Nvidia PAIR pools your home GPUs to run AI agents twice as fast

A new free tool from Nvidia called PAIR — Personal AI Router — routes AI agent tasks across multiple computers on your home network, treating your collection of gaming PCs, workstations, and even Macs as a unified inference cluster. The idea is simple enough: most AI-heavy computers sit idle most of the time, and local AI workloads are increasingly built around agents that spawn dozens of independent sub-tasks at once. PAIR connects the spare cycles.

The software works as a transparent proxy on top of Ollama and LM Studio, the two most popular local inference tools, so existing setups need no code changes. Once installed on each machine, PAIR uses mDNS — the same protocol your printer uses to appear on the network — to discover compatible devices automatically. A lead agent submits its work as usual, and PAIR decides which machine handles which sub-task based on which models each machine has available. No identical setup required across machines: a gaming PC running Llama 3.3 and a laptop with Mistral can cooperate on the same job.

The benchmark Nvidia published makes the difference concrete: a five-subagent task that took 18 minutes on a single RTX Spark laptop finished in 8 minutes 48 seconds across a three-device cluster. That is roughly a 2× speedup without touching a line of code or paying for cloud compute. The gain scales with the number of machines and the nature of the workload — tasks that break into many parallel pieces benefit most.

Privacy is the other selling point, and it is the more durable one. Every prompt, file, and piece of agent context stays on your home network. Nothing is routed to a cloud inference service; there is no API key, no usage metering, and no company receiving your queries. For anyone running agents on sensitive documents — legal drafts, medical notes, personal correspondence — that distinction matters more than benchmark numbers.

The caveats are real. PAIR offers no guaranteed quality of service: if the machine you were counting on wakes its GPU for a game or video export, PAIR redistributes that workload to whatever else is available. The tool is openly designed for long-running tasks where the deadline is flexible, not for applications where latency has to be predictable. It is also beta software. Nvidia is pitching this at enthusiasts with at least one compatible machine — a GeForce RTX 20-series card or newer, or Apple M4-series silicon — and the expectation that they already know their way around Ollama.

PAIR is free, open-source, and available in beta now across Windows, macOS, and Linux. Nvidia has not announced a roadmap for QoS guarantees or a stable release date. The practical floor for benefiting from it is two machines with capable hardware, which rules out most single-computer setups. For users who already run Ollama on more than one device at home, the barrier to entry is low enough to justify the experiment.

Tags: nvidia, AI agents, Open Source, Local AI, LM Studio, Ollama