Compute Overview

Grafilab Compute rents you a GPU machine by the hour. You pick the hardware, pick a software template, and Grafilab matches you with an available node from its supplier marketplace. A few minutes later you have a running container with a root shell and, for the web templates, a public URL.

The whole flow is three steps in the console at Compute.

Each rental is a Docker container running on a supplier’s GPU host, not a virtual machine you install yourself.

• A root shell. You log in as root over SSH, by password or by your own key. Everything in the container is yours to change — install packages, pull models, run anything.
• The GPU passed straight through. On a GPU node the card is attached to the container, so nvidia-smi and CUDA work with no extra setup.
• A public web URL. Templates that ship a web UI answer on port 80 at https://<your-slug>.instance.grafilab.ai, where the slug is three hyphenated words picked for your instance. You can publish extra ports yourself with Web Access.
• A prebuilt environment. Templates such as Open WebUI, JupyterLab, ComfyUI, code-server, and PyTorch save you the image build. You can also point a rental at your own image.
• 2 GB of shared memory (/dev/shm), which is what most PyTorch data loaders need.

See Templates for what each environment includes.

Step 1 of the wizard, Choose Your Power, shows a card per GPU family that suppliers currently have listed:

Card	Positioning	Memory
NVIDIA H100	Enterprise AI — LLM training	80 GB HBM3
NVIDIA RTX 5090	Ultra AI performance — Blackwell	32 GB VRAM
NVIDIA RTX 5080	Next-gen compute — high-efficiency AI	16 GB VRAM
NVIDIA RTX 4090	Consumer flagship — deep-learning workhorse	24 GB GDDR6X
NVIDIA L4	Enterprise inference — Ada Lovelace tensor	24 GB VRAM
NVIDIA RTX 3090	Value performance — general compute	24 GB GDDR6X
CPU	Web and microservices	No VRAM

The row is generated from live inventory, so a family only appears while a supplier has a node listed for it. If you don’t see the card you want, check again later or pick a neighbouring family.

Three cards carry a diagonal Contact Us ribbon instead of a price — NVIDIA B200 (180 GB HBM3e), NVIDIA L20N 72GB, and NVIDIA L20N 48GB. These are arranged with the sales team; the ribbon disappears automatically once self-serve inventory for that GPU exists.

Compute is a marketplace: the same GPU is offered by different suppliers at different rates. Step 3 of the wizard lists the matching nodes cheapest first, and each row tells you who you’d be renting from:

• Reputation — a coloured badge reading excellent or good, based on the supplier’s track record.
• DC Verified — a badge shown when the node is a verified data-centre host rather than a home machine.
• Location — the node’s country, which also matters for billing (some listings only accept renters in the same country).
• CPU model and clock speed, next to the GPU name.
• A spec grid with exactly four cells: vCPU, RAM, VRAM, Storage. Expand it with More.

Compute rentals are prepaid for a fixed duration, not metered while you run:

• The node’s rate is per hour. Your total is that rate multiplied by the rental period you choose, and it is charged in full at checkout.
• The minimum rental is 12 hours. Longer periods are offered per listing.
• The total must come to at least US$0.50. The check runs on the US-dollar total before conversion to your wallet currency, so a short cheap rental can be refused even when the converted figure looks larger.
• Pay from your credit balance or by card. The checkout shows your Balance: next to the credit option.
• Your profile needs a country set before you can rent, because that is what stamps your wallet currency.
• To keep an instance past its expiry, use Renew on the instance card before Expire at passes.

There is no early stop and no partial refund: once you launch, the whole period is paid for. Pick the shortest duration that covers your work and renew if you need more.

Every charge, including renewals, lands as a row on Billing → Invoice History. See Rentals & Billing History.