Nvidia Launches Free PAIR Tool to Share AI Workloads Across PCs and Macs

Nvidia has unveiled Personal AI Router, or PAIR, a free open-source tool that distributes local AI workloads across available computers on the same network, letting idle hardware help run multi-agent tasks. Announced at IFA 2026, PAIR works across Windows, macOS, and Linux systems without requiring users to build a dedicated AI server or cluster. PAIR does not combine multiple GPUs into one larger accelerator; instead, it finds available machines and sends separate AI requests to whichever system has spare computing capacity — an approach Nvidia is targeting particularly at multi-agent workloads. PAIR discovers devices via multicast DNS and secures inter-device communication through a six-digit verification PIN and mutual transport layer security encryption, with traffic kept on the local network. The tool targets agentic workflows, where a primary autonomous agent delegates assignments to multiple subagents. According to Nvidia, a five-subagent task analyzing a synthetic household inbox using the Qwen 3.6 35B A3B model took 18 minutes on an RTX Spark laptop alone, but finished in 8 minutes and 48 seconds when PAIR split the jobs across a three-node cluster including an RTX 5090 and a DGX Spark. The software officially validates hardware spanning Nvidia GeForce RTX 20-series GPUs and newer, workstation-grade RTX Pro chips, DGX Spark units, and Apple M4 silicon or newer. PAIR cannot pool VRAM or split a single inference request across multiple machines; each request runs entirely on one eligible node, making it most useful when a workload has several independent jobs that can run simultaneously. The bigger idea behind PAIR is making existing hardware behave like a flexible pool of resources — a gaming PC, work laptop, and Mac could have their unused capacity borrowed when available, making local AI more practical for households that already own several capable machines without buying another server or sending data to the cloud. For users with only one powerful machine, or workloads dominated by a single inference request, the benefit will be much smaller.