NVIDIA has released PAIR, or Personal AI Router, in beta as a routing layer for local AI inference across multiple computers on the same network. The September 3 launch is part of NVIDIA’s broader local-AI push at IFA 2026.
PAIR is not a new inference engine and it does not merge several GPUs into one larger virtual GPU. It sits in front of supported local runtimes, currently including Ollama and LM Studio, discovers paired machines and routes each independent inference request to an eligible node with available capacity.
Built for parallel agent workloads
The design is most useful when an agent generates several independent model calls at once. NVIDIA says PAIR considers node readiness, engine state, model availability and GPU utilization when deciding where to send a request. Tom’s Hardware independently confirms the important limit: PAIR does not pool VRAM and does not split one inference across several computers.
NVIDIA’s own technical demonstration reported a five-subagent Hermes Desktop workload completing in 8 minutes 48 seconds on a three-device PAIR cluster versus 18 minutes on a single RTX Spark laptop. That is a vendor demonstration, not an independent benchmark, so the result should be read as an example of the routing model rather than a universal performance claim.
Local-first, with explicit security controls
PAIR is open source and available for Windows, Linux and macOS. Supported hardware includes GeForce RTX 20-series GPUs and newer, RTX PRO workstation GPUs, DGX Spark systems and Apple M4-or-newer Macs. Nodes are discovered with mDNS, paired explicitly and protected with mutual TLS. NVIDIA says prompts, files and agent context remain on the local network rather than being sent to a cloud inference service.
For users already running local models, the practical value is reuse of idle machines for parallel work. The limitation is equally important: PAIR can reduce queueing across separate jobs, but it does not let a model exceed the memory or compute limits of the single node that ultimately runs that inference.