Sam Witteveen · 2026-09-06 · notable
Sam Witteveen — 'NVIDIA Doubles Down on Local AI With PAIR'
Sam Witteveen walks through NVIDIA PAIR, the Personal AI Router NVIDIA announced at IFA 2026 on 3 September. The Apache-2.0 tool finds the other PCs on your network and sends each local inference request to whichever machine has a free GPU.
A walkthrough of NVIDIA's new router that spreads local model calls across every PC in the house.
What is it?
NVIDIA PAIR — short for Personal AI Router — is the subject of this Sam Witteveen video. NVIDIA announced the free, open-source tool at IFA 2026 on 3 September 2026 and publishes it as NVIDIA/Personal-AI-Router under Apache-2.0. It is a public beta written in Go, with installers for Windows 11, Linux and macOS on both x64 and arm64, and it works with GeForce RTX 20 Series and newer, RTX PRO workstation cards, DGX Spark and Apple M4 or newer silicon.
How does it work?
PAIR runs no models of its own. It discovers compatible machines on the local network, and you pair them by typing a six-digit PIN shown on the first machine. Requests then arrive at an Ollama- or OpenAI-compatible endpoint, and the scheduler hands each one to a single node based on which systems have capacity. Mixed clusters are fine — a Windows box and a Mac can pair — and the supported inference engines are Ollama and LM Studio.
Why does it matter?
An agent run splits one task into several jobs that could go side by side, but on a single laptop they queue behind one GPU. NVIDIA's own example is an assistant asked to sort a cluttered inbox: it fans the work out to subagents, and PAIR spreads those jobs over the PCs already sitting idle on the network. Because PAIR speaks the Ollama and OpenAI-compatible APIs the tools already use, agent code does not have to change to pick up the extra capacity.
Who is it for?
people running local models on more than one machine
Try it
https://github.com/NVIDIA/Personal-AI-Router