Nvidia Connects Home Computers Into One AI Inference Cluster With PAIR

Nvidia on September 3, 2026 released the beta of the Personal AI Router (PAIR), free open-source software that links compatible computers on a home network into a single cluster for local AI inference, routing requests from agent workloads across whichever machines are available.

Despite its name, PAIR is not a hardware router and not a new inference engine. According to Nvidia’s technical blog post announcing the software, engines such as Ollama and LM Studio still run each model on a selected machine, while PAIR discovers participating systems, tracks whether each one is ready for a request, schedules independent jobs, and returns every response to the application that originated it.

What the Beta Supports

The PAIR beta runs on Windows 11, DGX OS, Ubuntu 14.04, and macOS Tahoe, with x64 and arm64 architectures supported across the three operating systems; the project’s documentation describes Windows on ARM as experimental. Nvidia’s product page lists the supported hardware as all GeForce RTX GPUs from the 20 Series onward, DGX Spark and GB10 systems, and Macs with Apple M4 chips or newer, alongside a minimum of 8 GB of RAM and 20 GB or more of recommended disk space. The technical blog post additionally lists RTX PRO workstation GPUs on the Turing architecture and newer.

No internet connection is required for operation, though one is needed to download models. The product page states that setting up a cluster requires no special cables or racks: users download the software, add their devices, and run their AI applications through it. Nvidia publishes the PAIR source code under the Apache License 2.0, and said developers can inspect the code, report issues, and contribute improvements to discovery, pairing, routing, engine integration, endpoints, and the user experience.

Routing Without Pooling GPUs

Nvidia describes PAIR as a virtual inference router rather than a way to merge hardware. Each request is assigned to one eligible node and remains there for its lifetime; the software does not pool GPU memory, combine GPUs into a larger logical accelerator, shard a single model across machines, or split one in-flight inference request between nodes.

Applications connect through Ollama-compatible and OpenAI-compatible proxy endpoints, which PAIR creates by taking over the default ports the two engines use. Nvidia said this means agent harnesses can keep using the interface they already understand, with no new cluster API to integrate. A node becomes eligible for a request only when a supported engine is enabled on it and the exact requested model is present there, and PAIR prefers nodes it already knows hold the model. Models do not have to be identical across the cluster; loading the same model tag on more nodes gives the scheduler a larger eligible pool.

For each new request, the scheduler weighs whether a paired node is online and ready, whether a supported engine is enabled, whether the requested model is present, the current workload including active jobs, and existing GPU utilization such as a graphics-intensive application running. Nodes can contribute capacity when available and drop away when needed, such as when a laptop sleeps, closes, or leaves the network.

Discovery, Security, and Privacy

After installation, PAIR uses local-network discovery through mDNS to find nearby systems automatically, and a node can also be added by IP address. Pairing two machines uses a six-digit PIN shown on the inviting machine and entered on the invited one. Nvidia said all node-to-node communication is blocked until the secure pairing is established, after which traffic is secured with MTLS and generated certificates. The company positions the software for private local inference, with prompts, files, and agent context remaining on the user’s home network rather than being sent to a cloud inference service. The repository’s documentation cautions users to read its security notes before deploying PAIR on an untrusted or shared network, since it includes local HTTP endpoints, LAN discovery, and cluster networking.

Target Workloads and an Unofficial Demo

Nvidia said PAIR is aimed at workloads that expose many independent requests at once, such as multi-agent applications in which a lead agent breaks a complex task into smaller jobs for subagents. In a demonstration using the Hermes Desktop agent and Ollama, the company reported that a five-subagent task using the Qwen 3.6 35B A3B model took an average of 18 minutes on a single RTX Spark laptop, while a three-device PAIR cluster combining an RTX Spark laptop, a DGX Spark, and an RTX 5090 completed the same workload in an average of 8 minutes and 48 seconds. Nvidia characterized the result as an unofficial, configuration-specific demonstration rather than a general benchmark or a promise of linear scaling, noting that outcomes depend on workload parallelism, model, engine settings, hardware, network, and node availability.

The company said highly sequential tasks, workloads dominated by one long model call, or configurations where only one node holds the requested model may see less benefit. The repository documentation adds that the software currently ships a single scheduling policy combining queued work with a coarse GPU-utilization signal, making it a better fit for similar machines than for a highly mixed cluster, and that Nvidia wants feedback on making the scheduler smarter without fixed commitments on what ships or when.