Launch your first instance

A guided walkthrough of the Create Instance flow.

This guide assumes you’ve signed up and topped up your wallet. If not, start with Quickstart.

Step 1 — pick a GPU

Open gpuoutlet.ai/app/create. You see six cards, sorted cheapest → most expensive:

GPU	VRAM	Arch	From $/hr	Counts
RTX 4090	24 GB	Ada	0.65	1, 2, 4, 8
RTX 5090	32 GB	Blackwell	1.15	1, 2, 4, 8
A100 80GB	80 GB	Ampere	1.40	1, 2, 4, 8
H100 80GB	80 GB	Hopper	2.80	1, 2, 4, 8
B200	192 GB	Blackwell	6.00	1, 2, 4, 8
B300	288 GB	Blackwell	7.50	1, 2, 4, 8

Click any GPU’s count pill (1×, 2×, etc.) to select it. The footer at the bottom updates with the live price.

Greyed-out pills (e.g. 8× on B300) mean no provider has that count available right now. We refresh availability every 30 seconds — re-check shortly if you need that exact config.

Step 2 — pick a template

Click Continue. The template grid loads:

• Ubuntu 24.04 — bare OS + CUDA drivers
• PyTorch — Ubuntu + PyTorch 2.x + transformers
• vLLM — inference server, pass --model at boot
• Ollama — daemon on port 11434
• A1111 — Stable Diffusion WebUI on port 7860
• text-generation-webui — Oobabooga on port 7860

Pick one. The footer updates with the per-hour cost based on your GPU choice

• the template (templates with preloaded models can be slightly more expensive due to disk image size).

Step 3 — launch

Click Launch. A modal appears:

🟢 Provisioning… RTX 4090 × 1 · Ubuntu 24.04 US-East · started 4s ago

ssh root@pending-host… -p 22001

20–60 seconds later the dot turns green and provisioning becomes running. The SSH command is now valid.

Step 4 — connect

Copy the SSH command shown. From your terminal:

ssh root@123.45.67.89 -p 22001
# Welcome to Ubuntu 24.04 (GNU/Linux 6.8.0-31-generic x86_64)
root@gpu-pod-01:~#

Verify the GPU:

nvidia-smi
# +---------------------------------------------------------------+
# | NVIDIA-SMI 550.54.14    Driver Version: 550.54.14    CUDA Version: 12.4 |
# |---+----------------------+-----------+---------+-----------|
# | 0 | NVIDIA RTX 4090       | 00000000:01 | 0%      | 35C       |
# +---------------------------------------------------------------+

Step 5 — work

The pod is yours until you stop it. Common patterns:

Quick PyTorch sanity check

python -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"
# True NVIDIA GeForce RTX 4090

Get your code in

git clone https://github.com/you/your-repo.git
cd your-repo && pip install -e .

Or rsync from local

# (from your laptop, in another terminal)
rsync -avz ./project root@123.45.67.89:/root/ -e "ssh -p 22001"

Step 6 — stop

Back in the dashboard, Instances → click the pod → Stop. The pod is unreachable within ~5 seconds. The final partial second is settled and you can see the total cost in the ledger.

Stopping wipes the pod’s local disk. Push your work somewhere persistent (S3, HuggingFace, your laptop via scp) BEFORE stopping.

What if I just close my laptop?

The pod keeps running. The meter keeps ticking. Don’t do this if you don’t want to burn through your wallet overnight. Either:

• Set an auto-stop when launching (Stop after N hours), or
• Lower your wallet balance to roughly the budget you’re OK with (when it hits $0, every running pod stops within 5 seconds)