According to Feber and unite.ai, Nvidia released a beta version of Personal AI Router (PAIR), a free open-source software designed to connect local computers into a single cluster for local AI inference and route agent requests to available machines. Released under the Apache License 2.0 on September 3, 2026, the software allows users to leverage idle hardware capacity across different devices without relying exclusively on cloud services or a single standalone machine.
Hardware Compatibility and Supported Systems
According to unite.ai and teknikgeek.se, PAIR operates across Windows 11, Linux, and macOS platforms, supporting both x64 and arm64 architectures with experimental support for Windows on ARM. Official product documentation listed by unite.ai specifies that compatible hardware includes any GeForce RTX graphics processing units from the 20-series onward, RTX Pro workstation GPUs based on the Turing architecture and newer, DGX Spark and GB10 systems, alongside Apple Mac computers equipped with an M4 chip or newer. The software requires a minimum of 8 GB of RAM and recommends 20 GB or more of free disk space, according to unite.ai and teknikgeek.se reports.
Pro Tip: According to unite.ai, setting up a cluster requires no special cables or server racks. Users simply download the software, add their devices to the network, and run local AI applications through the established proxy endpoints.
Routing Architecture Without GPU Pooling
According to technical details outlined by unite.ai, PAIR functions as a virtual inference router rather than a hardware-pooling tool. The software does not combine multiple GPUs into a single logical accelerator, pool GPU memory, or split a single active inference request across multiple nodes. Instead, tools like Ollama and LM Studio execute models on a designated machine while PAIR schedules independent jobs and routes requests to eligible nodes that already possess the exact requested model, according to unite.ai and teknikgeek.se coverage.
Network Discovery and Security Protocols
According to unite.ai, PAIR utilizes local network discovery via mDNS to automatically detect nearby systems after installation, though administrators can also manually add nodes via IP address. Devices pair securely using a six-digit PIN code displayed on the host machine and entered on the incoming device. All node-to-node communication remains blocked until this pairing occurs, after which traffic is secured using mutual TLS (mTLS) and generated certificates, as reported by unite.ai.
Frequently Asked Questions
Does Nvidia PAIR combine the VRAM of multiple computers?
No. According to unite.ai, PAIR does not pool GPU memory or combine hardware into a larger accelerator. Each request is assigned to a single eligible node for its entire lifespan.
What operating systems support PAIR?
According to unite.ai and teknikgeek.se, the software runs on Windows 11, Linux, and macOS, allowing devices running different operating systems to connect within the same local AI cluster.
Does using PAIR require an active internet connection?
According to unite.ai, an internet connection is only necessary to download AI models initially; running the software locally does not require constant web access.
How do AI applications connect to the PAIR network?
According to unite.ai, applications connect through Ollama-compatible and OpenAI-compatible proxy endpoints created by PAIR, allowing agent frameworks to utilize existing interfaces without integrating a new cluster API.