NVIDIA has announced the launch of PAIR, a free and open-source software that connects multiple home computers on the same network to jointly handle local AI inference tasks using idle computing power. The PAIR beta is now available, supporting Windows, Linux, and macOS.
PAIR stands for "Personal AI Router," but it is not a hardware router. This software can discover compatible computers on the network, complete device connections and task preparation, and work with tools like Ollama and LM Studio to execute local AI inference and agent workflows.
In terms of hardware, PAIR is primarily compatible with GeForce RTX 20 series and newer graphics cards, while also supporting RTX Pro GPUs and DGX Spark systems. Devices equipped with Apple M4 chips or newer can also join the computing network.
PAIR prioritizes calling upon idle devices to avoid interfering with other tasks the user is currently performing. Multiple computers can process split computational requests in parallel, thereby reducing the burden on a single GPU when handling complex agent workflows. PAIR also dynamically adjusts resources when devices join or leave the network; if a user starts running a game on their desktop, this change in device status is also taken into account during scheduling.
NVIDIA product manager Seth Schneider stated that the common usage scenarios envisioned by the company do not require a large number of high-end devices. Most users might only connect a MacBook or Windows laptop along with a gaming PC. For households with multiple high-performance devices, PAIR can centrally utilize computing power that would otherwise go unused.
Regarding security, devices need to be paired using a six-digit code, after which they establish an encrypted communication channel using the mutual Transport Layer Security (mTLS) protocol. In addition to PAIR, three AI agent applications—Perplexity Portable Computer, Hermes Agent, and OpenClaw—will provide simplified local installation processes for NVIDIA GPUs on the Windows platform, reducing the manual configuration required for the first run.