Nvidia launches PAIR, free software that pools idle home PCs for AI

Nvidia has released Personal AI Router, a free open source tool that finds idle PCs on a home network and spreads local AI inference jobs across them. The company announced it at IFA 2026 in Berlin on 3 September, alongside a wave of RTX Spark hardware landing in October. The beta runs on Windows, macOS and Linux, through both graphical and terminal interfaces.

Despite the name, PAIR isn’t a hardware router. It discovers compatible machines over mDNS or IP addresses, so it routes independent inference requests to whichever system still has spare capacity. Every participating node needs Ollama or LM Studio installed alongside PAIR, and SiliconANGLE reported that the models don’t have to match across machines.

That elasticity matters because the pitch rests on hardware people already own. More than half of US households have two or more PCs, Nvidia says, and much of that computing power sits idle through the day. Nvidia product manager Seth Schneider walked The Verge through a five machine household and put its unused capacity at roughly 165 teraflops.

It’s truly a treasure trove of free tokens just sitting in homes today.

Seth Schneider, product manager, Nvidia, via The Verge

That framing sets expectations high, which makes the constraints worth reading. PAIR supports GeForce RTX 20 Series GPUs and newer, RTX PRO workstation GPUs from Turing onward, DGX Spark, and Apple silicon from the M4 up. Devices pair using a six digit code and then talk over mTLS. Schneider told The Verge that most users are likelier to own one laptop and one gaming PC than a house full of GPUs.

The other half of the announcement is raw throughput, because a cluster only helps if the nodes in it are quick. Nvidia credits kernel optimisations, better speculative decoding and faster prefill for the gains below, and all three figures are its own rather than independently tested.

Backend	Hardware	Claimed gain
llama.cpp	GeForce RTX 5090	Up to 1.9x throughput
vLLM	RTX PRO 6000 Blackwell Workstation Edition	1.2x
vLLM	Two DGX Spark clusters	Up to 1.4x

Those backends feed three agent apps that are getting simplified local model setup on Windows, each built on llama.cpp. They are Perplexity’s Portable Computer, Nous Research’s Hermes Agent, and OpenClaw. Perplexity’s app runs on Linux today with RTX GPUs of at least 24GB of VRAM, with Windows support still pending, and the OpenClaw Windows app carries the same 24GB floor. If you’re weighing which runtime to sit underneath any of them, the choice between Ollama, llama.cpp and vLLM still turns on how many people are queueing.

The hardware side arrives in October. RTX Spark pairs a 1 petaflop Blackwell GPU with a 20 core Grace CPU and up to 128GB of unified memory, and newly announced designs join six OEMs already shipping in October. Wired got hands on with Lenovo’s 16 inch Yoga 9n 2-in-1 at the show and reported it at 0.69 inches thick, though that model tops out at 64GB.

But price is the gap in all of this. Nvidia and its partners haven’t published RTX Spark pricing, and Wired noted that Lenovo’s AMD based ThinkCentre X Ultra, which offers 128GB for agentic workloads, is expected to start at $3,699 in its base configuration. SiliconANGLE also flagged that a PAIR cluster can’t promise consistent quality of service, because a node’s share of the work gets redistributed the moment someone starts a game on it.

So the thing to watch isn’t the demo, it’s the arithmetic. Nvidia hasn’t published a benchmark showing what a mixed household cluster actually delivers on a real agentic task, and the 165 teraflop figure describes capacity rather than useful work. Until somebody outside the company measures a two machine setup, the closest thing to evidence is the 1.9x llama.cpp number on a single RTX 5090.