold out and whether one-year pricing stays near the March range of $2.00-$2.70 per GPU-hour. (nvidia.com) SemiAnalysis says the full data product is updated daily. (api.semianalysis.com)