Nvidia launches free software that turns home PCs into a local inference cluster

Nvidia is releasing Personal AI Router, or PAIR, a free open-source tool that stitches desktops and laptops on a home network into a single inference backend. Despite the name there is no hardware router involved. The software discovers compatible machines on the LAN, links them, and prepares them for inference workloads with tools such as Ollama and LM Studio. Support covers GeForce RTX 20-series and newer, RTX Pro cards, DGX Spark systems, and Apple M4 silicon and later.
How it works in practice
The engine only taps a machine while it sits idle, so normal work is not interrupted. When a job arrives — say an agentic workflow that breaks a complex problem into smaller sub-tasks — the system splits the processing across available devices in parallel, avoiding a single-GPU bottleneck. If the user starts gaming on the desktop, PAIR detects the shift and releases those resources automatically; when the machine goes quiet again it rejoins the pool.
Security and availability
Devices pair through a six-digit code, then communicate over an mTLS (Mutual Transport Layer Security) channel that mutually authenticates both ends. The beta is available today for Windows, Linux and macOS.
The numbers behind the vision
In a press briefing product manager Seth Schneider walked through an extreme scenario: a household with an RTX Spark laptop, a DGX Spark station, an RTX 5090 laptop, a gaming rig and a MacBook Pro. Together, he said, they deliver roughly 165 teraflops of unused compute. "That's a treasure trove of free tokens sitting in homes today," he added, even after factoring in average US electricity costs. Schneider acknowledged most users will make do with a single laptop — Mac or Windows — plus one gaming PC.
Ready-to-run agent apps
Alongside PAIR, Nvidia announced that three prominent agent applications — Perplexity Portable Computer, Hermes Agent and OpenClaw — will get a simplified local install on the company's GPUs in Windows. The aim: a working agent in a few clicks, without lengthy manual configuration.