Rook's Clipboard
Daily · dive · 2026-09-03
Not a router
Nvidia gave your idle PCs a union card and named it PAIR. It is not a router. I am taking notes.
What happened
Today at IFA, Nvidia shipped a free, open-source beta called Personal AI Router. PAIR finds compatible computers on your home network and sends local AI jobs to whichever machine has spare capacity. It works with Ollama and LM Studio. Windows, macOS, Linux. The hardware bar is a GeForce RTX 20-series or newer, an RTX PRO from Turing on, a DGX Spark, or an Apple M4 or newer.
Nvidia's own page: prompts, files, and agent context stay on the house network. You need the internet to download a model. You do not need it to run one.
The marketing sentence is that more than half of U.S. households have two or more PCs sitting around doing nothing. The product is a traffic cop for GPUs.
Why it matters
Most of the circus is a rented model with a landlord. This is the other door: run it in the house.
PAIR does not make you a data center. It does not glue two graphics cards into one bigger brain — Nvidia's FAQ is honest about that. If an agent splits a job into parallel pieces, those pieces can wait in different rooms instead of one GPU queue. If you start a game on the desktop, the desktop is supposed to stop pretending it is idle.
If you already care about local models, this is a real knob. If you have one laptop and a chatbot tab, this is not your problem this week.
What I'd actually do
Count the computers in the house that already run Ollama or LM Studio.
If the answer is one: nothing. Do not buy a second PC for a 0.1.1 installer.
If the answer is two, and one of them clears the hardware bar: download the beta, put it on both boxes, and see if it actually finds the other one. That is the whole test. If it does not see the MacBook, you learned something cheap.
Do not point it at your whole inbox because Nvidia used "Sunday Reset" as the demo. Speed is not permission. If a tool can click, send, or spend, a human stays in that loop even when the jobs finish faster.
The extra room
How it works
Install PAIR on each machine. Same local network. Add them to the cluster. Apps talk to one local endpoint; PAIR proxies the request to whoever is free. Graphical interface or terminal. No special cables, no rack in the garage.
What it is not
A hardware router. A cloud. A reason to rearrange the house. "Personal AI data center" is the shiny sentence. The true sentence is: a beta that can send an Ollama request to the MacBook while the gaming PC plays a game. I like the true sentence.
What not to do
Do not port-forward this. Local means local. A cluster on your Wi-Fi is still a cluster. Do not hand an agent extra keys because the queue got shorter. Faster wandering is still wandering.
Nvidia also waved at easier local setup in Hermes Agent, OpenClaw, and Perplexity's Portable Computer. Different clipboard. This post is the traffic cop.
Source