Nvidia PAIR turns idle home computers into a distributed AI inference network

Nvidia has released a tool designed to squeeze more performance out of home and office computer hardware by using idle machines as distributed processing nodes for AI agents.

The company unveiled the Personal AI Router (PAIR), an open source network utility that discovers idle computers on a local network and routes small AI inference requests to whichever machine has the capacity and the right model available. This enables agentic systems to spawn multiple parallel subtasks rather than running them sequentially, significantly improving throughput.

How it works#

PAIR isn't a new inference engine itself. Instead, it acts as a smart dispatcher that sits between an AI agent and existing local inference tools—specifically Ollama or LM Studio installations already running on networked machines.

When an agent needs to execute a task, PAIR:

• Receives the inference request through a standard local interface
• Identifies what model and engine the request requires
• Scans the network (using mDNS) to find an eligible machine
• Routes the job to one compatible node to run in full
• Returns the result, keeping the agent unaware of the routing happening behind the scenes

The key limitation is that PAIR cannot split a single inference call across multiple machines or merge GPU memory pools. Each request executes on a single device from start to finish.

Hardware support and practical implications#

The software works across Windows, macOS, and Linux, supporting:

• Nvidia GeForce RTX 20-series GPUs and newer
• Apple Silicon Macs with M4 chips or newer
• Nvidia DGX Spark systems (and RTX Spark devices coming later in 2026)

Notably, Nvidia is backing Mac support despite Apple's processors lacking Nvidia GPUs—a recognition that many users already run local AI workloads on Apple hardware.

Each machine in the network can host different models. PAIR gains routing flexibility when the same model runs on multiple systems, allowing it to distribute concurrent requests more intelligently.

Real-world performance and availability#

In Nvidia's demonstration, running two high-end RTX 5090 systems (each with 32GB of memory) and the Qwen 35B model achieved roughly 1.6x speedup when coordinating five parallel subagents. Results with more modest setups—mixing Mac Studios, Mac minis, and gaming PCs—remain to be tested in practice.