NVIDIA Pools Home PCs to Run AI Agents Locally
A free router tool spreads inference across idle machines on a home network, and Windows PCs with up to 128 GB of unified memory land in October 2026.
Illustration: three compact compute boxes with blinking status lights on a workbench as a hand plugs in a network cable.
At IFA 2026 NVIDIA introduced PAIR, a free open-source router that hands local AI requests to whichever machine on the network still has spare capacity, and said RTX Spark Windows PCs arrive in October 2026.
At a glance
- PAIR (Personal AI Router): free, open source, beta on Windows, macOS and Linux; works with Ollama and LM Studio.
- Supported: GeForce RTX 20 Series and newer, RTX PRO workstations from Turing on, DGX Spark, Apple silicon from M4 on.
- RTX Spark: 1 Petaflop RTX Blackwell GPU, up to 128 GB unified memory, 20-core Grace CPU; Windows PCs from October 2026.
- Vendor-reported speedups: llama.cpp up to 1.9x on a GeForce RTX 5090, vLLM 1.2x on RTX PRO 6000 Blackwell.
- Hermes Agent, OpenClaw and Perplexity Portable Computer promise faster setup; two of them need 24 GB VRAM or more.
The pitch NVIDIA brought to IFA 2026 on September 3, 2026 is that a household already owns enough silicon to run agents without a data center. The company released PAIR, a free tool that treats every capable PC on a local network as one pool, and said RTX Spark Windows PCs follow in October 2026.
One network, many GPUs
PAIR stands for Personal AI Router. It scans the local network for compatible machines and sends independent inference requests to whichever system has capacity to spare, reshuffling as devices come and go. Ollama and LM Studio serve as the backends.
The beta runs on Windows, macOS and Linux. Eligible hardware covers GeForce RTX 20 Series cards and newer, RTX PRO workstations from the Turing architecture on, DGX Spark and Apple silicon from M4 on. That mix matters: a gaming desktop and a laptop can carry the same job together instead of one machine stalling alone.
What RTX Spark actually is
NVIDIA describes the RTX Spark systems as one box for creators, gamers and agents, built around a 1 Petaflop RTX Blackwell GPU, up to 128 GB of unified memory and a 20-core Grace CPU. Acer showed a compact desktop at the show and Lenovo brought the Yoga Pro 9n and the Yoga 9n 2-in-1. A new Windows agent framework is mentioned for these machines.
No price appears anywhere in the announcement, and October 2026 is given without a specific day.
Setup friction and throughput claims
Three apps are named as the shortcut into local models. Hermes Agent offers one-click setup on Windows, with Linux described as coming. Perplexity Portable Computer runs on Linux with RTX GPUs carrying 24 GB of VRAM or more, and a Windows build is promised. OpenClaw ships a Windows app at the same VRAM bar.
The performance figures are NVIDIA's own. llama.cpp is credited with up to 1.9x more throughput on a GeForce RTX 5090 through kernel work, better speculative decoding and faster prefill. vLLM is put at 1.2x on an RTX PRO 6000 Blackwell Workstation Edition and up to 1.4x across two DGX Spark clusters. On the video side, FastH3, a four-step distilled open-weight take on MiniMax-H3 built with FastVideo, is credited with 7x.
Models named for local use
The examples cited as runnable on this hardware are Nemotron 3.5 Lightning at 30B parameters, Qwen3.8-Flash-Next and Qwen3.8-27B, the LTX 2.5 video model, Meta's Muse Glimmer at 30B parameters and DeepSeek v4 Flash at 284B MoE parameters.
What this report cannot confirm
Only one publisher stands behind these claims: NVIDIA's corporate blog. No independent testing backs the speedups, and the company publishes neither the measurement setup nor the baseline each multiplier is measured against. Pricing, volumes and country-by-country availability are absent.
Microsoft's specific contribution is also unresolved. The post frames the work as a collaboration and points to a new Windows agent framework, but names no particular Microsoft technology and makes no commitment a reader could check.
FAQ
How does NVIDIA PAIR split work across PCs?
It discovers compatible machines on the local network and routes independent inference requests to systems with free capacity, adjusting as devices join or leave. It uses Ollama and LM Studio as backends and is in beta on Windows, macOS and Linux. It is free and open source.
When do RTX Spark PCs ship and what do they cost?
NVIDIA says October 2026 and gives no exact day, no price and no country list. Acer showed a compact desktop, while Lenovo brought the Yoga Pro 9n and the Yoga 9n 2-in-1.
Which hardware can join a PAIR setup?
GeForce RTX 20 Series and newer, RTX PRO workstations from Turing on, DGX Spark, and Apple silicon from M4 on. Perplexity Portable Computer and OpenClaw additionally call for 24 GB of VRAM or more.
