Nvidia is pushing its networking prowess further into the local AI space with a software tool allowing developers to interconnect MacOS, Windows, and Linux systems.
Personal AI Router, or PAIR, connects multiple devices on the same local network. Nvidia bills PAIR as the means to build “a personal home AI cluster,” with the software capable of routing independent inference requests to an available system.
It works across systems housing Nvidia RTX GPUs and its lunchbox-sized desktop AI supercomputer line, DGX Spark. MacOS systems that are compatible with Nvidia’s offerings can employ PAIR to route local AI inference workloads to other machines on the network.
The virtual inference router routes workloads, while platforms like Ollama or LM Studio run the model on the machine the software selects. An Nvidia technical blog describes how it works: “PAIR discovers participating systems, tracks whether each system is ready for a request, schedules independent jobs, and returns each response to the application that originated it.”
Nvidia stressed that its newest software offering differs from a custom routing script with glossy green paint smeared over it in that it distributes workloads while keeping them secure. Where routing raw tokens across a home network creates security vulnerabilities, Nvidia contends PAIR employs mutual Transport Layer Security (mTLS) for encryption, while users employ a six-digit PIN between nodes, with the cluster refusing a machine that is not a member.
No fancy data center-esque cables are required to use it, with the software tool itself open source, so anyone can use it.
It does require a minimum of eight gigabytes of RAM to run, but does not need to be connected to the internet during operation. Windows devices need to be running Windows 11, while Apple systems need MacOS Tahoe. For Linux systems, PAIR needs a distro like Ubuntu 14.04 to run. At the time of writing, it’s only available in English.
PAIR seems perfect then for local AI enthusiasts deploying agentic offerings like OpenClaw – the agent platform CEO Jensen Huang enthusiastically billed as “the iPhone of tokens.”
Nvidia has gone all in on agentic AI in the wake of OpenClaw, unveiling its own equivalent at GTC in NemoClaw, as well as a toolkit to encourage enterprises and developers to build AI agents.
PAIR is the next step of that bet, with the vendor’s technical blog explaining how it works for local agentic workloads: “Agents can send a request through the familiar local interface it expects. PAIR receives the request through its proxy, identifies its engine and model requirements, and selects one eligible node. That node executes the request from start to finish and sends the response back through PAIR. The agent continues to see one connection while PAIR handles placement behind it.”