Renting a GPU

Choosing a GPU

Each GPU model has one fixed hourly price. The Rent a GPU page shows how many GPUs of each model are free right now and the largest number available on a single server — that is the maximum you can request in one server.

Number of GPUs. A server can have 1 to 8 GPUs, all on the same machine. The price is multiplied by the number of GPUs. See Multi-GPU for how several GPUs work together.
CPU and RAM are shared in proportion to the GPUs you rent: 1 GPU out of 4 on a machine gets a quarter of its CPU cores and of 90% of its RAM (at least 1 core and 2 GB). Shared memory (/dev/shm) is half of your RAM, up to 32 GB.
The platform picks the best matching machine for you: higher host level first, then higher reliability, then the tightest fit.

A template or model can also require a minimum CUDA version (the host’s NVIDIA driver). Machines with an older driver are skipped automatically.

Templates and models

A template is a ready Docker image with the right command, ports and settings — see the Templates reference. Some templates serve a model (vLLM, SGLang, embeddings require one; ComfyUI can optionally start with one). Only models whose total VRAM requirement fits the selected GPUs are offered — see Models.

With the Custom Docker image template you can run any public image (for example ghcr.io/me/app:1.2). Its default entrypoint and command are used.

Disk

Each template has a default disk size (20–100 GB). You can set 10–2000 GB. The disk is your persistent /workspace directory: it survives restarts and stopping.

Model weights are not counted in your disk: models are mounted read-only at /models/<model-id>.

Environment variables

You can add up to 50 variables (names in UPPER_CASE, values up to 4 KB). Common examples:

Variable	Why
`HF_TOKEN`	Your Hugging Face token. Required for gated models (Llama, Gemma).
`WANDB_API_KEY`	Logging training runs to Weights & Biases.

Values of variables whose name contains TOKEN, KEY, SECRET or PASSWORD are masked in the dashboard and the API.

The platform always sets:

Variable	Value
`GYSGA_INSTANCE_ID`	The server id, e.g. `i-abc12345`.
`GYSGA_TOKEN`	A random token for web interfaces (Jupyter uses it as its login token). Shown in the connection panel.

Lifecycle

Status	Meaning	Billed
`creating`	GPUs are reserved, waiting for the host machine.	no
`pulling`	The host downloads the Docker image and the model (progress is shown).	no
`starting`	The container is being created or restarted.	no
`running`	The container runs.	yes
`stopping`	Stop requested.	no
`stopped`	Container stopped, GPUs released, data kept.	no
`unreachable`	The host machine went offline.	no
`error`	Something failed (for example the image does not exist). The message explains why.	no
`destroying` → `destroyed`	The container and its data are deleted.	no

Every change is written to the server’s event log, visible in the dashboard and via the API.

Billing

You pay per second, only while the server is running.
If the host goes offline, billing pauses (unreachable) and resumes only when the server runs again. Offline time is never charged.
When your balance reaches zero, the server is stopped automatically (not deleted). Add funds and start it again.
Launching or starting a server requires a balance that covers at least one hour.
Prices of running servers are fixed at launch; price changes apply to new servers only.

Stop, start and destroy

Stop — stops billing immediately. The data in /workspace stays on the host machine.
Start — starts the server again on the same machine (that’s where the data is). If its GPUs were taken by someone else in the meantime, you’ll get “the GPUs of this machine are busy”; try later or create a new server.
Stopped servers are destroyed automatically after 7 days.
Destroy — deletes the container and /workspace. This cannot be undone.