A GPU marketplace connects people who own graphics cards with people who need to rent one, rather than owning every machine itself. You search listings by card, price and reliability, rent by the hour, and stop when the job finishes. The owner earns money on hardware that would otherwise sit idle.
Most people meet the phrase after they have already decided to rent something. You go looking for a machine to run a job on, land on a page full of cards with prices beside them, and it is not obvious whether you are buying from a company or from a stranger with a spare computer. The answer changes what you get for your money, so it is worth five minutes.
If you have not read how renting GPU compute works, start there. This piece assumes you know what paying by the hour means and deals only with the question of who is on the other end of it — and, further down, who does the work of turning a bare machine into something you can actually use.
A marketplace does not own the hardware
A traditional cloud owns its machines. It buys the servers, racks them, keeps them cool, and rents them out. Everything you can buy from it looks much like everything else, because one company bought all of it.
A marketplace does something different. It lists capacity that other people own, handles the matching and the money, and takes a cut. The comparison everyone reaches for is Airbnb, and it is close enough to be useful: the platform does not own the rooms.
Where the comparison stops being useful is that nobody picks a flat based on how fast its kitchen is. Compute is interchangeable in a way that property is not, so a marketplace competes on price and reliability rather than on charm.
GPUVault is a marketplace in that sense. The cards you rent here belong to providers we have vetted, not to us. What we own is everything between you deciding to run something and it actually running, which turns out to be the difficult part and is where most of this page ends up.
Two sides, and why each needs the other
The demand side is easy to picture. A studio with a deadline and four hundred clips to render. A startup that needs eight cards for a fortnight of training and then none at all. An agency that took a video job in March and will not touch a GPU again until June. All of them want a lot of compute briefly, which is the shape of demand that owning hardware serves worst.
The supply side is less obvious, and it is where the model actually comes from. A graphics card that is switched off earns nothing, and it cost its owner the same as one that is busy. That is the entire economic engine here, and the idea has a name worth knowing.
See the numbers
| Utilisation | Hours rented | A month's revenue |
|---|---|---|
| 25% | 183 h | $82 |
| 60% | 438 h | $197 |
| 90% | 657 h | $296 |
So the people listing capacity are a more varied group than the phrase "data centre" suggests. Small data centres between contracts. Hosting companies with headroom. Operators who bought rigs for cryptocurrency and needed a second act. Businesses whose machines run flat out for six hours a day and sit dark for the other eighteen. And plenty of individuals with one good card in a home office. Listing that time turns a sunk cost into revenue. On our fleet a provider earns $180 to $420 a month per card, which is not a living, and is a good deal better than nothing. If you are sitting on hardware that idles overnight, listing a card with us is the other half of this page.
What you are actually renting
Here is the part that catches people out. You are not renting a graphics card. You are renting a whole computer that happens to have one in it, and the rest of that computer decides whether the job goes well.
A listing is really a bundle: the card and its memory, the processor beside it, the disk, the network connection, and the owner's record for staying online. A cheap listing on a slow disk will spend twenty minutes loading a model that a dearer machine loads in two, and you pay for those twenty minutes.
This is why the hourly rate is a poor way to shop, and it is the same lesson as the one in our card-by-card pricing guide: you are buying the time it takes to finish, not the hardware.
The bit marketplaces hand back to you
On most platforms, renting the computer is where the help stops. You have a machine, an operating system and a login. Everything after that is yours: drivers, the right version of Python, the model weights, an interface to talk to them, and the library conflict that only shows up at eleven at night. People who do this for a living call that a Tuesday. People who just wanted four hundred clips upscaled call it a reason to give up.
It is not the renting that is hard. It is the twenty steps after it.
So we do those instead. The site is organised by the job rather than by the card. You tell us what you are trying to make, or pick one from the list of what people run here. We choose the hardware, install the model, wire up an interface and start the machine, and what comes back is a link to a working app in your browser, usually in about ninety seconds.
You never open a terminal, you never install a driver, and you never need to know that the model you picked wanted 12GB of memory and got 24. Choosing the card is our problem, and it is a problem we would rather solve once, properly, than ask every customer to solve badly.
Why the price moves
On a marketplace, price is not set by a committee. It falls out of supply and demand, which means it moves.
| 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 |
When a new card lands and everyone wants it, the rate climbs until enough owners bring more online. When a generation ages, the rate falls. A shortage anywhere in the supply chain shows up here within weeks. Every price on this site comes out of one table that every page reads, so when a rate moves, the guides move with it the same day.
A moving market price does not mean a moving bill, though, and those get confused. You are billed by the minute with no minimum and nothing charged while the machine is off. New accounts carry a hard spend cap of $25 that nothing is allowed to exceed, and a session that sits idle for fifteen minutes shuts itself down, with a warning at twelve.
Where the model gets uncomfortable
Three things, and any marketplace that does not mention them is selling you something.
Variance. When hundreds of people supply the machines, the machines are not equal. On a fully open platform anyone can list anything, so the cheapest option is often an anonymous machine with no track record and nobody answering for it. It might be fine. It might drop out at hour six of an eight-hour job.
Disappearance. A host who wants their hardware back can end your session. Good platforms make that expensive for the host and warn you first. It still happens.
Access. Whoever owns the machine is standing next to your data. That is true of every cloud on earth, including the biggest, but on a marketplace it is a wider group of people, so the question of who they are stops being academic.
Our answer is to vet the provider rather than to insist on the building. Cards on this site sit in a small data centre, in a hosting company's spare capacity, or on a desk in somebody's home office, and we are not fussy about which. A well-run rig on a good connection, with an owner who answers for it, beats an anonymous listing in a data centre every time.
What we do insist on is everything else. A provider is checked before a card is listed and carries an uptime record afterwards. Every job runs on a dedicated machine rather than a slice of a shared one. Your data is deleted within 24 hours. That means fewer listings and less spread on price than a fully open platform, and we would rather call that a trade than pretend it is free.
The short version
A marketplace exists because expensive hardware sits idle and someone else needs it briefly. You get compute at a price the large clouds cannot go near; the owner gets revenue from a card that was doing nothing. The thing to check is not the hourly rate but who is accountable for the machine, and how much of the setup you are expected to do yourself.
Our answer to that last one is all of it. Pick the job, get a link. If you want the same argument pointed at your own desk, the rent-or-buy breakeven is where it lands. And if you would rather just run something, these are the jobs people start with: