Most hardware loses value the moment it leaves the box. Nvidia’s H100 is not following that script.
The H100 is worth about 30% more than it was a year ago, fueled by heavy demand for AI compute. For a data center chip that debuted in Q4 2022, that is a strange kind of aging.
The rebound arrives after a rough 2025 for H100 pricing, when rental prices slumped. In 2026, the trend flipped. Demand for both AI training, which builds models, and AI inference, which runs them for users, has pulled rates back up.
What the rental market is saying
As of mid-October 2026, median H100 rental rates sit between $3.3 and $3.6 per GPU-hour. The median on-demand rate stood at $3.49 per GPU-hour as of October 16, 2026.
The year-over-year picture depends on which yardstick you use. One median shows rental prices up 79% from the lows recorded in 2025. Other indices show gains of 13% or more. Those figures line up with a reported average increase of around 30%, amid a supply chain that remains constrained.
Who pays more, and why it matters
Not every renter pays the same price. Hyperscale cloud providers charge premiums 1.7x to 2.7x higher than specialist cloud providers.
The secondary market tells a calmer but consistent story. Used or new-condition H100 units are trading between $25,000 and $40,000, depending on configuration and condition.
Why a 2022 chip still has pull
The H100 is built on Nvidia’s Hopper architecture and launched in Q4 2022 as the company’s flagship data center product. The chip carries 80GB of HBM3 memory, delivers up to 989 TFLOPS of FP16 tensor performance, and draws a 700W TDP.
Newer chips exist, but they have not yet arrived in volumes large enough to push the H100 aside.
What this means for buyers, Nvidia, and the AI trade
Recent price hikes from service providers, combined with persistent interest in AI applications, point to pressure on compute costs for end users. Organizations investing in AI infrastructure may face higher operating costs in the near term. Teams that budgeted around 2025’s cheaper rental rates could find their assumptions outdated.
For GPU owners and rental providers, a rebound in rental prices improves the value proposition of holding GPU capacity, especially when the hardware itself is trading at $25,000 to $40,000 on the secondary market.
The variation across indices matters. A 13% gain and a 79% gain describe very different markets. Anyone making decisions on H100 pricing should look at which benchmark they are using and what period it compares against.
The gap between hyperscaler and specialist pricing is another variable. A 1.7x to 2.7x premium gives cost-conscious buyers a reason to shop around.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.