RunKeyRunKey
A4000 · 3090 · 4090 · 5090 — from 0.0127 credits/sec

Rent the whole GPU.Pay by the second.

A dedicated GPU with CUDA ready on it, up in about a minute, root over Direct SSH, and spending stops the moment you terminate it. A fixed credit rate per second, no minimums, no subscription.

Deploy an InstanceSee Pricing
~60s
Typical time to running
Per second
Credits, no minimums
Direct SSH
Root on the real GPU

Four cards. One rate each.

Every instance of a given model spends the same credits per second, whichever node it lands on. Pick the card that fits the workload, then pick the node that fits the job.

RTX A4000

16 GB · 448 GB/s · 19.2 TFLOPS FP16

0.0127 credits/sec
45.71 credits per hour

The efficient workhorse. Ideal for SD 1.5 and SDXL at moderate batch sizes, LoRA training and always-on inference where cost per hour matters more than raw speed.

  • ›SD 1.5 / SDXL
  • ›LoRA training
  • ›Long-running inference
Deploy RTX A4000

RTX 3090

24 GB · 936 GB/s · 35.6 TFLOPS FP16

0.0198 credits/sec
71.43 credits per hour

24 GB of VRAM at the lowest price on the platform. A dependable choice for SDXL, Flux at reduced precision, and 13B-class model inference.

  • ›SDXL at scale
  • ›Flux (quantised)
  • ›13B inference
Deploy RTX 3090

RTX 4090

24 GB · 1.01 TB/s · 82.6 TFLOPS FP16

0.0294 credits/sec
105.71 credits per hour

The default pick. Ada-generation throughput makes Flux, SDXL and video pipelines roughly twice as fast as a 3090 for well under twice the price.

  • ›Flux at full precision
  • ›Video pipelines
  • ›Fast iteration
Deploy RTX 4090

RTX 5090

32 GB · 1.79 TB/s · 104.8 TFLOPS FP16

0.0393 credits/sec
141.43 credits per hour

Blackwell, 32 GB, and the fastest memory bandwidth we rent. For long video generation, high-resolution batches and workflows that will not fit in 24 GB.

  • ›Long video generation
  • ›High-res batches
  • ›30B+ inference
Deploy RTX 5090

From nothing to a running GPU

Four steps, and the slow one is the image pull.

1

Add a key and some credits

Paste your SSH public key and buy a credit pack. Nothing recurring, and credits never expire.

2

Pick a GPU and a node

Choose the card, then the specific node it runs on — vCPU, RAM, disk, network and region are all on the table before you commit.

3

Start the instance

Ubuntu comes up with CUDA and cuDNN matched to the node's driver. Typically running in under a minute.

4

Connect over Direct SSH

Copy the connection line from the instance page. Direct SSH, not a proxy, so uploads and downloads run at the node's full speed.

The environment

CUDA, ready on boot

Ubuntu with the NVIDIA stack already in place — CUDA and cuDNN, built to match the driver the node is actually running, on every card from Ampere to Blackwell. Nothing else is installed and nothing is listening: you land as root and the environment is yours.

  • CUDA and cuDNN matched to the node's driver
  • Root over Direct SSH, no proxy in between
  • A desktop panel in the browser — terminal, files, port forwarding, monitoring
  • 80 GB at /workspace for models and outputs
  • Install anything — ComfyUI, a training stack, your own image
Deploy an Instance
once you are connected
# the driver is already there
nvidia-smi --query-gpu=name,memory.total --format=csv,noheader
NVIDIA GeForce RTX 4090, 24564 MiB

# so is CUDA — build or install whatever you need
pip install torch && python -c "import torch; print(torch.cuda.is_available())"
True

# move your own files in
scp -P 41277 sdxl.safetensors root@198.51.100.24:/workspace/

# and when you are done
# hit Terminate in the console — spending stops there

Built to be predictable

Renting compute should be boring. These are the parts we refused to make interesting.

Direct SSH, never a proxy

We only take capacity from nodes that publish real ports, so your connection is a straight line to your instance. Copy one command from the instance page and you are in.

Metered by the second

The meter starts when the instance reports running and stops the instant you terminate it. There are no minimums, no rounding up to the hour and no idle charges.

One fixed rate per GPU

An RTX 4090 spends the same credits per second whichever node you land on. You pick a node on its specs, not on a shifting spot price.

CUDA ready on boot

The image carries CUDA and cuDNN, and picks the build that matches the node's driver. No image building, no driver archaeology — nvidia-smi works the moment you are in.

Capacity worth renting

Every listed node clears a reliability, bandwidth and open-port bar before it reaches you. The ones that do not never appear in the picker.

Spend that cannot run away

Watch the burn rate and runway in the console. When the credits hit zero instances are released automatically, so a forgotten instance cannot drain an account.

Questions worth asking

The things people actually want to know before renting a GPU they do not own.

How does billing actually work?
You buy credits in a pack, and every instance has a fixed credit rate per second shown before you deploy — 0.0127 for an RTX A4000, 0.0294 for a 4090. We meter from the moment the instance reports running until you terminate it, and draw the credits down as it runs. An instance you run for eleven minutes costs eleven minutes. Your statement groups usage into one line per instance per hour so it stays readable.
What do I get access to?
A whole GPU and the environment around it, as root, over Direct SSH — not a slice of a card and not a queue you submit jobs to. The instance page gives you the exact connection line plus an scp example for moving models and outputs. Nothing runs there but what you put there.
Why is the rate the same for every node of one GPU model?
Because a shifting per-node rate makes it impossible to plan. We hold a fixed credit rate per GPU model and absorb the variation upstream. You choose a node on the things that actually affect your work — cores, RAM, disk, bandwidth, region — rather than on arbitrage.
What happens when my credits run out?
The console shows your burn rate and how many hours of runway you have left, and warns you well before the end. If the balance does reach zero, running instances are terminated automatically so the account can never go deeply negative.
Is my data kept after I terminate an instance?
No. Terminating releases the resources and everything on the instance's disk goes with it. Download anything you want to keep first — the scp command on the instance page is there for exactly that.
What is installed?
Ubuntu, the NVIDIA stack — CUDA and cuDNN, matched to the node's driver — and an SSH server. That is on purpose: you get a clean environment as root and install exactly what your work needs, whether that is ComfyUI, a training stack or your own container.

Your GPU is about a minute away

Add a key, add some credit, pick a card. Terminate whenever — you only pay for the seconds it ran.

Get Started

No subscription · No minimum spend · Terminate anytime