For renters · Updated
GPU rental marketplaces compared: Vast.ai, RunPod, TensorDock, Salad and more (2026)
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: Vast.ai is usually cheapest per hour and has the widest choice, but leaves picking machines, extra fees and failures to you. RunPod reportedly no longer takes new home hosts, and TensorDock now accepts only data-center hardware. Salad is cheap but jobs can be interrupted without warning. io.net, Clore.ai and Nosana pay hosts in their own crypto tokens. We found no marketplace that refunds failed rentals automatically.
GPU marketplaces connect people who need a GPU with people and companies who have spare ones. They fall into two camps. In host-priced markets (Vast.ai, Clore.ai, TensorDock), hosts set their own prices and compete on a list. In platform-priced networks (Salad, Nosana, io.net), the platform sets the rate. Everything below was checked in early October 2026, and each fact links to its source.
For renters
| Platform | How you choose | Watch out for |
|---|---|---|
| Vast.ai | Browse and filter hundreds of host offers | No refunds once credit is spent; storage billed on stopped machines and bandwidth billed per byte on many machines (billing); services on random port numbers (networking) |
| RunPod | Pick a GPU type from RunPod's list | A stopped machine can lose its GPU to another renter, leaving files stuck (docs); one disk type costs $0.20 per GB a month while stopped, double the running price (pricing); basic SSH doesn't support SCP or SFTP (SSH docs) |
| TensorDock | Host offers, now data-center hardware only | 1.6 out of 5 on Trustpilot, with machines that never start and unanswered support (Trustpilot) |
| Salad | Platform-set prices on home PCs | Jobs "will be interrupted with no warning," and starts are slow (FAQ) |
| Clore.ai | Host offers | Payment in crypto only; disputes start with the host (docs) |
| io.net | Platform-priced clusters | About 1.8 million fake GPUs appeared in a 2024 attack, since fixed with a compute check (Cointelegraph) |
| Akash | Bid on providers' offers | Few consumer cards: three RTX 4090s on the whole network on 6 October 2026 (Akash feed) |
What renters pay
Median renter price per card-hour among the top 64 offers in Vast.ai's public feed on 6 October 2026 (Vast.ai feed):
| Card | GPU memory | Median price per hour |
|---|---|---|
| RTX 3070 | 8 GB | $0.096 |
| RTX 3090 | 24 GB | $0.155 |
| RTX 4090 | 24 GB | $0.428 |
| RTX 5090 | 32 GB | $0.524 |
The hourly price isn't the whole cost. Add storage, bandwidth, time lost to machines that don't start, and work lost to machines that vanish. What to check before you rent covers each one.
For hosts
| Platform | Who can host | Setup | Fee and payment |
|---|---|---|---|
| Vast.ai | Anyone with suitable hardware | Install Ubuntu, partition disks, install drivers, open router ports, fixed public address, Secure Boot off (verification) | Fee no longer published since June 2024 (update); weekly payouts from $20 (payouts) |
| RunPod | Reportedly closed to new home hosts since early 2026 (RunPod's announcement page has since moved, so we couldn't re-check it) | n/a | n/a |
| TensorDock | Data-center hardware only (host page) | Hardware must sit in a data center | 25% fee |
| Salad | Windows PCs | A Windows app running jobs in Windows' built-in Linux layer; may need startup-setting and memory changes, and can clash with games' anti-cheat (Salad) | Platform sets rates, e.g. $0.173 an hour for an RTX 4090 while a job runs on 6 October 2026 (Salad feed) |
| Clore.ai | Anyone | Open router ports, install the driver by hand (install docs) | About 5% on standard rentals, plus 15% when the renter pays in Bitcoin or stablecoins; paid in the CLORE token (fees) |
| io.net | Anyone who locks up a token deposit | New machines can wait up to 12 hours as "Pending" (docs) | Paid in the IO token (rewards) |
| Nosana | Linux PCs, one card each | Install the driver, Docker and NVIDIA's container tools by hand (docs) | 10% fee; paid in the NOS token |
| Akash | Server clusters, or Windows PCs via Homenode (since February 2026) | Homenode sets up dual-boot from a Windows installer (Homenode) | Not covered here |
How much a host actually clears after electricity depends mostly on the card. Our earnings guide works it out for the RTX 3070, 3090, 4090 and 5090.
Which one should you pick?
- Cheapest consumer cards, and you're happy to check machines yourself: a host-priced market such as Vast.ai. Read the billing page first.
- Batch jobs that can restart at any time: a platform-priced network such as Salad can be very cheap.
- Long jobs where failure is expensive: data-center machines, and a provider that bills from when your app works and refunds failures.
- You want to pay or be paid in dollars, not tokens: avoid the token networks.
Where KernelMachine fits
KernelMachine is preparing an invite-only beta; no launch date is set. It's built around the gaps above:
- Renters pick what they want to do and a size, or a card by name; we pick a checked host and open it in the browser. If a host drops, the rental can recover from its last complete save when another card of the same model is available, and nobody pays for the work lost in between. Billing is to the second from when the rental is ready.
- Hosts paste one line on Ubuntu, with no router ports and no driver hunting. They keep 80% (85% as top hosts), paid in dollars through Stripe, PayPal or Wise: monthly once $25 is ready, weekly to PayPal from $20. How hosting works.