NVIDIA PAIR Virtual Inference Router Expands Available Compute on Your Local Network

This NVIDIA Developer blog, written by product manager Seth Schneider, is the technical description of Personal AI Router. PAIR is not a new inference engine. Ollama or LM Studio still runs the model on a selected machine. PAIR discovers participating systems, tracks readiness, schedules independent jobs, and returns responses through the interface the agent already uses. No harness changes are required.

Hardware support: GeForce RTX 20 Series and newer, RTX PRO workstation GPUs from Turing on, DGX Spark, and Apple M4 or newer. Beta clients exist for Windows, macOS, and Linux, GUI and terminal. Discovery is mDNS, pairing is user-approved, traffic is mTLS. Eligibility for a request requires the engine enabled and the exact model present on that node. In an unofficial Hermes five-subagent demo using Qwen 3.6 35B A3B, a single RTX Spark laptop averaged 18 minutes; a three-device cluster (RTX Spark laptop, DGX Spark, RTX 5090) averaged 8 minutes 48 seconds. Nvidia says this is not a general benchmark.

Label this vendor-reported. The useful sentences are the negative ones. PAIR does not merge GPUs, does not pool VRAM, and does not shard a single request across machines. Every call is assigned to one eligible node for its lifetime. Sequential work, or a model that lives on only one box, will not get faster. That is the correction to every 'personal AI data center' headline sitting next to this post.

The architecture is a port hijack. PAIR takes over the default Ollama and LM Studio ports, inspects engine and model, and places the job. Agents keep one connection. Home hardware is treated as elastic: laptops sleep, gaming PCs take the GPU back, models are not identical across nodes. Scheduling looks at online status, engine state, model presence, active jobs, and GPU utilization including graphics-heavy apps.

Signal Desk's read: this is a good piece of plumbing for people already running local agents, and a weak story as a mass-market product. The 18-to-8:48 demo needed three heterogeneous Nvidia boxes and a workload that already split into five subagents. Most households do not have that. The open-source repo and the Jobs view that shows where each request ran are the trust features. If those stay accurate, PAIR is a scheduler. If they become a funnel toward Spark SKUs, it is a storefront.

The Apple M4 inclusion is the interesting concession. Nvidia is willing to route work onto silicon it does not sell, because the alternative is a router nobody installs.

Multi-agent local workflows have started to bottleneck a single GPU as subagents fire dozens of independent calls. PAIR is Nvidia's answer at IFA, shipped as a free beta rather than a new engine, sitting on top of the Ollama/LM Studio stack developers already use.