Metered per second
From the moment it runs to the moment you stop it, in euro. A stopped instance costs nothing.
Every block you need,
billed by the second.
Compute, storage, networking and serving. Take one or all of them, keep them for ninety seconds or a month, and pay for exactly that.
SINGLE-TENANT GPUSNAPSHOT + RESTOREWIREGUARD ROUTE
The fleet
Every card, one command
Pick any card in the catalogue and you launch it the same way: one command, your image, your key, billed per second.
every card, one command
The cookbook
What do you need to run it?
Browse every model against the fleet: what it needs to run, the cheapest card that holds it, its speed, and roughly what a million tokens comes to. No account, no sign-in.
| Model | Params | What it needs to run | VRAM | Fits on | Tok/s | Per 1M tok | Per hour |
|---|---|---|---|---|---|---|---|
Qwen3 235B-A22B Mixture of experts. Needs datacentre memory, then generates at a 22B model's pace. | 235B22B active | at int8 · 8k ctx | 295 GB | A100×44 cards | ~65 | €19.18 | €4.48Find |
Qwen3 32B Dense, and about the largest that fits one 48 GB card at full precision. | 32B | at int8 · 8k ctx | 40 GB | A100fits | ~45 | €6.98 | €1.12Find |
Qwen3 30B-A3B The local sweet spot: 30B of weights to hold, 3.3B of work per token. | 30.5B3.3B active | at int8 · 8k ctx | 38 GB | A100tight | ~330 | €0.66 | €0.78Find |
Qwen3 8B Long context at a size a desktop card holds without quantising. | 8B | at int8 · 8k ctx | 10 GB | RTX 4090fits | ~88 | €0.98 | €0.31Find |
QwQ 32B A reasoning model: it emits far more tokens per answer, so cost per token dominates. | 32B | at int8 · 8k ctx | 40 GB | A100fits | ~45 | €6.98 | €1.12Find |
Gemma 3 27B Dense, long window, and takes images as well as text. | 27B | at int8 · 8k ctx | 34 GB | A100fits | ~40 | €5.37 | €0.78Find |
Mistral Small 3.1 24B Sized to fit a single 32 GB card once quantised, which is what it does. | 24B | at int8 · 8k ctx | 30 GB | A100fits | ~45 | €4.78 | €0.78Find |
Gemma 3 12B The 27B model's smaller sibling. Fits a 16 GB card at full precision. | 12B | at int8 · 8k ctx | 15 GB | RTX 4090fits | ~59 | €1.46 | €0.31Find |
Gemma 3 4B Small enough that the context cache, not the weights, is the larger half. | 4B | at int8 · 8k ctx | 5.0 GB | RTX 4090fits | ~176 | €0.49 | €0.31Find |
DeepSeek R1 Mixture of experts, and the clearest case for two counts: 671B has to fit, 37B runs. | 671B37B active | at int8 · 8k ctx | 841 GB | H200×66 cards | ~91 | €68.64 | €22.44Find |
57 models · 57 run on a card in the catalogue at int8, 8k context. Cheapest serving card, largest saving first. Estimates that err upward, not measurements.
1 / 6
compute · rental · example
Launched on a matched cardH100 80GB · price frozenus-east
Streaming usageusage_events seq 413 · metered
Metered per secondstopped costs nothing
Stop when you're donegiving the card back is the same action
This rental · metered per second · example
H100 · 14m 22s elapsed€0.69
idle in that window€0.00
billed so far€0.69
Settles as · sample rows
Last 60 seconds · one tick a second
€0.000803 each
Rate
frozen at match
Split
85 / 15
One compute layer
A GPU market that bills like a meter
Rent a card by the second, launch it with your own key, and stop paying the moment you stop. Whatever providers have online, at today's price. Follow one rental end to end.
Your image or ours
Launch a notebook or a container on somebody's idle card, with your own SSH key.
Any region, no reservation
Whatever providers have online right now. A machine already rented is not offered to you.
Frozen at match
The rate is locked onto the rental, so a provider re-pricing never changes what you pay.
Every metered second splits
exact integer, no cent leaks
Metered per second
What a job costs, per second vs per hour
Metered per second, a job pays for the minutes it ran. Rounded up to the hour, it pays for the rest of the hour too, and the waste is worst just past each one.
[€] a 41-minute job, per second€1.97
[↑] the same job, billed hourly€2.89
// X-AXIS: JOB LENGTH// Y-AXIS: EUR · 1 ■ = €0.48
- A 7m job on an H100: €0.34 per second, €2.89 billed hourly.
- A 20m job on an H100: €0.96 per second, €2.89 billed hourly.
- A 41m job on an H100: €1.97 per second, €2.89 billed hourly.
- A 52m job on an H100: €2.50 per second, €2.89 billed hourly.
- A 1h10 job on an H100: €3.37 per second, €5.78 billed hourly.
- A 1h34 job on an H100: €4.53 per second, €5.78 billed hourly.
- A 2h04 job on an H100: €5.97 per second, €8.67 billed hourly.
- A 2h30 job on an H100: €7.22 per second, €8.67 billed hourly.
- A 2h52 job on an H100: €8.28 per second, €8.67 billed hourly.
per secondwhat hourly rounding adds
illustrative · list rate × job length, from the one price list · 1 square is 10 minutes of the card
The fleet
Rent the latest accelerators
From flagship to cost-efficient consumer cards, every GPU is metered by the second with marketplace pricing.
7 classes / metered by the second
~ nvidia-smi
$ nvidia-smiCUDA 12.4
GPU 0 · NVIDIA H100 80GBrunning
util
87%
mem
70/80G
temp 64°Cpower 420W/700W
example output · your own shell, your own numbers
NVIDIA H100
Datacenter GPU · HBM memory
€2.89 /hr
show
6 of 6
NVIDIA H200
latestTop of the stack for the biggest models.
141 GBVRAM
vs. fleet max94%
- Best for
- large-context
- HBM3e / NVLink
- 1,979 FP16
- low stock
€3.74/hr
Rent an NVIDIA H200 €29.92 for 8 hours
NVIDIA H100
hotBuilt for serious training runs.
80 GBVRAM
vs. fleet max82%
- Best for
- training-ready
- SXM5 / NVLink
- 1,513 FP16
- available
€2.89/hr
Rent an NVIDIA H100 €23.12 for 8 hours
NVIDIA A100
The dependable workhorse.
80 GBVRAM
vs. fleet max76%
- Best for
- steady batch
- SXM4 / PCIe
- 624 FP16
- available
€1.12/hr
Rent an NVIDIA A100 €8.96 for 8 hours
NVIDIA L40S
Great value for inference and media.
48 GBVRAM
vs. fleet max58%
- Best for
- inference lane
- Ada / PCIe
- 733 FP16
- available
€0.64/hr
Rent an NVIDIA L40S €5.12 for 8 hours
NVIDIA RTX 4090
Fast, flexible, surprisingly capable.
24 GBVRAM
vs. fleet max42%
- Best for
- flex node
- consumer
- 330 FP16
- available
€0.31/hr
Rent an NVIDIA RTX 4090 €2.48 for 8 hours
NVIDIA A10
Lean and cheap for steady inference.
24 GBVRAM
vs. fleet max34%
- Best for
- serve-ready
- inference
- 125 FP16
- available
€1.12/hr
Rent an NVIDIA A10 €8.96 for 8 hours
Spec sheet
| GPU | Architecture | VRAM | Bandwidth | Interconnect | € / GPU-hour | Source |
|---|---|---|---|---|---|---|
| NVIDIA H200 | Hopper | 141 GB HBM3e | 4.8 TB/s | NVLink 4 | €3.74 | nvidia.com |
| NVIDIA H100 | Hopper | 80 GB HBM3 | 3.35 TB/s | NVLink 4 | €2.89 | nvidia.com |
| NVIDIA A100 | Ampere | 80 GB HBM2e | 2.0 TB/s | NVLink 3 | €1.12 | nvidia.com |
| NVIDIA A100 | Ampere | 40 GB HBM2e | 1.6 TB/s | NVLink 3 | €0.78 | nvidia.com |
| NVIDIA L40S | Ada Lovelace | 48 GB GDDR6 | 864 GB/s | PCIe 4 | €0.64 | nvidia.com |
| NVIDIA RTX 4090 | Ada Lovelace | 24 GB GDDR6X | 1.0 TB/s | PCIe 4 | €0.31 | nvidia.com |
| NVIDIA A10 | Ampere | 24 GB GDDR6 | 600 GB/s | PCIe 4 | €0.22 | nvidia.com |
In practice
What people do with Compute
Renting a card by the second changes what you can afford to try. Each of these is a job it is built for. Open one to see how it holds up.
Take an H100 for the length of a training run and give it back. No reservation, no minimum, and no card sitting idle between experiments.
- per-second
- no reservation, no minimum
- give it back
- idle between runs costs nothing
One console
Launch and watch every job stream
Provision GPUs, stream deploy logs, and monitor utilisation. Switch between instances without leaving the page.
A rental
Kracht Console
Instances
↑ click an instance to switch. Preview data
Two sides of the same market
One of them is probably you. Both run on the same ledger, the same regions and the same per-second meter.
FIG 01 the marketplace
renter cli · kracht launch
The renter's own machine. One command becomes a signed request naming a GPU model and region, never a machine.
Earn from idle
What could your cards earn?
You set the price, you keep 85%, Kracht takes 15% and handles the renter, the metering and the payout. Drag to your setup.
Number of cards4
Your listed price
€/ GPU-hour
● in the band: priced to rent
Assumed utilisation65%
How much of the month a card is actually rented.
Electricity
Hardware costoptional
€/ card
What you paid for a card, to see payback time.
Your take-home, per month
€4,662/mo
After the 15% Kracht fee, on 4 cards at 65% utilisation.
take-homeKracht 15%
Gross rental 4×H100 80GB @ 65%€5,485
Kracht fee, 15% €823
Your earnings, 85%€4,662/mo
≈ €1,166 per card · €55,949 a year across 4 cards.
Cumulative take-home€55,949 in year one
Bars are illustrative monthly take-home; it moves with demand.
Every primitive, wired together
Instances, storage and keys are one account and one meter, not four products you join up yourself.
One meter
Every primitive bills against the same per-second clock.
One ledger
Append-only entries, unique on an idempotency key.
One account
Instances, volumes and keys hang off the billable account.
Start where it runs
Rent a card in one line
No reservation, no minimum, your own SSH key. This is the real path, not a preview: Compute runs end to end, metered per second, in euro.
Launch a GPU
$ kracht login
Paste an API key from https://kracht.ai/app/settings/keys:
Signed in as Ada Lab (balance 20.000000 EUR)
Key saved to ~/.config/kracht/config.json
$ 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...
root@kracht:~# nvidia-smi -L
GPU 0: NVIDIA H100 80GB HBM3
- €2.89
- h100, per hour
- 85%
- to the provider
- €0
- to hold nothing
Resources
Where to go next
Compute is the part of the marketplace that runs today. These are the three places worth going once you have seen how it launches.
Documentation
How it works, endpoint by endpoint
Pricing
Every rate, and what a month of one card costs
Browse live capacity
What is actually available, at today's prices
Changelog
What shipped, dated, including what did not
Status
What is up, and what was not
Help centre
The questions that come up more than once
market · rates
Per GPU-hour, in euro
youmarkethostlisted
- H200 141 GB€3.74
- H100 80 GB€2.89
- A100 80 GB€1.12
- A100 40 GB€0.78
Frequently
asked questions
It runs. Rentals have been placed, metered and settled end to end on real hardware, in euro, with the split applied. The parts of Kracht that are not finished are Studio and Autopilot, and each of those pages says so at the top.
Continue through the platform
The next part of the same job
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