.png){kind=logo}
NVIDIA is turning idle computers into a distributed AI workforce with Personal AI Router, or PAIR, a free, open-source software tool that routes AI requests to compatible machines on a home network.
PAIR works with Ollama and LM Studio and supports Windows, macOS, and Linux. Instead of combining the computing power of multiple systems into a single larger machine, it routes each independent AI request to the single computer with available capacity.
PAIR essentially acts as a traffic manager for AI workloads. It discovers compatible machines on the network and decides which system should handle each request. The software is licensed under the Apache 2.0 license and supports Windows 11, Linux, and macOS on x64 and arm64 machines.
Supported hardware includes GeForce RTX 20 Series graphics cards and newer; RTX PRO workstation cards built on Turing or later; DGX Spark; and Apple silicon from M4 onward. Windows on ARM support remains experimental.
Machines discover each other through mDNS, while pairing requires a six-digit PIN. Traffic between paired devices uses mutual TLS, with certificates used to establish trusted connections.
A key limitation separates PAIR from conventional multi-GPU systems. Four PCs with 16GB graphics cards remain four separate 16GB machines. PAIR does not pool GPU memory, combine graphics cards into a single logical GPU, shard models across systems, or split a request while it is running.
The benefit instead comes from parallel workloads. Five independent jobs can run across several machines rather than queue behind a single computer.
Also Read: DLSS 5 Debut: NVIDIA Brings Neural Rendering Graphics to NBA 2K27
NVIDIA's approach becomes particularly relevant for AI agents, which can break a larger task into multiple subagent requests. Those requests can run simultaneously across available machines.
In NVIDIA's demonstration, a five-subagent workload took 18 minutes on one RTX Spark laptop. Across three machines, an RTX Spark laptop, a DGX Spark, and an RTX 5090, the workload took 8 minutes 48 seconds, a 51 percent reduction.
PAIR is currently available in beta for Windows, Linux and macOS, giving households and developers a way to put otherwise idle computing hardware to work on local AI workloads.
