vibescoder
cd ..

NVIDIA PAIR Looks Promising for Multi-User Homelabs

·5 min read
homelabaillmagentsopen-sourcebuilding-in-public

I opened NVIDIA’s developer blog on September 3rd expecting another driver update and found a router instead. Not a network router. A router for inference requests, built to sit between my agents and whichever GPU in the house happens to be free. I read it twice, then went back to my own posts from the last few weeks, because I already knew the exact problem it claims to solve. I wrote about that problem myself, twice.

PAIR turns Ollama and LM Studio into one address

NVIDIA Personal AI Router, PAIR for short, is not a new inference engine. Ollama or LM Studio still does the actual work. PAIR is the layer in front of them. Install it on every machine you want to contribute, pair those machines together over your local network, and PAIR gives you one address on the machine in front of you. Point an agent at that address the same way you always have. PAIR decides, request by request, which paired machine actually serves it.

The mechanics are straightforward once you separate them from the marketing. Discovery runs over mDNS. Trust is a deliberate step, a six digit PIN typed on the machine you are inviting, and every byte between paired nodes travels over mutual TLS afterward. A node becomes eligible for a given request only when it is online, running a supported engine, and already holding the exact model requested. The scheduler then picks one eligible node based on current load, and that node runs the whole request start to finish. PAIR does not merge GPUs, does not pool VRAM, and does not split one model or one request across machines. It is deliberately narrow. It only decides where the next independent job goes.

The five-subagent demo is my own bottleneck

NVIDIA’s own demo is what sold me. They ran Hermes Desktop against a synthetic household task, split five ways into subagents, entirely on Ollama. On one RTX Spark laptop the run took 18 minutes. Add a DGX Spark and an RTX 5090 to a PAIR cluster, same task, same subagent count, and it finished in 8 minutes 48 seconds.

I have been circling this exact shape of problem since I wrote about why two agents beat one on this homelab. My own conclusion was architectural: coding agents are ephemeral and invoke driven, while home automation agents like OpenClaw are persistent and event driven. What I did not have an answer for was what happens when either kind spins up several subagents at once and they all queue behind the same GPU. PAIR is a direct answer to that specific gap, and it works with the exact engines I would already be running, no new API for an agent harness to learn.

Strix Halo just lost its shot at the third box

I have been debating a third machine for weeks, a dedicated box sized to run something like DeepSeek V4 Flash. The candidates were NVIDIA’s DGX Spark, an Apple Mac Studio, and AMD’s Ryzen AI Max+ 395, better known as Strix Halo. PAIR’s supported hardware list settles part of that debate before I spend a dollar. It covers NVIDIA GeForce RTX 20 series and newer, RTX PRO workstation GPUs, DGX Spark, and Apple silicon starting at M4. AMD is not on the list. No ROCm path, no ambiguity to argue with.

That single fact rules Strix Halo out of a PAIR based homelab entirely, independent of anything I already found about its prefill gap against a Spark. If I want a router across my own hardware, the third box has to be a Spark or a Studio.

One GPU today, a router with nothing to route to

Here is the honest state of this homelab right now. One box does all the local inference, the RTX 5090 in AI-NT-No-Problem, confirmed as recently as this week when an agent audited every piece of software running on it. My wife runs her own separate homelab on an M4 Mac mini, which happens to already qualify as a PAIR node. Neither fact adds up to a cluster today. A router with one machine to route to is not doing anything yet.

PAIR also only proxies Ollama and LM Studio. My current inference stack is a hand tuned llama-server running directly under systemd, the exact setup I fought to get right in earlier posts, flags and all. Nothing about PAIR touches that setup without a decision to run a different engine somewhere in the mix. And PAIR itself is young, open source under Apache 2.0, born in July, still in beta, with no fixed roadmap NVIDIA has committed to yet.

None of that changes what I want to build next. It changes what building it actually requires.

What would you offload first if every GPU in your house suddenly answered to one address?

By the Numbers

  • 18 minutes to 8 minutes 48 seconds — NVIDIA’s own five subagent demo, one RTX Spark laptop versus a three device PAIR cluster
  • 2 inference engines PAIR proxies, Ollama and LM Studio, and nothing else
  • 0 AMD systems on PAIR’s supported hardware list, which is what rules out Strix Halo
  • 1 GPU currently doing all local inference on this homelab, the RTX 5090 in AI-NT-No-Problem
  • 1 node that already qualifies for a future cluster without buying anything, my wife’s M4 Mac mini
  • ~2 months old — PAIR’s age as an open source project, created in July 2026
  • 1,115 stars / 190 forks on the GitHub repo as of this post

Comments