Nvidia H100 Price: Retail, Deals & Price History

The Nvidia H100 is a data center GPU built for training and running large AI models. It is sold mainly to companies and cloud providers, not through a standard consumer price list. Nvidia does not publish a list price for it, so what you pay depends on the seller, the form factor, and how many cards you buy at once.

There are two ways to get H100 capacity: buy the hardware, or rent it by the hour from a cloud provider. Buying figures below are estimates from CloudZero and live listings, checked Oct 12, 2026. Cloud rates are the providers' own published prices, quoted per GPU per hour unless noted.

Single cards and full servers are priced very differently. A single 80GB card is only part of the cost of a working system, because the card needs a compatible server, power, and cooling.

Two things stand out. Cards on the secondary market are listed in a range, which reflects condition, warranty, and how much the seller wants to move stock. A full 8-GPU HGX system costs many times what a single card does, since it bundles eight GPUs with the board that connects them.

These are estimates, not Nvidia prices. Quotes from system builders and resellers usually include configuration, networking, and support, and can differ from what you see on a marketplace.

Renting avoids the upfront purchase and lets you pay only while a job runs. Providers publish hourly rates per GPU, and the rate often depends on how many GPUs you reserve in one instance and which version of the card you choose.

Lambda lists its on-demand SXM 80GB rate at $3.99/hr on an 8-GPU instance, and the same card costs $4.09/hr on a 4-GPU instance and $4.19/hr on a 2-GPU instance. The PCIe version on a single GPU is listed at $3.29/hr. Runpod lists SXM at $3.99/hr and PCIe at $2.89/hr.

The 1-Click Cluster option on Lambda is priced differently because it is a commitment product: $6.16/hr per GPU for a 16-GPU cluster reserved from two weeks up to a year. The per-hour number looks higher than on-demand here, so read what the cluster package includes before comparing.

CoreWeave prices differently again, by the whole node rather than per GPU.

An 8-GPU HGX H100 node is listed at $49.24/hr. To compare it with per-GPU rates, divide by eight and check what each provider includes.

H100 comes in two main forms. SXM cards sit on a specialized board and connect to each other at high speed, which suits large multi-GPU training. PCIe cards plug into a standard slot and are easier to place in ordinary servers. On the rental lists above, PCIe is cheaper than SXM at both Lambda and Runpod. Which one you need depends on whether your workload spreads across many GPUs.

• Buy or rent: Hardware is a large upfront cost; rental spreads cost over time but adds up if the GPU runs constantly.
• Form factor: SXM and PCIe are priced differently.
• GPU count: Instance size changes the hourly rate at some providers.
• Commitment length: Reserved or cluster plans differ from on-demand pricing.
• Condition and warranty: Secondary market cards may have limited or no warranty.
• Supply and demand: AI demand and the release of newer Nvidia GPUs can move prices.

The card is not the whole bill. If you buy, plan for the server, power delivery, cooling, networking, rack space, and someone to maintain it. If you rent, check for charges outside the GPU rate, such as storage, data transfer out, and idle time while an instance is running but not working. Billing granularity also varies, so read how each provider rounds usage.

• Is the quote for the card alone, or for a complete system with support?
• Is the GPU new, refurbished, or used, and what warranty comes with it?
• Is the rate per GPU or per node?
• Is the price on-demand, or does it require a reservation?
• What are the storage and data transfer fees?

This page is general price information, not investment or purchasing advice. For a large purchase, get written quotes from several vendors.