What a month of GPU time costs

A month of GPU time is the hourly rate times 730 hours if the machine never stops. Most bills are smaller or larger than that for reasons you can count.

INFRATOAPP · PUBLISHED 26 OCTOBER 2026 · 5 MIN READ · 4 SOURCES · PRICES LAST CHECKED 18H AGO

The month is 730 hours

A month of one GPU running without a break costs the hourly rate times 730, the average number of hours in a month. For the H100 SXM 80GB, 730 hours at the cheapest on-demand rate on record is shown as $1,226.40 per card. Multiply by the number of cards, then adjust for the hours the machine is really on and for the storage and data transfer the rate leaves out.

The figure is per card and on-demand. A node with several cards costs that many times as much, and a reserved or interruptible rate would give a different month altogether.

Treat it as a starting sum. The useful version uses your own card count, your own hours and the rate of the seller you will actually choose, and the rest of this post is about getting each of those right.

The month for an H100, an H200 and an RTX 4090

Always-on months at the cheapest rate on record, for cards at different memory sizes. The H100 SXM 80GB: $1,226.40 (hourly: $1.68). The H200 SXM 141GB: $1,525.70 (hourly: $2.09). The RTX 4090 24GB: $87.60 (hourly: $0.12). Each figure uses the cheapest seller on record for that card, and every other seller charges the same or more.

The cards are not interchangeable. The H200 has 141 GB of memory to the H100's 80 GB, and the RTX 4090 has 24 GB, so a model that fits on one may not fit on another. Work out which card your job needs first, then look at its month. The GPU table has every card with a rate and every seller.

Note also the spread inside each card. The dearest H100 SXM rate on record is $6.16 an hour; multiply it by 730 for the other end of the month. The lowest rate is a floor to check against, not an expected price.

Always-on versus part-time

The 730 hour figure assumes the instance never stops. A development box used eight hours a working day is on for roughly 170 to 180 hours a month, under a quarter of the always-on figure, if you actually stop it each evening. A training run of a known length costs the rate times the hours of the run, plus the time spent setting up and copying data.

Serving is where always-on is real. An endpoint that must answer at three in the morning runs 730 hours whether or not anyone calls it, so its cost is the full month per replica. That is the figure to compare with a token-priced API for the same traffic.

Write down which of these your work is before you multiply. A part-time job left running all month is billed as an always-on one.

Doing the arithmetic for your own case

The rental cost calculator takes a card, a number of GPUs and a number of hours and multiplies them by each seller's published rate. Set the hours to what you expect, not to 730, unless the machine really will run all month.

Then add what it leaves out. Persistent storage is billed per gigabyte per month and keeps billing when the instance is stopped. Moving data out of the provider may be charged. Neither is in the rate. Ask the provider for the price of a volume per gigabyte per month and for any charge on data you move out, then estimate your own sizes and add them to the month.

Keep the result as a range rather than one figure: the hours if everything goes to plan, and the hours if the run fails once and has to restart.

A node is not a card

Training and serving often run on more than one card. Where a provider rents only whole nodes, the month is the per-card figure times the cards in the node, whether or not your job uses them all. A model that fits on half a node still pays for the whole one if the half is not for sale.

Check the smallest unit each provider sells before you work out a month. Some rent a single card, some rent fixed bundles, some rent only full machines. The per-card rate on a card page is a fair comparison of price, but the smallest unit decides what you can actually buy.

For serving, also check how many replicas the traffic needs at its busiest and at its quietest. An always-on floor of replicas is 730 hours each; replicas added for busy hours cost only those hours if they are stopped again.

Where the month changes under you

Rates move. The figures above are read from provider pages on a schedule and any older than 21 days are hidden, so the month you see today may not be the month you pay. Recheck the rate on the day you launch and note the date beside it.

Billing increments change short months too. A job that runs many short sessions on a provider that rounds each up to the hour costs more than the hours suggest. For a full month of continuous use the increment hardly matters; for a month of stop-start testing it can matter a lot.

Currency can move it as well. When the bill is in a currency other than your budget's, the month you calculated and the month you pay can differ for reasons that have nothing to do with the provider. Budget with some room for that, or ask whether billing in your own currency is possible.

What we could not answer

⚑The monthly figures use the cheapest on-demand rate on record only and leave out storage, data transfer and reserved discounts.
⚑We cannot say how many hours a typical renter keeps an instance running, so the part-time example is arithmetic, not an observed pattern.

What this rests on

infratoapp, “What a month of GPU time costs”, read 27 October 2026. lowest listed H100 SXM rate for a 730 hour month: $1,226.40. https://www.infratoapp.com/blog/what-a-month-of-gpu-time-costs

Questions people ask about this

How much does it cost to rent an H100 for a month?

Multiply the hourly rate by 730, the hours in an average month, if it runs without stopping. The H100 SXM 80GB card page lists each seller's hourly rate to multiply. Storage, data transfer and any extra cards are on top, and part-time use costs less.

Why use 730 hours for a month?

A year has 8,760 hours and twelve months, so the average month is 730 hours. Months of 28, 30 and 31 days differ slightly, but 730 is the figure providers and calculators commonly use for monthly estimates, and it keeps comparisons consistent.

Is it cheaper to rent a GPU part-time?

Yes, if you stop the instance when you are not using it, because on-demand billing counts the hours the instance is running. A stopped instance may still bill for attached storage. A machine that is left running overnight costs the same as one doing useful work.

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