For builders
Rent the compute you need
- Launch an in seconds. One GPU, or a whole host
- Per-second billing with no egress fees or minimums
- Bring your own image and framework. Zero lock-in
Marketplace live · billed per second
Idle GPUs become AI infrastructure.
Rent AI compute by the second from people with spare capacity. You pay only for what you use.
The owner cannot open a shell into your job. How isolation works
- Billed per second
- Providers keep 85%
- Six regions, no lock-in
Runs your entire stack
Bring your framework, image, or orchestrator. It runs unchanged on marketplace GPUs.
Frameworks
Runtimes & tooling
Why it costs less
Because someone else already bought it
Every GPU here is spare capacity someone else paid for. They keep 85% of what it earns, you pay by the second for the time you actually use, and the marketplace is what puts the two together.
FIG 01 two sides, one meter
renter · pays per second
Names a GPU model and a region, never a machine, and pays by the second it runs.
example · H100 80GB at €2.89/h
€0.000000 after 0s
- provider · 85%
- €0.000000
- kracht · 15%
- €0.000000
For providers
Earn from idle hardware
- List the GPUs you're not using and set them to auto-rent
- Earn 85% of every rental, metered per second
- Every rental is sandboxed from your host and your files
One rental
Launch, run, settle, and the split at the end
A renter's request is matched and its price frozen, runs in their own container, and settles when the meter closes. The provider's share and ours are cut from the same charge, to the microeuro.
A worked example at the catalogue rate, not a rental that happened.
The math
The same card, a lot less.
~62%lower on average,
vs. hyperscaler list price
vs. hyperscaler list price
A marketplace of idle cards has no datacenter to amortise, so the rate is the rate, not a markup on one.
measured: observed once, at a stated time, as of 2026-06-01, source hyperscaler on-demand list prices, in USD
[↓] an H100, per hour, here€2.89
[≈] hyperscaler list price≈€6.09
// SOLID: KRACHT · GHOST: HYPERSCALER LIST// RATE: $1.1463 to €1 · ECB euro reference rate, 22 September 2026// LIST PRICES READ: 2026-06-01
- H200: €3.74 per hour here, ≈€9.25 list price, 60% less.
- H100: €2.89 per hour here, ≈€6.09 list price, 53% less.
- A100: €1.12 per hour here, ≈€3.20 list price, 65% less.
- L40S: €0.64 per hour here, ≈€1.70 list price, 62% less.
- RTX 4090: €0.31 per hour here, ≈€0.96 list price, 68% less.
Billing is always in euros. This switch only changes how the comparison is shown; a converted figure carries ≈.
Marketplace rates move with supply and demand.
One control plane
Rent it, run it, and see what it cost
Three products, one meter, on hardware somebody else already bought.
Compute
power
Somebody else's idle card, by the second, in euro. Scale it to a cluster, and see what every instance costs or earns from one console.
- install$
- launchwaiting
- runwaiting
- settlewaiting
Studio
plannedorganise
Studio draws a workflow as many rentals, gives every step its own card, and lets a step waiting on a person hold nothing at all.
Studio · agent workflow
every GPU step gets its own card · one waiting on a person holds none
Autopilot
plannedoptimise
Autopilot reads the same metering as your bill, finds a card it is not using, prices the fix against what you already paid. Nothing moves until you agree.
Autopilot · orchestrator
reads your bill, then calls each card: keep, release, or rightsize
The console
Sort the market by whatever the job actually needs
The real table from the launch screen, not a picture of it. Sort it.
The cheapest card that can hold your model
Providers set their own price; billing runs per second against the price frozen at match time.
- Per-second metering, from launch to settle
- Price snapshotted when the match is made
- No minimum term and no egress charge
kracht · launch
| Launch | |||||
|---|---|---|---|---|---|
| RTX 40908.9 | 24 GB | eu-nl | 0.94 | €7.44/day | |
| L40S | 48 GB | us-east | 0.97 | €15.36/day | |
| A1008.0 | 80 GB | eu-central | 0.96 | €26.88/day | |
| A108.6 | 24 GB | ap-south | 0.95 | €26.88/day | |
| H1009.0 | 80 GB | eu-west | 0.99 | €69.36/day | |
| H200 | 141 GB | us-west | 0.98 | €89.76/day |
From the terminal
Rent it, run it, stop it
Three commands from a card to a settled bill. What the window prints is what the kracht CLI prints.
A rental
$ kracht launch --gpu 1 --region eu-nl --ssh
Rented 8fa2c1d0 on H100 80GB (eu-nl) at 2.890000 EUR/hr
Billing has started. Stop it with: kracht stop 8fa2c1d0
Waiting for it to finish provisioning...
$ python train.py
epoch 3/3 · saved ./out/model.safetensors
$ kracht stop 8fa2c1d0
8fa2c1d0 is stopping. The provider picks this up on its next heartbeat, so `kracht ls`
may still show it running for a few seconds. Billing stops at teardown.
Under the hood
Core capabilities
The guarantees the marketplace is built on, enforced in the code, not features bolted on top.
Per-second metering
Billed to the microeuro, from launch to settle.
Price locked at match
Snapshotted when you match; a re-price cannot touch a running rental.
Bring your image
Your container or ours, with the Dockerfile untouched.
Your SSH key
Connect straight to the machine. No proprietary shell in between.
Snapshot and restore
Save a rental and resume it later, files intact.
Dead-node rescue
A lost rental reschedules; no bill for a window the host could not serve.
Region-pinned
Six regions, fixed at match time, not best effort.
Exact-integer ledger
Append-only, 85% to providers, no cent leaks.
Fleet preview
One marketplace, both sides of the trade
Illustrative utilisation, temperature and VRAM. Flip the switch to see the fleet from either seat.
Renting. By the second, at the price frozen at match. No minimum, no egress.
Illustrative utilisation and temperature · 6 cards
Global marketplace
Compute everywhere you need it
Capacity is pooled across six regions and routed to wherever it's cheapest and closest to your data. Automatically.
6 regions7 GPU classesper second billing
- US
- US
- IE
- DE
- NL
- SG
FIG 02 region mesh
Region mesh
6 live
Illustrative routing preview
Why Kracht
The marketplace advantage
Kracht gives AI teams the speed and scale to take models from concept to production. Without the lock-in or the markup.
Up to 68% lower cost
Market-driven pricing puts idle capacity to work, often a fraction of hyperscaler rates for the same silicon. You rent hardware that already exists.
€0.22lowest GPU-hr in the catalogue
85%
Provider share
you keep 85% of what your card earns
Per-second billing
No minimums, no egress.
Instant capacity
Nothing to reserve or queue.
7
GPU classes
6
Regions
Earn from idle hardware
List the GPUs you are not using and turn depreciating silicon into a steady, automatic income stream.
computed · from this repository
No lock-in
Nothing here holds you
The strongest thing about renting somebody else's idle card is how little of it you adopt. Four ways out, all true today.
Your image
Bring your own container or use ours. Nothing here needs a Kracht-shaped build.
Your key
Launch with your own SSH key and connect straight to the machine. No proprietary shell in between.
Any framework
CUDA, Docker, PyTorch and vLLM run unchanged. A rented GPU is a machine, not a platform.
One command out
Stop a rental and the meter stops with it, to the second. No term, no minimum, and nothing to cancel first.
Developer experience
One platform, three surfaces
Console, terminal, or raw HTTP. The same marketplace underneath. Try the real shell right here.
you@laptop. Drag or arrow-key meArrow keys move the terminal window, shift moves it further, Home recentres it.demo
kracht v0.1.1 · you@laptop
run `kracht login` to start, or `help` to see every command.
you@laptop $
Recently shipped
- Aug 2026
Three products, and a site that says so
- newProducts live under /products/*: compute, studio, autopilot, and the two audience pages beneath compute. Every old address still answers, permanently redirected.
- newTwo capability pages for Studio: observability and evaluation, describing console surfaces that exist on example data.
- Aug 2026
Renters can log in, and the console stops lying about it
- fixedRenters can actually log in to a rental over SSH (#199).
- fixedThe console no longer hangs on loading skeletons once a session goes stale, which had left a renter unable to see spend or press Stop while metering ran (#205).
- Jul 2026
Isolation hardening, and a design pass with the claims checked
- securitySessions can no longer reach the provider's own machine (#183, #184).
- securityA guard that fails the build on the next database view which bypasses RLS (#171).
Two sides, one marketplace
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