How it works

Renting Compute on w.ai

Renting compute on w.ai provides a dedicated Jupyter Notebook or SSH environment backed by real GPU hardware from our distributed compute network. You gain full access to a sandboxed container, choosing between CUDA for NVIDIA GPUs, ROCm for AMD GPUs, or MLX for Apple Silicon, billed by the minute using w.ai points.

---

Step-by-Step Guide

Browse Available GPUs

Visit the Compute page at https://app.w.ai/compute to see live GPU inventory. Each card displays:

• GPU Model — e.g., NVIDIA RTX 5090, Apple M4 Pro

• VRAM — Total GPU memory available

• Backend — CUDA, ROCm, Metal, or Vulkan

• Price — Cost in w.ai points per hour

• Availability — Number of workers offering this hardware

Use the search bar and category filters (NVIDIA, Apple, AMD, Intel) to find the right GPU for your workload.

Select Duration & Rent

Choose your rental duration from available options:

Duration	Use Case
5–30 minutes	Quick experiments, testing code
1 hour	Standard development sessions
2 hours	Training runs, larger experiments
4–12 hours	Extended compute jobs
24 hours	Long-running training or processing

The UI shows your total cost before confirmation. Ensure you have sufficient w.ai points to cover the duration. Click Rent to confirm.

Wait for Provisioning

Upon confirmation, your rental enters a Pending state as:

1. The network matches a worker with your requested hardware.

2. The worker provisions an isolated container environment.

3. The Jupyter Notebook and SSH server starts and becomes accessible, typically taking under ~1 minute.

Access Your Jupyter Notebook or SSH terminal

Once active, your rental card shows:

• Session URL — Your unique Jupyter Notebook endpoint

• Access Token — Authentication token

• Time Remaining — Live countdown

For SSH, you will see your SSH credentials (username, host, etc.)

Click Details on My Rentals for full access credentials, then Open Jupyter Notebook or Open Web Terminal to launch your environment in a new tab.

---

Your Environment

Each rental provides a fully configured Jupyter Notebook or SSH environment with Python 3.11, pip, and a comprehensive set of pre-installed packages.

Note: The environment is ephemeral. All files are deleted when the session ends. Download any results before your rental expires.

Pre-Installed Python Packages

Linux / Windows (CUDA & ROCm)

Category	Packages
Notebooks	JupyterLab, ipywidgets, tqdm
Data Science	NumPy, Pandas, SciPy, scikit-learn
Visualization	Matplotlib, Seaborn, Pillow
ML / Deep Learning	PyTorch (CUDA/ROCm/CPU), TorchVision, TorchAudio, MLX-CUDA, Unsloth
Hugging Face	Transformers, Datasets, Hugging Face Hub, Accelerate, Diffusers, Safetensors

macOS (Apple Silicon)

Category	Packages
Notebooks	JupyterLab, ipywidgets
Data Science	NumPy, Pandas, SciPy, scikit-learn
Visualization	Matplotlib
ML / Deep Learning	PyTorch (MPS), TorchVision
Hugging Face	Transformers, Diffusers, Accelerate, Safetensors
Apple MLX	MLX, mlx-lm, mlx-vlm, mlx-audio, mlx-video, MFlux

Base Container Images

Each environment is built on a minimal, production-grade base image depending on the backend:

Backend	Base Image
CUDA (NVIDIA)	nvidia/cuda:12.1.1-devel-ubuntu22.04
ROCm (AMD)	rocm/dev-ubuntu-22.04:7.1.1
Vulkan	ubuntu:22.04

All images also include: curl, git, ca-certificates, glib, Python 3.11.14, and uv (fast Python package installer).

Need an additional system package? Contact us to request additions to the base image.

---

Managing Your Rentals

Visit My Rentals in the sidebar to manage all sessions.

Extending a Session

Need more time? Click Extend to add minutes or hours. Extensions charge the same per-minute rate and require a sufficient point balance.

Stopping Early

Click Stop Rental to end a session early. Billing is pro-rated by the minute — you are charged only for the minutes used and remaining reserved points are refunded.

---

Pricing

Rental pricing depends on:

• GPU Performance — Higher-tier GPUs cost more points per hour

• VRAM — More memory increases the base price

• Network Earnings — Prices reflect actual GPU earning rates on the network

A minimum charge of 1 point per minute applies to all sessions. Pro-rated billing refunds unused reserved points if a session ends early. Prices are displayed on each GPU card in the Compute marketplace.

---

Networking Restrictions

For security, outbound network access from rental environments is restricted to an allowlist of trusted domains required for ML development:

Category	Domain
Package Registries	pypi.org
	files.pythonhosted.org
ML / Dataset Platforms	huggingface.co
	hf.co
	kaggle.com
	openml.org
	archive.ics.uci.edu
ML Frameworks	download.pytorch.org
	tensorflow.org
Research Data Repositories	zenodo.org
	figshare.com
	osf.io
	dataverse.harvard.edu
	datadryad.org
	archive.org
Source Code Hosting	github.com
	githubusercontent.com
	gitlab.com
	bitbucket.org
	sourceforge.net
Cloud / CDN Storage	dl.fbaipublicfiles.com
	storage.googleapis.com
	storage.cloud.google.com
	s3.amazonaws.com
	blob.core.windows.net
	cloudfront.net
	r2.cloudflarestorage.com
	download.microsoft.com
File Sharing Services	drive.google.com
	drive.usercontent.google.com
	dropbox.com
	dl.dropboxusercontent.com
Government / Research Labs	nersc.gov
	nasa.gov
	cern.ch
	noaa.gov
	nist.gov
Web-Scale Datasets	commoncrawl.org
	dumps.wikimedia.org
Computer Vision Datasets	image-net.org
	cocodataset.org
	cs.toronto.edu
Biomedical Data	ncbi.nlm.nih.gov
	ebi.ac.uk
	physionet.org
Speech / Audio Data	openslr.org

All other outbound network traffic is strictly blocked. This includes general web browsing, SSH connections, and access to arbitrary external services.

Need access to an additional domain? Contact us on Discord to request additions to the network allowlist.

---

Security

w.ai compute sessions are designed with defense-in-depth isolation to protect both renters and workers.

Linux / Windows

• All Linux capabilities dropped — containers run with the minimal possible privilege set

• Non-root user — system package installation is disabled inside the container

• Network isolation — outbound traffic restricted to the allowlist above; all other traffic denied

• Resource limits — CPU, memory, and ulimits are enforced to prevent resource abuse

• Container filesystem is ephemeral and fully destroyed on session termination

macOS (Apple Silicon)

• Network isolation — very limited networking restricted to the allowlist above

• Locked-down sandbox — all system access is blocked outside the sandbox directory

• Restricted shell — only a very limited set of shell commands are enabled; all others are blocked

• Minimal kernel access — only essential kernel services are available to the sandbox environment

---

Worker Requirements

To contribute your machine as a compute rental worker, you must meet these minimum requirements:

All Platforms

• 20 GB available storage — required for container images, models, and session data

Linux

• System packages: uidmap, iptables

• NVIDIA GPUs additionally require: nvidia-container-toolkit

Windows

• NVIDIA GPU required

• WSL (Windows Subsystem for Linux) must be installed

  • The w.ai worker will attempt to install WSL automatically if missing — a system reboot is required to complete setup

macOS

• Apple Silicon (M1 or later) required

• No additional system packages needed

Running the Worker

Start the worker in the foreground:

wai run w.ai

Or run in the background (detached mode):

wai run w.ai -d

Manage background workers with:

wai list              # List running instances
wai logs -f           # Follow live logs
wai stop <name>       # Stop a worker (or "wai stop all")
wai restart <name>    # Restart a worker

Workers earn w.ai points for every minute a rental session is active on their hardware.