How can we
help you?

Compute is the live marketplace that rents and meters GPUs. Studio is the planned workflow surface that assigns hardware per step. Autopilot is the planned optimisation surface that reads the same metering, simulates a change, and waits for approval.

Compute. Rentals have been placed, metered and settled end to end on real hardware, in euro, with the split applied. Studio and Autopilot are still being built, and each of their pages says so at the top.

No separate product price exists today. Compute usage is metered per second. Studio and Autopilot are planned around that same meter and euro ledger, and their commercial terms will not be advertised before they exist.

From providers who list machines they own. Kracht runs no data centre of its own: the control plane matches, meters and settles, and the GPU you rent is a provider's card, passed through to your container for the length of the rental and handed back when it ends.

No. Payments run against a mock provider, so every balance today is credit granted by hand, and nothing is charged to a card or paid out to a bank. The ledger underneath it is proven: per-second metering, the provider split and settlement in euro all run for real. The payment rail is the part still to connect.

Yes. Stop a rental with a snapshot and its scratch volume is captured; launch a new rental from that snapshot and it comes up on a fresh volume with your files in place. This has been exercised end to end on real hardware. Snapshots are held on platform storage for now rather than in object storage.

In the Netherlands, which is why the ledger is in euro and the payment provider is an EU one. The control plane that matches, meters and settles runs in Frankfurt. The GPUs are wherever the providers are, and each listing shows its region.

Launch to a running container is typically under a minute, most of which is pulling your image. The figure quoted on the marketing pages is illustrative, and stays labelled that way until there is enough real traffic to publish a median.

No. You can sign up and look around without one, and only add billing when you move to paid capacity. There are no quota forms or sales calls in the way.

H200, H100, A100, L40S, RTX 4090 and A10 today, across six regions. The catalogue on the users page is the live list, with the on-demand rate we bill on each card we rent.

Two keys, not one. The SSH public key is your login; the WireGuard public key is the route in. Providers sit behind NAT, so without the WireGuard key nothing is published on any address and the machine bills without ever being reachable. The launch screen refuses a rental missing either.

Yes. Pass any public or private registry image and it is pulled onto the host before your job starts. The default image exists because it ships an sshd, which stock framework images do not.

Two ways. Several cards on one host as independent shards, which is what a hyperparameter sweep wants, or a whole host so that its own GPUs and its interconnect are yours and nobody else is on the box. A batch job can fan across as many free hosts as it needs.

Whichever is closest to your data. Rates do not vary by region today, so the choice is about how long it takes to get your dataset onto the machine rather than about price.

The worker notices the missing heartbeats, marks the host offline, and releases the rental so it stops billing. Work in progress is lost unless you wrote it somewhere durable yourself: there is no automatic checkpointing, so on a long run, checkpoint to your own storage on a schedule you choose.

The card is yours for the whole rental. No time-slicing, no partition shared with a stranger, and the workload runs in a sandbox the host does not get a shell into. The container and its scratch volume are destroyed when the rental ends, so nothing you wrote to disk survives it unless you moved it off.

Stop the instance. Billing is metered per second from usage events, so the meter stops when the container does and settlement charges only the seconds that actually ran.

Yes. You get an SSH shell, your own logs, and whatever tooling your image ships. There is no console to learn and nothing to click: if it works over SSH on a box you own, it works here.

Nothing on our side stops. The container is still rented and still metered until you stop it, because the meter follows the rental rather than the process inside it. A crashed script on a card you are holding is a card you are holding.

Not a formal one yet. Nothing measures uptime to a number we would publish, and the status page says plainly what the platform watches and what it does not, rather than quoting a figure it cannot stand behind.

Outbound, yes. The container reaches the internet and resolves DNS, so a pip install, a git clone or a registry pull all work. Inbound is only over the WireGuard route you peered with, and the provider's own LAN and sshd are blocked from inside the rental, both ways.

Everything is metered per second and drawn from your prepaid balance, in euros. You pay only while an instance runs: no hourly minimum, no egress charge. The price is frozen when the match is made, so a provider re-pricing cannot change what a running rental costs you.

A running rental holds credit against your balance as it meters, so it cannot quietly overspend. When there is not enough left to cover the next interval the instance is stopped rather than allowed to run up a debt.

Not yet. Everything is on-demand and per-second today. If a commitment is what makes the numbers work for you, tell us what shape it needs to be and it moves up the list.

The settled decision is Mollie, which is EU-native and does iDEAL and SEPA. Card payment is part of that. Payments are not switched on yet, so nothing is charged today; everything above describes how it works once they are.

We are an EU company billing EU customers, so the ledger is EUR end to end, stored at six decimal places and settled to the microeuro. Where a competitor's price appears on this site it is converted at a rate printed next to it, never relabelled.

No. Topped-up credit is spendable on compute and is not withdrawable; only money earned as a provider can leave to a bank account. Card in and bank out for the same person is close enough to money transmission that a Netherlands company should not do it, so a payout is capped at what the account earned, less payouts already taken, and never above the balance actually held.

No hourly minimum and no rounding up to a block. Usage is metered per second from the rental's own heartbeats, and settlement charges the seconds that ran at the rate frozen when the match was made. The provider's share is split exact-integer, so no cent is created or lost between the two sides.