NVIDIA Pools the Idle Macs and PCs in Your House into One Inference Pool: PAIR Ships as an Open-Source Beta — but It Does Not Pool VRAM

On September 3, during IFA, NVIDIA released the Personal AI Router (PAIR) beta, free open-source software (Apache-2.0, written in Go, repo at NVIDIA/Personal-AI-Router). It is not an inference engine itself: it reuses the Ollama or LM Studio install already on each machine and exposes Ollama-compatible and OpenAI-compatible proxy endpoints, so existing agents need no configuration changes. Nodes are found over mDNS on the local network or by IP, pairing uses a six-digit code, and node-to-node traffic runs over mTLS. NVIDIA is explicit about what it does not do: it does not merge GPUs into one accelerator, does not pool VRAM, and cannot split a single inference request across machines. Supported hardware spans GeForce RTX 20 Series and newer, RTX PRO from Turing on, DGX Spark, and Apple M4 or newer silicon, on Windows, macOS and Linux.

It Is a Router, Not an Engine

NVIDIA calls it a "virtual inference router," and the qualifier is worth taking literally. PAIR ships no model and no kernels; the thing running the model is still the Ollama or LM Studio install on your machine. What PAIR does is find which other machines on the same local network can take work, hand a whole request to one of them, and return the result to whoever asked. Scheduling looks at four things: whether the node is up, whether its inference engine is running, whether that model is already on disk there, and how busy its GPU is right now. Because the endpoints it exposes are Ollama- and OpenAI-compatible, nothing changes on the agent side — the agent still decides what to do, and PAIR only decides which machine does it. Start gaming on one of those machines and it drops out of the pool. Security comes in two parts: a six-digit code to confirm pairing by hand, then mTLS with generated certificates between nodes, with prompts and context staying on the local network. The repo is Apache-2.0 and written in Go, with both a graphical and a terminal interface on Windows, macOS and Linux; the floor is 8 GB of RAM with 20 GB of storage recommended.

That 2× in the Demo — How Much of It Is Actually Distribution?

The comparison shown at IFA: a five-subagent workflow on Hermes Desktop with Ollama finished in 8 minutes 48 seconds on a three-device PAIR cluster, against 18 minutes on a single RTX Spark laptop. The numbers themselves are fine, but there is exactly one comparison point, and NVIDIA has not said what the other two machines were. So how much of that 2× comes from requests being distributed and how much is simply a faster machine joining the pool is not something the public information settles. Anyone evaluating this should stand up three machines and measure it again.

When It Is Worth Installing

Rule out one class of need first: it will not let you run bigger models. If you do not have the VRAM for a 70B, three machines will not give it to you — PAIR raises how many requests you can serve at once, not the ceiling on any single one. What it actually fits is subagent fan-out: a lead agent splits off five or six subtasks, each of which is an independent small-model call, and extra machines turn into real throughput. If you have two or three discrete-GPU machines around a home or small office, idle through the workday, already running Ollama or LM Studio, the cost of trying it is basically zero. One machine, or a workload that is mostly about getting one large model running, and this will not help. Perplexity Portable Computer, Hermes Agent and OpenClaw already ship simplified Windows setup for it. It is a beta — don't put it in a production path yet.

via: NVIDIA technical blog, PAIR product page and FAQ, GitHub: NVIDIA/Personal-AI-Router, The New Stack, Neowin