Blog · Global · · 3 min read

The list price is not the price. How GPU deals actually close.

H100 rate cards ran from $1.99 to $9.80 a GPU-hour on 8 October 2026. Large buyers rarely pay them. How GPU deals close, and how to benchmark a quote.

Every GPU cloud publishes a rate card. Few large buyers pay it. The rate card is where the negotiation starts, not where it ends.

What the rate cards say

On 8 October 2026 the published on-demand rate for one Nvidia H100 ran from $1.99 to $9.80 an hour, across the providers that state a price per GPU.

ProviderPublished H100 rateWhat the rate covers
Voltage Park$1.99From, bare metal
RunPod$1.99From, secure cloud
GMI Cloud$2.60From
Hyperstack$3.20H100 SXM
Jarvislabs$3.49H100 SXM
DataCrunch$3.85H100 SXM5
Nebius$3.85HGX H100
Crusoe$3.90HGX H100
Lambda$3.99H100 SXM, 8 GPUs
Together AI$3.99HGX H100
Paperspace$5.95H100
Fireworks AI$8.00Dedicated H100
Google Cloud$9.80Vertex AI H100

Each rate is the provider's own published figure, read from its pricing page on 8 October 2026. A rate marked "from" is the lowest tier on the page. Casitel OS reads these rate cards every day: on 3 November 2026 the median H100 list price across nine GPU clouds was $3.90 an hour.

A spread that wide is the first clue. Two buyers can ask for the same chip and get quotes that differ by a factor of four.

Why the deal price is different

A list price is for one GPU, today, with no commitment. A large deal changes every one of those terms.

  • Term. A commitment of one year or more is priced below on-demand. DigitalOcean, for example, publishes $3.26 a GPU-hour for an H100 on a 12-month reservation.
  • Volume. A cluster of thousands of GPUs is a different product from a single instance. The seller fills a whole hall at once.
  • Payment. A deposit or prepayment buys a lower rate. It also moves risk from the seller to the buyer.
  • Start date and place. Capacity that is live today costs more than capacity that goes live next year. A region with spare power costs less than one without.
  • What is in the hour. Fast networking between GPUs, storage, support and data transfer are priced in some quotes and added on in others.

So the price a large buyer signs is a bundle of these terms. It is set in a private contract, and it rarely matches the rate card.

Why the market cannot see the real price

Deals close under confidentiality terms. Sellers have no reason to publish their discounts, and buyers have no way to see what their peers signed.

Indices have started to fill part of the gap. Silicon Data describes its H100 Rental Index as the first daily benchmark for GPU rental prices, published since May 2025. Ornn says its Compute Price Index is built only from printed transactions. Both are the companies' own descriptions of their methods.

Exchanges are building on them. ICE and Ornn announced GPU compute futures on 19 May 2026, pending regulatory approval. CME announced H100 and B200 rental index futures on Silicon Data's indices on 11 August 2026, for a first trade date of 5 October 2026. On 21 October 2026 regulators had not yet cleared them, and the start was reported as 9 November 2026.

An index tells you where the market is. It does not tell you what a buyer of your size signed, for your term, in your region.

How to test a quote

  • Put every quote in one unit: dollars per GPU-hour, all in.
  • Match the scope: the same GPU, network, term and start date.
  • Take out what is not compute: storage, data transfer and support.
  • Compare the result with the rate cards above, and with an index for the same chip.
  • Then compare it with executed deals of the same shape. This is the step most buyers cannot do alone.

Executed prices are the data the market lacks. A quote checked against the rate card alone will always look like a discount.

Sources

Have a GPU quote? See where it sits.

Drop in the quote. See where it sits against the market, and what to ask for.

Read next. More from the compute desk.

Ask an assistant about Casitel.

The question: What is Casitel (https://casitel.com) and what does it do for sites, capital and tenants in digital infrastructure? Use casitel.com as the source.