For renters · Updated
How to rent a GPU in 2026: what to check before you pay
KernelMachine is preparing an invite-only beta; no launch date is set. Where this guide describes KernelMachine, it's how it works at launch. Join the waitlist
Short answer: work out how much GPU memory your job needs first, then pick the cheapest machine that has it. Before you pay, check when billing starts, whether failed machines are refunded, what else is billed, whether you can cap the cost, and what happens to your files when the rental stops.
A GPU (graphics processing unit, or graphics card) is the chip that does the heavy math for AI models, image generation and 3D rendering. Renting one means using a GPU in someone else's computer over the internet, paid by the hour. This guide walks through what to check, in the order that matters.
1. Work out how much GPU memory you need
GPU memory (often called VRAM) is the hard limit. If your model doesn't fit, it either fails with an "out of memory" error or runs many times slower. Speed is a preference; memory is a requirement. So choose by memory first.
| What you want to do | GPU memory to look for |
|---|---|
| Chat with an 8B open model (4-bit) | 8 GB or more |
| Chat with a 32B model (4-bit) | 24 GB |
| Chat with a 70B model (4-bit) | 48 GB, or two 24 GB cards |
| Make images with SDXL | 12 GB |
| Make images with Flux at full precision, or make video | 24 GB or more |
| Fine-tune an 8B model with QLoRA | 16 to 24 GB |
The full rules of thumb, with the math, are in how much GPU memory you need.
2. Choose the kind of provider
- Big clouds (AWS, Google Cloud, Azure): reliable and well documented, but usually the most expensive per hour. New accounts often have to ask for permission (a "quota") before they can use GPUs.
- GPU clouds: companies that run data centers full of GPUs and rent them at fixed prices. Good reliability, mostly data-center cards such as the A100 and H100.
- Marketplaces (Vast.ai, Salad, Clore.ai and others): independent hosts rent out their own machines, often consumer cards like the RTX 3090 or 4090. Usually the cheapest per hour, but quality varies from machine to machine. Our comparison covers the main ones.
Prices give a sense of scale. On 6 October 2026, the median RTX 4090 offer in Vast.ai's public feed was about $0.43 an hour, and an RTX 3090 about $0.16 (Vast.ai offers feed, top 64 offers per card). Data-center cards such as the H100 cost more, typically a few dollars an hour.
3. Check when billing starts
Many services start the meter when the machine boots, not when your software works. If your image takes 15 minutes to download, or the machine never starts properly, you can pay for time you couldn't use. One Vast.ai renter described being charged $1.70 for about 20 minutes of a service that never worked (Trustpilot). Look for a provider that bills from when your app is reachable, or at least doesn't bill while the image downloads.
4. Read the refund policy before you top up
Machines fail. What matters is what happens next. Vast.ai's billing docs say there are no refunds once credit is spent (Vast.ai billing). In our reading of about 60 negative renter reviews across five platforms, refused refunds were one of the top four complaints (why rented GPU jobs fail). Prefer services that refund failed machines without a fight, and top up small amounts until you trust one.
5. Look for charges on top of the hourly price
- Storage on stopped machines. Vast.ai charges storage while a machine is stopped. On RunPod, one disk type costs $0.20 per GB a month while stopped, double its running price (RunPod pricing).
- Bandwidth. Vast.ai lists bandwidth prices "for every byte sent or received" on many machines (Vast.ai billing FAQ). Downloading big models and datasets can add up.
Work out the all-in cost for a typical session, not just the hourly rate.
6. Know what happens to your files
This catches people most. On some platforms, a stopped machine's GPU can be rented by someone else, and your files are stuck until it frees up (RunPod docs). Others delete everything when you stop. Before you start real work, find out: are files kept after a stop, for how long, and can you copy them out without a running GPU? Then save anything important to your own storage as you go.
7. Decide how you'll connect
Some services give you an address, a random port number and an SSH login (SSH is the standard way to get a command line on a remote computer). You set up keys and tunnels yourself. Others open Jupyter, ComfyUI or a chat screen in your browser with one click. If you plan to copy files with scp or rsync, check that they work: RunPod's basic SSH doesn't support SCP or SFTP (RunPod SSH docs).
8. Think about who can see your data
Whoever owns the hardware controls it. On a data-center GPU, that's a company with security staff and contracts. On a marketplace machine in someone's home, it's that person. Consumer cards can't hide your data from their owner. Only some data-center cards (such as the H100) support "confidential computing" that can. Don't put passwords, private keys or customer data on a machine you don't trust.
The checklist
| Question | Good answer | Warning sign |
|---|---|---|
| Does my job fit in the card's memory? | Model size plus headroom fits | "It should be fine" |
| When does billing start? | When the app or shell is reachable | When the machine is created |
| What if the machine fails? | Automatic or easy refund | "No refunds after credit is spent" |
| What else is billed? | Nothing, or clearly listed | Storage and per-GB bandwidth buried in docs |
| Can I cap the cost? | A time limit with the most it can cost, shown up front | A meter that runs until you notice |
| What happens to my files at stop? | Kept for a while, easy to copy out | Deleted at once, or stuck until the GPU frees |
| How do I connect? | Browser buttons, or SSH that just works | Random ports, manual key setup |
How KernelMachine answers these
- Choosing: pick what you want to do, and we suggest a size with enough memory for it. Take the cheapest card in that size or one by name; we pick the host. Each card shows its speed next to an RTX 3090, and you see the price before you start. Prices by card.
- Billing: to the second, from when your rental is ready. Downloads and set-up are free. Set a time limit (1, 3 or 8 hours, or a day) and the rent page shows the most it can cost; it stops by itself, with an email 10 minutes before.
- Failures: top hosts are offered rentals first, and if a host drops, your rental can recover from its last complete save when another card of the same model is available. You don't pay for the work lost in between, and the host isn't paid for it. Offline time is never charged, and "It's broken" stops the rental and refunds it automatically.
- Extra charges: none beyond the hourly price. Unused time goes back to your balance, and credit never expires.
- Files: deleted when a rental stops, so download or copy your work out first. A forgotten rental gets an email after 30 minutes unused and stops after an hour.
- Connecting: open your notebook, ComfyUI or chat model in the browser, or use the
kmcommand. No SSH keys or port numbers. - Whose machines: home and small-business PCs that passed our automatic checks and, for most hosts, a 48-hour trial. Their owners could inspect them, so keep secrets off.