Nvidia's Free PAIR Tool Routes Local AI Requests Across Idle PCs and Macs

Nvidia has released PAIR, a free, open-source tool that lets idle computers on a home or office network share local AI inference work. Each request still runs on a single machine: the tool spreads out a queue of requests, not a single model.

What Happened

PAIR, short for Personal AI Router, discovers compatible machines on a local network and routes each inference request to whichever one has capacity, so apps that fire off many independent requests at once, such as multi-agent workflows, are no longer stuck waiting on one GPU. It works with Ollama and LM Studio and adapts as devices join or leave the network. Machines find each other with local mDNS discovery, and traffic between them is encrypted with mutual TLS so prompts and responses stay on the local network.

The beta is available for Windows, macOS, and Linux with both a desktop app and a terminal interface, and the source is published on GitHub under the Apache 2.0 license. Supported hardware includes GeForce RTX 20 Series GPUs and newer, RTX PRO workstation GPUs, DGX Spark, and Macs with Apple M4 or newer silicon.

What It Does Not Do

Early coverage described PAIR as a way to pool machines into a personal data center, but Nvidia's own documentation is explicit about the limits. PAIR "does not pool GPU memory, combine GPUs into a larger logical GPU, shard one model across machines, or split an in-flight inference request between nodes." A model that does not fit on one machine will not fit on a PAIR cluster either, and the scheduler checks that the exact model is present on a node before sending work there.

Why It Matters

For people running open-weight models locally, the practical win is throughput rather than model size. Agent frameworks that spawn several sub-tasks at once can spread those across a desktop, a laptop, and a workstation instead of queuing on one card, and nothing leaves the local network. Anyone hoping to run a 70-billion-parameter model on a stack of small GPUs still needs a different approach, such as layer-splitting setups built on llama.cpp's RPC backend.

The Bottom Line

PAIR is a scheduler for local inference, not a way to build a bigger GPU out of smaller ones. That is still useful for anyone whose bottleneck is concurrency, and it fits Nvidia's push to keep more AI work on RTX hardware at home. Expectations should be set by the documentation rather than the headlines: it spreads requests, not models.