If you have more than one capable PC sitting around the house, NVIDIA has a new idea for putting all that hardware to work, and this tool is called the NVIDIA Personal AI Router, or PAIR.
The new utility can connect compatible PCs, Macs, and NVIDIA DGX Spark systems on the same local network and route AI inference requests between them. Instead of leaving a second gaming PC or workstation idle while another machine is struggling with several AI tasks, PAIR can make use of that available compute.
The idea is particularly interesting for local AI agents, where one application can generate several independent inference requests at the same time. Rather than forcing everything through a single GPU, PAIR looks at the connected machines and sends each request to a suitable system based on factors such as GPU utilisation, model availability, and the inference engine running on that machine.
That means a setup with multiple PCs can effectively behave like a small personal AI cluster – without requiring specialised networking equipment or a complicated server setup. And yes, existing applications do not necessarily need to be rewritten to support it.
NVIDIA designed PAIR to work with existing Ollama and LM Studio interfaces, allowing compatible AI applications and agents to send requests through PAIR while the router decides which machine should handle them. This is especially relevant as tools such as local AI agents increasingly perform multiple tasks in parallel. If one machine is already busy processing an inference request, another available PC can take the next one instead of leaving its GPU sitting unused.
This ain’t your data center “one big giant GPU” tool
PAIR does not combine multiple GPUs into one massive pool of VRAM. You cannot take a 12GB GPU, another 16GB GPU, and somehow turn them into a single 28GB graphics card. Instead, PAIR distributes independent inference requests between separate machines. Each request is handled by one system, with the router deciding which node is best suited for the job. That makes PAIR much more useful for workloads involving multiple simultaneous tasks than for trying to run a model that is simply too large for one GPU.
But thanks to the cross-platform compatibility layer, a user could have a desktop with a GeForce RTX GPU, another PC elsewhere in the house, and perhaps a MacBook. Install PAIR on the compatible machines, connect them to the same network, and pair them together. From there, PAIR can monitor which systems are available and route inference requests accordingly.
It also means the cluster can adapt as machines become busy. If someone starts gaming on one of the PCs, for example, PAIR can take its current workload into account instead of blindly continuing to send AI requests to it. NVIDIA says its scheduling considers node readiness, engine state, model availability, and GPU utilisation
Maximum data control
For users that want to keep their AI workloads away from the cloud, NVIDIA PAIR helps them with that, as models and paired systems are running locally, prompts, files, and agent context can remain within the user’s local network. NVIDIA also uses secure pairing and mTLS encryption between nodes.
Although there is still an internet connection requirement when downloading models, PAIR itself does not require the internet to operate once the local setup is ready. For people experimenting with local LLMs, that makes the concept particularly appealing. Instead of paying for additional cloud inference every time an AI agent needs more compute, existing hardware around the house can potentially handle some of that workload.
NVIDIA PAIR is currently available for download here.
