H100 SXM rental prices across providers

Where H100 SXM 80GB on-demand rates sit across providers, read from each seller's page, and what a seller below or above the range should check.

INFRATOAPP · PUBLISHED 25 OCTOBER 2026 · 5 MIN READ · 4 SOURCES · PRICES LAST CHECKED 8H AGO

Where H100 SXM 80GB rates sit

The H100 SXM 80GB has a current on-demand rate at 13 providers we hold. The lowest listed rate is $1.68 per GPU-hour, at Vast.ai, and the highest is $6.16, at CoreWeave. Highest divided by lowest gives 3.7×. Each figure is a seller's cheapest on-demand list rate, read from its own pricing page, and none of it recommends one seller over another.

Run around the clock for a month of 730 hours, the lowest rate comes to $1,226.40 per card. It is rate times hours and nothing more: no storage, no egress, no support. Every seller, rate and source page sits on the H100 SXM page.

This post reads the range from the side of the people selling the card. The names at either end follow the rates and nothing else. A low rate is a fact about a price list, not a verdict on the service behind it.

How a rate gets into the range

Each figure starts on a provider's own pricing page. We take one rate per seller for this card: the cheapest on-demand list rate on that page, quoted per GPU per hour. Reserved terms, interruptible capacity and negotiated discounts stay out of this range, which covers on-demand list rates only.

A rate older than 21 days is hidden rather than shown. The range is therefore made of sellers whose pages were read recently, and a seller whose page could not be read lately drops out until the next good read. The method, including what we refuse to publish, is on how the data is gathered.

That rule changes what the two ends mean. They are the cheapest and dearest current list rates among the sellers we hold, which is a narrower thing than the cheapest and dearest H100 SXM capacity anywhere.

The PCIe card is a different purchase

The H100 also comes as a PCIe card with the same 80 GB of memory. SXM modules sit on a board with a fast GPU-to-GPU link; PCIe cards sit in an ordinary slot. Buyers searching for an H100 meet both, and a price table that blurs them invites the wrong comparison.

On the H100 PCIe page the lowest listed rate is $1.66, at Shadeform, and the highest is $3.50, across 7 sellers. Set beside the SXM figures above, they let a seller see whether its own SXM rate is being read against the right card.

If your page lists an H100 without saying SXM or PCIe, a reader has to guess, and so does anyone reading the page to build a table. The fix costs one word in the product name.

A seller cheaper than the table shows

Some sellers publish an on-demand H100 SXM rate under the low end shown above ($1.68) and still do not appear there. The reasons are likely to be mechanical: the rate may be on a page that has not been read, it may be quoted per node or per instance rather than per GPU, or it may sit behind a sign-in.

Check the unit first. A node price without a plainly stated card count cannot be turned into a per-GPU rate by guesswork, and a per-GPU rate is what the range shows. Check the date next: a page whose last good read is older than 21 days has its rate hidden.

Then check the label. A rate filed under a generic H100 heading may be matched to a different SKU or not matched at all. A claimed listing lets a seller correct a figure we read wrongly, and a dispute can be filed against any value with evidence.

A seller at or past the top of the range

A rate at the top of the range, or past the highest shown ($6.16), is not a mistake by itself. Sellers bundle CPU, memory, local disk, network and support into an hourly price in different amounts, and a bundled rate will look high next to a bare one. The page has to say what the rate includes, so a buyer can see why.

Commitment is the other cause to check. Where contracted customers pay less than the on-demand rate on your page, the table still shows the on-demand rate, since it is the published one. Showing both on the page, each labelled, is how a buyer gets to see the second number.

Put the parts of a bundle beside the rate: vCPUs, system memory, local storage, and whether egress is billed. The rental cost calculator multiplies a listed rate by hours and adds nothing, so anything outside the rate stays outside the sum a buyer sees.

What the H100 SXM range leaves out

The range says nothing about how many hours were sold at any rate. A seller at the low end may be full or empty, and the data cannot tell which. It holds nothing on availability either, so the table cannot separate a rate on cards that are free today from a rate on cards that are not.

It cannot show discounts that are not on a public page, and it does not weight sellers by size. Every seller counts once, at its cheapest on-demand rate, whatever share of the hours it sells.

A seller reading this table for its own pricing should treat it as the view buyers start from, and then check its own page against the points above, in that order: card name, unit, commitment, what the hour includes.

What we could not answer

⚑We hold no data on hours sold at any rate, so the range describes list prices, not the prices most capacity trades at.
⚑Availability is not in a rate and is not measured, so nothing here says who has H100 SXM cards free today.

What this rests on

infratoapp, “H100 SXM rental prices across providers”, read 27 October 2026. lowest listed H100 SXM 80GB rate per GPU-hour: $1.68. https://www.infratoapp.com/blog/h100-sxm-rental-prices-across-providers

Questions people ask about this

What is the H100 SXM rental price per hour?

Across 13 sellers, H100 SXM 80GB on-demand list rates run from $1.68 to $6.16 per GPU-hour. Each figure is a seller's cheapest on-demand list rate, read from its own pricing page, and any rate older than 21 days is hidden.

Is the H100 SXM priced differently from the H100 PCIe?

They are listed as different cards with separate rates. Both carry 80 GB of memory, but the SXM module sits on a board with a fast GPU-to-GPU link and the PCIe card sits in an ordinary slot, so buyers treat them as different purchases.

Why is my H100 rate missing from a price comparison?

The usual reasons are a rate quoted per node or per instance rather than per GPU, a page that could not be read, a rate behind a sign-in, or a last good read older than 21 days. A claimed listing lets a seller correct a figure read wrongly.

These render as FAQPage structured data · answers are filled from the same live figures