NVIDIA has released PAIR, short for Personal AI Router, as a beta tool for distributing local AI inference across multiple computers on the same network.
PAIR sits in front of runtimes such as Ollama and LM Studio, discovers paired machines and sends each independent request to an eligible node with available capacity. It does not merge GPUs into one virtual device and it cannot split a single inference job across several computers.
That makes the tool most useful for parallel agent workloads. If one task spawns several independent model calls, PAIR can send them to different machines instead of queuing everything behind one GPU.
NVIDIA says PAIR is open source and supports Windows, Linux and macOS on compatible hardware. Its own demonstration showed a substantial speedup on a multi-agent workload, but that remains a vendor benchmark. The architectural limit is clear: each individual model still has to fit and run on the single node that receives it.