Ubuntu Server Setup for GPU Hosting: A Host's Checklist

Most GPU host problems trace back to the first hour of setup: a desktop install where a server belonged, a driver from the wrong source, or a network that only works from inside the house. This checklist covers the OS, drivers, Docker, and networking decisions that decide whether your rig earns quietly for months or pages you at 3 AM.

Start with Ubuntu Server, not Desktop

Install Ubuntu Server 22.04 LTS. The Desktop edition drags in a display manager, a compositor, and a tray full of background services that eat RAM and occasionally grab the GPU. A rental machine needs none of it. Server gives you a clean SSH box where the only thing touching your cards is the renter's workload.

Before anything else, bring the base system current:

sudo apt update && sudo apt upgrade -y

Do this once, deliberately, at install time. Then turn automatic updates off. A background upgrade that swaps your kernel under a running rental is the single most common self-inflicted outage in GPU hosting. The full reasoning and the exact commands are in our guide on disabling unattended upgrades.

NVIDIA drivers: two paths depending on your card

For anything up to the RTX 40 series, the Ubuntu driver tooling does the job:

ubuntu-drivers devices
sudo ubuntu-drivers autoinstall
sudo reboot

The first command lists available drivers and marks one as recommended. If you would rather pin a version, install it directly with sudo apt install nvidia-driver-575 (or 535 for older cards). Details and version notes are in the NVIDIA driver install guide in our open source Linux tools collection.

The RTX 5090 is a different animal. On Ubuntu 22.04 it needs driver 575.64.05 or newer from NVIDIA's own installer, plus build-essential and, on some kernels, GCC 12 to compile the modules. Run the installer with the DKMS option so the modules rebuild automatically after kernel updates. The step-by-step walkthrough, including the GCC fix, is in the RTX 5090 on Ubuntu 22.04 guide.

Either way, the finish line is the same: nvidia-smi shows every card with the full VRAM count. A 5090 should report 32,607 MiB. If a card is missing here, it will be missing on the marketplace too.

Docker and the NVIDIA container toolkit

Renters run inside containers, so Docker plus the NVIDIA container toolkit is the layer that actually earns you money. Install Docker with the official convenience script:

curl -fsSL https://get.docker.com/ -o get-docker.sh
sudo sh get-docker.sh
rm get-docker.sh

Then add NVIDIA's package repository, install nvidia-container-toolkit, point the Docker runtime at it, and restart Docker. The exact copy-paste block, including the repository setup and the GPU test container that proves the whole chain works, is in the Docker + NVIDIA toolkit guide. The test matters: a container that can run nvidia-smi and see your cards is the same path every renter workload takes.

Networking: make the machine reachable from outside

Renters connect from the open internet, so your machine needs a stable address and open ports. Three things to lock down:

1Give the machine a static LAN IP, either in your router's DHCP reservation table or in netplan. A host that changes address after a power cut silently breaks its own port forwards.
2Forward a port range from your router to the machine. Vast.ai assigns renters ports from the range you expose, so forward a generous block (several hundred ports) rather than a handful.
3Verify from outside your own network. Checking from the LAN proves nothing about NAT. Start a listener with nc -l -p 49200, confirm it locally with ss -tuln | grep 49200, then test the port from an external checker against your public IP.

If a port fails from outside, work from the machine outward: listener, firewall, router, then the marketplace's view of you.

The rest of the checklist

SSH hardening. This machine is now a public server and the login attempts start within hours. Key-only authentication and fail2ban are covered in SSH security for GPU hosts.
Disk layout. Renters download large images and datasets. Put Docker's data directory on your biggest, fastest drive and leave headroom; a full disk takes the machine off the market as surely as a power cut.
Power and cooling. Sustained rental load is nothing like a gaming session. Set a fan curve and consider a power limit before you list, not after the first thermal throttle.

From checklist to first listing

With the OS clean, drivers verified, containers running, and ports reachable, the machine is ready for the marketplace side: installing the Vast.ai software, verification, and your first listing. That half of the journey is covered in how to become a Vast.ai host. Before you go live, give the hardware a real workout with a burn-in test so the first crash happens on your watch, not a renter's.