Help

Whichever way suits you — they are all at checkout, and nothing is charged until you press Pay.

There is no fee on top: you pay the hourly price, and it is held safely until the GPU delivers. Time you do not use comes back. What each GPU typically costs: Pricing.

1. While your GPU is rented, the renter’s money waits in escrow. When the rental ends, what your machine delivered goes to your balance — you keep 90% of the price.
2. A short hold: each earning can be withdrawn 72 hours after it lands. It protects you and the renter against payment disputes.
3. Withdraw from Dashboard → Earnings, to:

Moving money out always asks for a 6-digit code — by email, or from an authenticator app if you set one up — so nobody else can send your earnings anywhere.

The rental stops on its own. You booked a number of hours; when the last one is up, the clock reaches zero and everything shuts down. You do not have to do anything, and nothing renews.

About ten minutes before the end you get a notification — the bell in the top bar, and an email unless you have turned those off — saying the rental is nearly over, so you can save your work or add more hours.

The moment it ends, three things happen:

• Anything still running is stopped: your jobs, your terminal, your notebook, any web app you started on the machine.
• The money is settled. The host is paid for the time they really delivered; the rest comes back to you. (See escrow.)
• The sandbox on the host's computer is deleted, and the files inside it go with it. (See where your files go.)

If you want to keep going, use + Extend on the rental page before the time is up. Extending is a separate payment that you make yourself — it is never taken automatically.

You are only ever charged for time the machine actually worked. That is not a promise somebody keeps by hand — it is how the billing is built.

While your rental is running, RentGPUPower checks the host's machine roughly every ten seconds and writes down one of two answers: delivering or not delivering. “Not delivering” covers the obvious case — the host's computer went offline — and three less obvious ones: it is switched on but ignoring the work you sent it, it is accepting your work and instantly failing all of it, or the app you asked it to run never started at all.

At the end, the maths is simply the fraction. Delivered 100% of your two hours? The host gets paid for two hours. Delivered half? The host is paid for half and the other half comes back to you, automatically, with no form to fill in and nobody to email.

If the machine goes dark while you are using it:

• After about two minutes of the machine being unreachable you get a notification telling you so. The Stop button on your rental page works at any time — you never have to wait for permission to stop — and pressing it refunds everything the host did not deliver, right then.
• If you are asleep or away, RentGPUPower does it for you: after thirty minutes of a machine being continuously unreachable, the rental is cancelled on your behalf and the undelivered time is refunded. You do not have to be watching.

Where the money lands. Unused time comes back to your RentGPUPower balance immediately, ready for your next rental — whichever way you paid, so there is nothing to wait for at your bank. Need what a card paid back on that card instead? Write to support. A rental that never starts is different: its card payment goes back to the card, in full.

What that balance can do. A rental you paid from your balance simply puts the money back the way it was. Money that came in from a card or SOL payment — a refund, or a payment that could not become its rental — is credit: it pays for any GPU time, but it cannot be withdrawn as cash. Only money you earn is cashed out — by hosting, or by a phone's work for API customers where that is offered. (That is what stops a stolen card being turned into cash through us.)

Two different places, and the difference matters a lot.

The rented machine's workspace — the folder called /workspace that your notebook and terminal open in — lives inside a sandbox on the host's computer, and it exists only for as long as your rental does. When the rental ends, that sandbox and its disk are deleted. Files left there do not survive. This is deliberate: the next renter must not be able to read your data off that machine.

My Drive is storage that belongs to your RentGPUPower account, not to any one rental. Files you put there stay there — between rentals, and when you have no rental at all — and you can pull them onto the next GPU you rent. It holds up to 2 GB, with each individual file up to 200 MB.

So the rule is simple: anything you want to keep, copy it to My Drive or download it before the time runs out. The ten-minute warning exists for exactly this.

One useful detail: when you paste a link into the workspace to import a dataset or a model, those bytes travel straight from the internet to the host's machine. They do not pass through RentGPUPower, and they do not eat your home connection.

On your rental page there is a button that says Stop rental, and it tells you roughly how much you will get back before you press it. You can stop it at any second — there is no minimum, no notice period and no cancellation fee.

The moment you press it: charging stops, everything running on the machine is shut down, and the time you paid for but did not use is refunded automatically.

You do not need to stop a rental to avoid extra charges. Nothing keeps running up a bill after the hours you booked — an unstopped rental simply reaches its end and settles itself. Stopping early is about getting money back, not about avoiding a surprise.

And if you walk away and forget: when a rental sits with nothing running on it for long enough, RentGPUPower stops it for you and refunds the rest. You get a warning about five minutes before that happens. Your own limit is an hour by default, and you can change it or switch it off entirely in Dashboard → Account. A rental with a job actually running is never stopped this way, however quiet it looks.

A cluster is several machines from one host — graphics cards, processors, on one computer or on several — rented together in one step. Each machine opens its own workspace, and from the cluster's page you can run a command on all of them at once or stop them together.

It is paid from your balance (add credits first, any way you like). Each machine is paid only for the time it actually delivers: if one stops, its unused time comes back to you on its own while the others carry on. Where the cluster’s page says so, the machines share a private network — each reaches the others by name, node1, node2, … — for coordinating work and moving moderate amounts of data.

Hosting several machines? Sign in to the same account on each, then combine them under Dashboard → My GPUs → Clusters — or in the app's My GPUs tab.

No. There is one charge, for the rental you booked, at the moment you pay for it. You are not charged again — not when the rental ends, not when it runs long, not next month.

There is no subscription here, no membership and no automatic top-up. A card is kept only if you tick Remember this card when you pay — Stripe keeps it, not us, and it is still only charged when you press Pay. Every payment is one you start yourself, on a page that shows you the exact amount first.

Adding hours to a running rental is a new payment you make deliberately, with its own checkout page and its own total. Money can also move the other way without you doing anything — refunds — but never towards us.

There is also no service fee for renting. You pay the hourly rate the listing shows, multiplied by the hours, and nothing else. RentGPUPower's cut (10%) comes out of what the host earns, not out of your price.

It means your money is held — by RentGPUPower, not sent to the host — until the rental is over and we know how much of it the host earned.

When you pay, the host gets nothing yet. Nothing at all. Their payout only exists after the rental ends, and it is worked out from the time their machine actually delivered. A host whose computer goes offline an hour into a four-hour rental is paid for one hour. A host who never delivers anything is paid nothing, and you get all of it back.

This is the whole reason you can safely pay a stranger up front. The host cannot take your money and disappear, because they never had it — and the software, not a support queue, decides who gets what.

On the rental page there is a panel called Escrow & delivery that shows the live figure: what percentage of your time has been delivered so far, and roughly what you would get back if you stopped right now. Treat that refund figure as approximately right — the delivery clock is a few seconds behind, and the final settlement rounds in your favour at the edges.

Be honest with yourself about this one: you are renting somebody else's computer.

The host owns the machine. Your work runs in a sealed-off sandbox, and the marketplace gives the host no button to open it — but it is their hardware, sitting in their house, and nobody can lock an owner out of their own computer. A determined host could look. So: do not put passwords, customer data, medical records, private keys or anything you would hate a stranger to read onto a rented GPU. Bring the work, not the secrets.

The links are the keys. When you open a notebook, a terminal or a web app on your rented GPU, the address contains a long secret. Anyone you send that address to can open the same thing while the rental is running, with no password asked. Do not paste it into a group chat. When the rental ends, the link dies with it.

What passes through RentGPUPower. Your terminal and notebook traffic is relayed through our servers to reach the host's machine (that is what saves the host from opening a port on their home router). Files you import by link do not — those go straight from the internet to the GPU. We keep the account and billing records you would expect: what you rented, when, and what it cost.

We never ask for your password inside an app. RentGPUPower signs you in only on its own pages. A sign-in form inside a notebook or an app on a rented machine is not ours — the machine's owner could have put it there — so never type your password into one.

Other renters cannot see anything. A rented GPU is yours alone for the whole rental. When one computer is rented to two people at once — a card each, or a card and its processor — each gets a sandbox and a network of its own, and neither can reach the other. Each rental's sandbox is destroyed when that rental ends, before the next renter arrives.

Its owner's identity was checked. A verified host set up bank payouts with Stripe, which checked who they are, the way a bank does, before it would pay them. We never see or show those details — only that the check passed. If Stripe closes their account, the badge goes with it.

What it does not mean. It tells you who the host is, not how good their machine is — the hardware check, reliability and reviews tell you that. Every rental is protected the same way, verified or not: your payment is held in escrow, and time the machine does not deliver comes back to you. On the marketplace, ✔ Verified hosts only shows just these machines.

Hosting? Set up bank payouts in your dashboard and the badge appears on all your machines once Stripe has checked you.

We measured that GPU ourselves. While nobody rents a machine, RentGPUPower sends its GPU a fresh puzzle — a long chain of matrix maths, the work AI does — and times the answer on our own clock. We then redo part of the work on our side to check the answers are right. No card can answer without doing the work, so a host's app cannot claim a faster GPU than it has.

What you see. The badge, and on the GPU's page the speed we measured (in TFLOPS) and when. A GPU that sent back wrong answers is taken off the market until it passes again. The test runs again every week, never while someone is renting the machine.

Hosting? Nothing to do: the test runs by itself while your GPU is idle, in an environment your machine already keeps for renters, and takes a minute or two.

Not your files. A renter's code runs in a sealed sandbox (Docker) that sees only its own workspace — never your documents, your other programs or your accounts. It is deleted when the rental ends.

Not your home network. The sandbox reaches the internet — so renters can download models and packages — but the app blocks everything from it to your router, your other computers, a NAS, printers or cameras, and to your computer itself. It checks this when it starts, before each rental and every few minutes while one runs, and puts it back if Docker restarts.

Never open by mistake. If the app cannot close your home network, nobody is let in: that machine is not offered to renters until it can, and the app keeps trying by itself. You don't have to set anything or read a warning for this to hold.

No spam from your address. Renters' code cannot send email straight from your connection (port 25 is closed to it), so nobody can get your internet address blacklisted that way.

You can see it. On your dashboard, each machine says 🔒 home network closed — or, when it is held back, why. No network for renters, not even the internet in the app's Settings rents a machine out meanwhile, with no network at all.

One thing to avoid: don't host from a computer connected to a work VPN. Renters' internet traffic goes wherever your computer's does — through the VPN too.

Yes — as the model. Run any app that serves an OpenAI-compatible API on the device (Ollama, LM Studio, llama.cpp's server, vLLM, PS5LM) with a model loaded, then add it in the RentGPUPower app: AI devices on your network → Add an AI device. The app checks which model it serves and how fast, and lists it as that model.

Renters reach the model, nothing else: three OpenAI calls — the model list, chat and completions — with your model pinned. No shell, no files, nothing else on the device or on your network. If the device stops answering, it leaves the marketplace until it is back, and renters get that time back.

Renting one? It works like any rental — per second, held in escrow — with a chat on the rental page and an OpenAI-compatible address for your own code. The device's owner runs the model server, so keep passwords and private data out of what you send.

Use your own device, and software you have the right to run on it.

The short version: you can lose the time the machine really did work for you, and your unsaved files. You cannot lose the rest of your money.

What RentGPUPower does not do is compensate you for lost time or lost work beyond refunding what was not delivered. If a run mattered, checkpoint it to My Drive as it goes. That is the honest limit of what a marketplace between two strangers can promise.

A little. A phone or tablet cannot host a rental — a browser stops running the moment you switch apps or lock the screen — but it can do small pieces of AI work while you watch. Open the earn page in the phone's browser and sign in. It downloads a small AI model once (about 115 MB, so use Wi-Fi), then works through short batches that the server checks against known answers.

• Only checked work counts, and only while the page is open with the screen on.
• The test batches pay spend-only credit, roughly a fifth of a cent an hour (about 0.2¢/hour): it pays for GPU time here, can never be withdrawn, and shows on your dashboard as phone credit. It is a way to try the marketplace, not an income.
• The phone works flat out and gets warm, so keep it plugged in.

A computer with a graphics card earns far more than a phone — see Earn with your PC.