H100 Rental Prices Rise by 40%, Splitting the GPU Market

NVIDIA's ($NVDA) one-year rental contract price for the H100 has reportedly increased by about 40% from October 2025 to March 2026. As demand for AI model training and inference grows, the GPU rental market is shifting from a single price structure to one that varies based on contract duration, supply channels, and customer segments.

SemiAnalysis reported on April 2 that the one-year contract price for the H100 rose from $1.70 per hour (approximately 2400 KRW) in October 2025 to $2.35 (approximately 3330 KRW) in March 2026. The report described the increase as 'almost 40%'. The 50% increase mentioned in some headlines was not corroborated by accessible primary data and cross-reports.

The same report explained that on-demand GPU rental capacity is sold out across all GPU types. The H100 and H200 eight-node clusters are also mentioned as being in short supply. Short-term users find it difficult to locate available capacity, while long-term contract holders have a greater incentive to retain existing supplies.

Price indicators do not converge into a single figure. Silicon Data aggregates H100 rental prices separately for NeoCloud and hyperscalers. The current price on the H100 page is listed as $2.53 per hour (approximately 3580 KRW).

The upper index screen of the same platform presented H100 prices as $2.77 (approximately 3920 KRW) for NeoCloud and $7.18 (approximately 10160 KRW) for hyperscalers as of July 30. NeoCloud refers to a group of emerging cloud providers that specialize in renting GPU infrastructure, while hyperscalers refer to large cloud providers.

Another public index, gpurentprice.com, reported the H100 price as $4.57 (approximately 6470 KRW) as of July 29. However, the three indicators differ in contract duration, provider groups, sample composition, and regions, making it difficult to compare absolute prices one-to-one. Even for the same H100, prices vary depending on whether it is on-demand or a one-year contract, and whether it is from NeoCloud or a large cloud provider.

Silicon Data stated in a blog post on August 4 that the NeoCloud price for the H100 rose by 7.0% over 12 weeks, while the H200 increased by 14.4%. The price difference between the H100 and H200 was reported to have widened from 5% to 0.4% and then to 16.2%. This suggests that the demand for AI compute is spreading across multiple GPU generations rather than being limited to a shortage of a specific chip.

There are counterarguments as well. Industry writings have claimed that the median price for on-demand H100 rentals has not significantly changed since March, remaining around $2.73 as of June 2026. The simultaneous observation of a sharp rise in contract prices and stagnation in spot median prices indicates that market conditions should be interpreted with a nuanced understanding of the indicators.

GPU-hour is a cost unit for using one GPU for one hour. AI companies bear this cost directly when training large language models or operating inference services, or it is reflected in cloud fees. When rental prices rise, companies that do not purchase GPUs directly are also exposed to fluctuations in service costs.

This trend is also relevant to domestic readers as it relates to the cost structure of AI services. Companies operating model training or inference services based on the cloud have different cost structures depending on how they secure GPUs. We previously reported that long-term rental prices for the H100 have risen by about 40% since October 2025.

At that time, the investigation also revealed that the price range for the H100 varied significantly by provider. An analysis comparing 38 clouds and 32 GPU products indicated that the price range for the H100 differed by as much as 18 times. This suggests that the GPU rental market is closer to a fragmented price structure than a standardized single market.

Price fluctuations have also led to discussions about financial products. Barron's reported that the CME Group is preparing to launch compute futures aimed to start on October 5, 2026, with Silicon Data mentioned as the reference data.

We also previously reported that the CME is pushing for the launch of GPU rental-linked futures. The GPU rental market has a structure where price differences arise based on providers, regions, contract durations, and availability. Once a public benchmark price is established, users and suppliers can manage rental cost fluctuations separately.

In the cryptocurrency market, there is a connection with decentralized compute projects. The rise in GPU rental prices can be linked to the cost structure of networks providing or utilizing compute resources. However, this does not necessarily indicate a rise in specific token prices or increased demand for projects.

The crux of this issue lies not in how much the H100 rental price has risen at once, but in how the market is being segmented. Public indicators show that AI compute is solidifying into a market where different prices are formed based on contract duration, supply channels, and customer segments. The CME Group's compute futures are being prepared with a target start date of October 5, 2026.