Rent a GPU in seconds.
Pay for the seconds you use.
Capacity from people with idle hardware, at a fraction of hyperscaler pricing. No quota request, no reservation, no minimum term.
Payments are not switched on yet, so nothing is charged today. Everything below describes how it works once they are.
To start
What you need to rent
An account
Sign up in seconds, no card on file. Nothing is charged until you launch, and payments aren't switched on for anyone yet.
An SSH key
Paste a public key at launch, the same one you already use. Kracht never sees or handles your private key.
A workload
Your own container image, or the default with CUDA and OpenSSH. SSH in and point a script at the GPU.
The catalogue
Find the card you need
Every GPU, searchable in one place: ours at the rate we bill, the rest as reference specs.
| GPU | Mem BW | Regions | Book | ||
|---|---|---|---|---|---|
| A10 | 24 GB | 600 GB/s | 6 regions | 1–8 GPUs | €5.28/day |
| RTX 4090 | 24 GB | 1.0 TB/s | 6 regions | 1–8 GPUs | €7.44/day |
| L40S | 48 GB | 864 GB/s | 6 regions | 1–8 GPUs | €15.36/day |
| A100 | 40 GB | 1.6 TB/s | 6 regions | 1–8 GPUs | €18.72/day |
| A100 | 80 GB | 2.0 TB/s | 6 regions | 1–8 GPUs | €26.88/day |
| H100 | 80 GB | 3.4 TB/s | 6 regions | 1–8 GPUs | €69.36/day |
| H200 | 141 GB | 4.8 TB/s | 6 regions | 1–8 GPUs | €89.76/day |
Every card we rent, at the on-demand rate we bill, across 6 regions. You book a whole host, one to eight GPUs, for as long as you hold it. What is free at any moment depends on who else is renting.
Straight answers
The three things people ask first
Renting a stranger's GPU raises fair questions. Pick the one on your mind.
The launch
What a launch sets up
One command brings a machine up. You are welcome to see exactly what it puts in front of you: an isolated box, reached with your own key, running the image you choose.
kracht launch --gpu 1 --region eu-nl --ssh- A container of your own: single-tenant, so nobody shares the card, the VRAM or the memory bus.
- Reached with your key, installed at launch. A box that could not be reached is refused before you are charged.
- The image you choose: your own container, or the default with CUDA and OpenSSH, ready to SSH into.
- The network you set: outbound internet, on or off, before it starts.
Two things it does only if you ask:
--image REFYour own container instead of the default. Any image the host can pull; pair it with --command so the container has something to run.
--max-price EURRefuse any listing dearer than this, per hour. The match fails rather than placing you on a card above your ceiling.
Isolation
What you are protected from
The card is yours alone
Kracht rents whole hosts, not slices: nobody shares your GPU, VRAM or memory bus while you hold it.
The provider cannot read your work
Sessions run in containers the host user cannot enter, and a provider gets no shell into yours.
Where we are honest about this
A determined host with physical access is a stronger adversary than a container boundary. Read the isolation page before you decide.
The full threat model, including what we do not defend against, is on the security page.
Billing
What you are actually charged
No hourly rounding, no minimum term, no reservation. A seven-minute job costs seven minutes, and the meter stops the moment the container does.
- H100, 41 minutes
- €1.97
- Same run, billed hourly
- €2.89
- You keep
- €0.92
Illustrative rate. Per-GPU pricing is on the pricing page.
- Billed per second. From the moment the container starts to the moment it stops. No rounding up to the hour.
- Your price is frozen. The rate is locked when you launch, so a provider re-pricing cannot change what a running rental costs.
- Credit is held, not spent. A hold is taken at launch and released when you stop, so a long run cannot quietly overdraw.
- Every charge is a line. In a ledger you can read, tied to the rental it came from.
On your terms
How to stop, and stop paying
Stopping is giving it back
End a rental with one command. There is nothing to cancel:
kracht stop <instance-id>Metering ends at the last second the container ran. The container and its scratch volume are destroyed at teardown, and nothing continues to bill after you stop.
The cost
What it would cost you
- Rate
- €2.89/GPU-hr
- Time
- 8 h
- GPUs
- 1
- Total
- €23.12
The CLI
Rent from the terminal
- 01
Make an API key
Create one in settings. It is shown once, and kracht login saves it on your machine.
- 02
Launch a card
Placed against a listing, the price frozen at match. Add --max-price to refuse anything dearer.
- 03
You are in
--ssh opens a shell once it is ready. kracht stop ends it, and the meter with it.
kracht login --key <YOUR_KEY>
kracht launch --gpu 1 --region eu-nl --ssh<YOUR_KEY> is an API key from your settings, shown once. Your SSH private key never leaves your machine; only the public half is installed on the box.
What a launch does
- Matches your request against a listing and freezes that listing's rate onto the rental.
- Holds credit for the rental when it starts and releases it when you stop.
- Starts your container, or the default with CUDA and OpenSSH, on a host nobody else is renting.
- Installs your SSH public key. A box that could not be reached is refused before you are charged.
- Meters per second from the moment the container starts to the moment it stops.
- Name
- kracht - rent a GPU by the second from the command line
- Synopsis
- kracht <command> [flags]
- Commands
- $ kracht login
- Save your API key
- $ kracht launch --gpu 1 --region eu-west --ssh
- Rent a card and open a shell
- $ kracht ls
- What you are running
- $ nvidia-smi
- Check the GPU
- $ kracht stop <instance>
- Stop, and stop paying
In the wild
What people do with compute
Four jobs the platform runs end to end today, argued from how the billing works.
A training run needs a large card for a few days and nothing for the weeks either side. Buying for the peak means paying for the trough.
Take it for the run, give it back
Metering starts when the container does and stops when it stops. There is no minimum and no notice period, so an aborted run costs what it used.
No reservation at all
Capacity is whatever providers have online right now. A machine already rented is not offered to you, so what you see is what you can have.
Change your mind for the price of an hour
Seven GPU models, from a 24GB 4090 to a 141GB H200. Trying the larger one costs the hour it takes to find out.
The other side
Have hardware instead?
The GPUs here belong to people who had them sitting idle. If that is you, the provider side pays 85% of every rented hour.
Resources
Where to go next
Frequently
asked questions
Per second, in euro, from the instance running to you stopping it. The rate is frozen onto the rental when it starts, so a provider re-pricing their card cannot change what you are paying mid-run.
Spin up a GPU, or earn from your own
Rent on-demand compute by the second, or list idle hardware and let it pay for itself. Same marketplace, both directions.
No card required
- 85%
- revenue to providers
- Per-second
- billing, live
- EUR
- metered to the microeuro