Preparing an invite-only beta
Powerful compute.Responsible by design.
Rent a GPU by the second from independent hosts, at prices made for real AI work. We check every card, set up your app and save your work, so you can focus on your models, images and ideas.
Prepaid credit. Billed by the second, once your app is ready.
- Checked before you payEvery card passes a hardware check and a test rental before anyone rents it.
- No surprise billsJust the card's rate, by the second. No storage or transfer fees. Set a time limit and see the most it can cost.
- Your work, savedYour workspace is saved every few minutes. If a host drops, you pick up from the last save on another card. How
- The right cardWe suggest a size for your app and show how fast each card is next to an RTX 3090.
01 / Choose your work
One card can go a long way.
For jobs that fit on a single GPU. Memory needs depend on your model, data and settings.
Fine-tune a model
LoRA / QLoRA · around 8B and smallerJupyter with Unsloth, TRL and vLLM. Keep checkpoints in/workspace and use training code that can resume.Fine-tuning guide
Make images & video
ComfyUI · Image makerBuild image and video workflows with compatible models. Choose a card with enough memory for your workflow and outputs.Image maker guideQueue a batch job
Your script · the km commandUsekm run --wait 6h -- <command> to wait up to six hours for a card. The rental ends when your command finishes.Developer guide
Run open chat models
Ollama · Open WebUITry open models in a browser chat app. Allow memory for both the model and the length of your conversations.Chat model guide02 / Find your card
Expected prices.
Real numbers to plan with.
Choose the cheapest available card in a memory size, or an exact model. We choose the host. Your rental keeps the rate it started at.
RTX 3090
$0.155/hour
A low-cost 24 GB starting point24 GBRTX 4090
$0.428/hour
24 GB for a range of AI workloads32 GBRTX 5090
$0.524/hour
32 GB for more memory headroomCompare all 10 cards
| Card | Memory | Expected price | Elsewhere, Oct 2026 |
|---|---|---|---|
| NVIDIA RTX 4090 | 24 GB | $0.428/hr | $0.33–$0.47/hr elsewhere |
| NVIDIA RTX 5090 | 32 GB | $0.524/hr | $0.47–$0.69/hr elsewhere |
| NVIDIA RTX 3090 | 24 GB | $0.155/hr | $0.13–$0.22/hr elsewhere |
| NVIDIA RTX 4080 | 16 GB | $0.25/hr | $0.20–$0.27/hr elsewhere |
| NVIDIA RTX 3090 Ti | 24 GB | $0.175/hr | $0.18–$0.24/hr elsewhere |
| NVIDIA RTX 5080 | 16 GB | $0.30/hr | $0.25–$0.39/hr elsewhere |
| NVIDIA RTX 5070 Ti | 16 GB | $0.26/hr | $0.16–$0.53/hr elsewhere |
| NVIDIA RTX 4070 Ti SUPER | 16 GB | $0.19/hr | $0.17–$0.27/hr elsewhere |
| NVIDIA RTX 5060 Ti 16GB | 16 GB | $0.15/hr | $0.12–$0.27/hr elsewhere |
| NVIDIA RTX 4060 Ti 16GB | 16 GB | $0.14/hr | $0.14–$0.27/hr elsewhere |
USD per GPU-hour, billed by the second. Expected launch prices, not live availability. The comparison table uses October 2026 market ranges.
03 / A look inside
The essentials, in one place.
Your running time, spend and latest complete save. An illustrative rental, with sample data.
Image maker
ComfyUI · NVIDIA RTX 4090 · 24 GB
/workspace04 / From idea to app
A few choices.
Then you're in.
Choose the work and the card. We handle host selection and open the app in your browser.
When rentals open, eligible new Google or GitHub accounts can try templates with $5 credit, while the monthly allowance lasts. No payment card needed.
Explore GPU rentals- Sign inUse Google, GitHub or an email link.
- Choose your appNotebook, Image maker, Chat or Fine-tune.
- Choose a card, then startWe suggest a size for your app. Take the cheapest card in it or an exact model, and set a time limit if you like.
- Open your workspaceUse your browser, or
km openfrom your computer.
05 / Designed for interruptions
A host can drop.
A complete save can help.
Home machines can lose power or internet. Recovery restores saved files on another available card of the same model. It does not preserve a running program's memory.
Save the workspace
Keep outputs and checkpoints in
/workspace. Periodic copies go to our storage; only a fully uploaded copy counts.Find a matching card
A replacement must be free and match the original GPU model. Capacity is limited in beta, so recovery is not guaranteed.
Restore what was saved
Restore the last complete copy at the booked rate. Paid time lost since that save is refunded; offline time is not billed.
What a save cannot keepChanges since the last complete save and work held only in memory can be lost. Training needs resumable checkpoints; notebook cells may need to run again.
Download before you stopRecovery protects an active rental. After it stops, its last save is kept for a day so you can start again with it, then deleted. Download anything you want to keep longer.
Common questions
Questions, answered plainly.
What is KernelMachine?
A GPU rental service for work that fits on one gaming card: small fine-tunes, image generation and batch jobs. Choose your task and card; we pick the host. Your app opens in a browser, and developers can run scripts with km.
How much does it cost to rent a GPU?
Expected launch prices are about $0.155/hr for an RTX 3090, $0.428/hr for an RTX 4090 and $0.524/hr for an RTX 5090. These are expected prices, not live offers. Billing is by the second once your app is ready, using prepaid credit with no automatic top-up. When rentals open, eligible new Google or GitHub accounts can try templates with $5 of credit, while the monthly allowance lasts. No payment card is needed for that trial.
Do I need SSH, Docker or Linux to rent a GPU?
No. Pick Notebook, Image maker, Chat or Fine-tune and open it in your browser. There are no SSH keys to upload or ports to configure. The km command is available for developers who want to work from a terminal.
Which GPU should I rent?
Choose enough memory for your model, data and settings, then the cheapest card in that size or an exact model. We suggest a size for each app (16 GB for Notebook and Chat, 24 GB for Image maker and Fine-tuning) and show each card's speed next to an RTX 3090. Cards have at least 16 GB; an RTX 3090 or 4090 has 24 GB, and a 5090 has 32 GB.
Can I cap what a rental costs?
Yes. Choose how long it runs (1, 3 or 8 hours, or a day) and the rent page shows the most it can cost. It stops by itself once that much time is used, with an email 10 minutes before. You can add time or remove the limit from the rental's page.
What if a host drops or something goes wrong?
Recovery needs a complete workspace save and another available card of the same model. Unsaved changes and a running program's memory can be lost; training code must support checkpoint recovery. Paid time lost since the save is refunded, and offline time is not billed. For other problems, "It's broken" stops the rental and refunds its current lease, or the whole rental if it ends within 30 minutes of becoming ready, up to three automatic refunds a month.
What happens to my files when I stop?
Stop deletes the rental's files. Its last save is kept for a day so you can start again from it, then deleted. Download anything you want to keep longer.
Whose machines will I be renting?
The first beta uses our own home machines. Outside hosts follow only after the required checks are ready. These machines can lose power or internet, and a host can inspect renter data. Keep passwords, keys and private data off rentals.
When does KernelMachine open?
No launch date is set. We are preparing an invite-only beta on our own machines and must verify recovery from a dropped host before rentals open. Join the waitlist for one email when access opens; joining does not guarantee a beta invitation.
Can I rent out my own graphics card?
Join the host waitlist. You need a 64-bit desktop or server running Ubuntu 24.04 or 26.04, an NVIDIA RTX 30-series or newer card with at least 16 GB, and 250 GB free disk (500 GB recommended). Outside hosts follow the beta on our own machines. Hosts keep 80% of rental revenue, or 85% as top hosts. At the expected price, an RTX 4090 earns about $0.34 per rented hour before electricity, hardware and internet costs. Unrented hours earn nothing; demand is not guaranteed.
Our principles
How we run it matters too.
Four principles guide what we build. What's in place today, and what isn't yet
- Responsible useCompute for real work, under a clear policy everyone agrees to: no pornographic content, nothing that exploits anyone, nothing illegal.
- PrivacyWe don't look inside your work, and hosts get no extra access to it.
- FairnessClear prices by the second, rules written down, and a fair deal for renters and hosts.
- AccessibilityCards already in homes and offices make serious AI work more affordable.
For the other side of the card
Your idle GPU
could earn its keep.
Hosts keep 80% of rental revenue, or 85% as top hosts, before electricity, hardware and internet costs. Earnings depend on demand.
Explore hostingA compatible desktop or server
- NVIDIA RTX 30-series or newerAt least 16 GB of GPU memory
- Ubuntu 24.04 or 26.0464-bit desktop or server; no laptops or Macs
- 250 GB of free disk500 GB and wired internet recommended
Our own machines come first. Outside hosts follow gradually after the required checks are ready.
Preparing an invite-only beta
Get an email when access opens.
No launch date is set. We'll email you once when KernelMachine opens. Nothing else.