GPU rental: how renting compute actually works

Somewhere between "my laptop can't do this" and "I should buy a $1,600 graphics card" there is a third option that most people never seriously consider, mostly because nobody explains it in plain language. You can rent the card. By the hour, by the minute, for as long as the job takes and not a second longer.

This guide is the whole thing end to end. What you are actually renting, what the bill looks like, what happens when it goes wrong, and the point at which you should stop renting and go buy the hardware. No assumed knowledge, and no pretending the annoying parts are not annoying.

What you are actually renting

You are renting a computer with a graphics card in it. Not a slice of a card, not a subscription to a service, not credits. An actual machine, sitting in an actual building, that exists for as long as you are using it and then stops existing.

A note on the words, because this corner of the industry is careless with them and nobody stops to explain. A GPU is a graphics card: a physical board about the size of a paperback that slots into a computer. People who work with them every day just call it the card, and so does the rest of this guide. The computer it slots into — in a data center rather than under your desk — is a GPU server. When a provider advertises a dedicated GPU server, they mean that whole machine, with the card in it, reserved for you alone rather than shared with three strangers. What you rent from us is a dedicated machine, for as long as your job runs and no longer.

That distinction matters because it explains everything else. It is why you can install whatever you like on it. It is why your files disappear when it shuts down unless you save them somewhere that persists. And it is why the clock is running the whole time it is switched on, whether you are typing or making a sandwich.

The graphics card is the part you care about. A GPU is very good at doing thousands of small sums at the same time, which happens to be exactly what every AI model does. A CPU does a handful of sums very quickly, one after another. For AI work the difference is not subtle: a job that takes eight seconds on a rented GPU can take six or seven minutes on a decent laptop CPU, if it runs at all.

What you are not renting is a service that does the thinking for you. The model still has to be installed, configured and pointed at your files. On some platforms that is your job and the meter is running while you do it. On ours the machine arrives with the model already loaded, which is the difference between a 90-second start and a first evening spent reading installation notes.

Two shapes of work sit on top of that machine, and it is worth knowing which one you want before you start.

A session is you, at a keyboard, with an app open in your browser. You type, you look at the result, you change something. Good for figuring out what you want. The meter runs the whole time, including while you think.

A batch job is you handing over a list and leaving. Four hundred clips to upscale, a spreadsheet of product descriptions, ninety hours of interviews to transcribe. The machine works through the pile, writes each result as it finishes, shuts itself off, and emails you a link. The meter runs only while work is happening, which makes this by far the cheaper way to do anything at volume.

Most people start with a session to find the settings they like, then run a batch job with those settings against the real pile. That pattern costs a fraction of doing the whole thing by hand.

What "per hour" really means

Every provider advertises an hourly rate. Almost none of them charge you by the hour.

We bill by the minute. If your job takes 23 minutes, you pay for 23 minutes. The hourly number is a rate, not a unit, in the same way a car hire firm quoting a daily rate would still be strange about charging you for a full day when you had the car for twenty minutes.

Here is what the rate looks like against a real job. Generating 500 product images on Stable Diffusion XL takes about eight seconds each, so roughly 67 minutes of GPU time:

See the numbers

Model	Card	Time	Cost
SDXL Turbo	RTX 3090	10 min	$0.04
Stable Diffusion XL	RTX 4090	67 min	$0.50
FLUX.1 dev	L40S	150 min	$2.23

Fifty cents. That number does more work than any paragraph we could write about value. It is also the honest number, not a promotional one, and it is why the calculator on every model page shows you the total before you start rather than after you finish.

The thing the hourly rate does not tell you is how long your job will take, and that is the number that actually decides the bill. A faster card at twice the price can easily be cheaper overall, because it finishes in a third of the time. This is the single most common mistake people make when comparing providers on price alone.

What it costs, card by card

GPUVault hourly rental prices by card

Card	Memory	Class	Per hour	An 8-hour day
RTX 3090	24GB	Value	$0.22	$2
RTX 4090	24GB	Standard	$0.45	$4
RTX 5090	32GB	Standard	$0.69	$6
RTX A6000	48GB	Large memory	$0.79	$6
L40S	48GB	Large memory	$0.89	$7
A100 80GB	80GB	Data center	$1.35	$11
H100 80GB	80GB	Data center	$2.25	$18
H200 141GB	141GB	Data center	$3.10	$25

Most people should stop at the RTX 4090. It has 24GB of memory, which is enough for every image model in common use and for chat models up to about 30 billion parameters. The cards above it exist for two reasons: models that physically will not fit in 24GB, and jobs where finishing sooner is worth paying for.

The cards below it are worth knowing about. An RTX 3090 at $0.22 an hour is the same 24GB of memory as a 4090 for half the price, and it is maybe 40% slower. For an overnight batch job where nobody is waiting, that is a good trade.

What happens if you forget to turn it off

This is the question people are too polite to ask, so we will ask it for them: what stops a rented GPU quietly billing you for four days because you closed the laptop lid and went on holiday.

On most platforms, nothing does. This is the number one way people get hurt renting compute, and the horror stories are real. A machine you forgot about at $0.45 an hour is about eleven dollars a day. On an H100 it is fifty-four.

Three things are switched on by default here, and you do not have to configure any of them.

• A hard spend cap. New accounts start at $25. Nothing starts that would take you past it, and you raise it deliberately rather than discovering you already have.
• Auto-shutoff when idle. Sessions stop after 15 minutes with no activity, with a warning at 12. If you walk away mid-job, the job finishes and then the machine stops.
• A live cost meter. What you have spent and what is left, on screen, the whole time. Not buried in a billing dashboard you have to go looking for.

What the bill actually looks like

Abstract pricing is useless, so here is a realistic month for somebody doing this seriously but not full time.

A freelance photographer takes on a catalogue job. She spends two evenings working out a look — call it four hours in a session on an RTX 4090, at $0.45 an hour. That is about a dollar eighty. Then she runs the actual catalogue as a batch job: 1,200 images at roughly eight seconds each, which is about two hours forty minutes of GPU time, so another dollar twenty. Later in the month a client asks for the same shots upscaled, which is a short batch on a cheaper card, about forty cents.

Her month is under four dollars.

That is not a marketing example, it is just what the arithmetic gives you when the machine is only on while it is working. The bills that shock people are never the jobs. They are the machine somebody left running, or a platform where the meter starts when you press create and stops when you remember.

The other honest thing to say: at four dollars a month, the cost is not the interesting part of your decision. Whether the tool does what you need is. We would rather you spent the evaluation on that.

Where your files live, and who can see them

This is the question every business asks second, right after the price, and most providers answer it with a badge instead of a sentence.

While you are working, your files sit on the rented machine's disk. When the session ends, that machine is terminated and the disk is wiped. Anything you deliberately saved into your workspace folder is kept in your account and is there next time. Batch job inputs and outputs are deleted from our storage 24 hours after you download them.

We do not read your files. We do not use them to train anything, ours or anyone else's. There is no model in this business that gets better because you uploaded a client's photographs to it.

The part people do not think to ask about is who else has been on that hardware. Machines are wiped between tenants, and on our fleet a card is never shared between two customers at the same time — you get the whole card, not a slice of one. Some cheaper marketplaces do share, and that is worth knowing before you put anything confidential on them.

If your work is covered by a contract with specific requirements, read the actual terms rather than the badge on the homepage. That advice includes reading ours.

Rent or buy? The real number

Here is the maths, and we are not going to fudge it in our favour.

An RTX 4090 costs about $1,600 to buy. Running one pulls roughly 450 watts, which at the US average electricity price is about $0.07 an hour. Renting the same card from us is $0.45 an hour, all in.

So every hour you rent costs you $0.45, and every hour you run a card you own costs you $0.07 plus a share of the $1,600 you already spent. Divide it out and buying wins after about 4,183 hours of actual use.

That is 20 hours a week for four years. Or 27 hours a week for three.

See the numbers

Use	Hours over 3 years	Renting	Buying
2 hrs a week	312	$140	$1,621
5 hrs a week	780	$351	$1,653
10 hrs a week	1,560	$702	$1,705
20 hrs a week	3,120	$1,404	$1,811
40 hrs a week	6,240	$2,808	$2,021

Try your own number:

Be honest with yourself about the input. Most people who ask us this question say "oh, a few hours a week" and are right. At six hours a week you would need to keep the card for thirteen years to break even, by which point it will have been obsolete for a decade.

The people who should buy are the ones who already know they should: studios rendering every day, a shop with a constant fine-tuning workload, anyone whose GPU is busy while they sleep. If that is you, buy the card. Renting is not a religion.

There is a version of this argument that is not about money at all, and it is a fair one. A card you own has no meter running, and a meter running changes how you work. Some people experiment more freely when the clock is not ticking. That is a real cost and it does not show up in any spreadsheet.

The one thing to know before your first hour

Memory. Specifically VRAM, which is the memory on the graphics card itself. Think of it as the desk the model has to spread its work out on. Too small a desk and the job does not run slowly, it simply refuses to start.

This trips everyone up once. Speed determines how long a job takes; memory determines whether it happens at all. If a model needs 40GB and your card has 24GB, no amount of patience fixes it. You need a bigger card, a smaller version of the model, or a compressed one.

The practical version: check the memory requirement of the model before you check the price of the card. We put it on every model page for exactly this reason.

When renting is the wrong answer

Three cases, said plainly.

Your job is tiny and constant. If you want a small chat model answering questions all day, every day, a rented GPU sitting idle between requests is the wrong shape. You want something that charges per request, or a small card you own.

It already runs on your laptop. Modern Macs with a lot of unified memory run small models perfectly well. If yours does, use it. We would rather you did than pay us for something you already have.

You need it in a specific building. If your data genuinely cannot leave a particular room for legal reasons, no amount of encryption paperwork changes that. Buy hardware.

Everything else — the occasional big job, the thing you want to try once, the 400 clips that arrived on Friday, the model that is too big for your machine — is what renting is for.

There is a fourth case that is less clear-cut, and it is worth naming. Some people simply work better without a meter. If knowing the clock is running makes you cautious, and being cautious makes your work worse, that is a real cost even though it never appears in a spreadsheet. We have watched people talk themselves out of the experiment that would have worked because it felt expensive, when it would have cost eleven cents. If that sounds like you, either buy a card or make peace with the fact that a bad hour costs less than a coffee.

What to check before you pick a provider

Five questions, in the order that matters. None of them is the hourly rate.

1. What happens when I stop paying attention? Idle shutoff, spend caps, and whether either is on by default or something you have to go and configure.
2. How long from clicking start to doing work? Ninety seconds and forty minutes are both normal answers in this market, and the second one is expensive at any hourly rate.
3. What happens to a failed job? Whether you are charged for work that did not complete, and whether a batch that dies at item 380 of 400 leaves you 379 usable files or nothing.
4. Where do my files go? Not the badge. The sentence describing what is wiped, when, and whether the card is shared.
5. Can I get help from a person? At three in the morning, mid-job, this is the difference between a delay and a disaster.

Price is the sixth question, and by the time you have answered the first five it usually answers itself.

Where to start

Pick the job rather than the hardware. You do not need to know which card you want; that is our problem, and every model page tells you which one we put it on and why.

Or browse everything by the kind of work you do and work backwards from the task.