Nvidia Launches PAIR to Route AI Tasks Across Local GPUs

Nvidia has launched the beta version of its Personal AI Router (PAIR), a tool designed to aggregate the inference capacity of multiple local computers. By automatically distributing artificial intelligence requests across a local network, PAIR aims to solve the hardware bottlenecks that occur during multi-agent AI workloads. When a lead agent dispatches multiple subtasks to sub-agents, a single local GPU can easily become overwhelmed. PAIR acts as a proxy, receiving requests, identifying model and engine requirements, and routing individual tasks to eligible nodes without requiring changes to the underlying agent architecture.

The router integrates with popular local inference services like Ollama and LM Studio. It supports Windows 11, Linux, and macOS across both x64 and arm64 architectures, allowing users to pair nodes running different operating systems. However, Nvidia clarified that PAIR does not pool VRAM or merge multiple GPUs into a single larger accelerator. Instead, it manages the placement of individual, complete inference requests.

To demonstrate the tool's capabilities, Nvidia showcased a demo featuring Hermes Desktop, Ollama, and PAIR. The setup split a task into five independent specialist analyses. By distributing these tasks across an RTX Spark laptop, a DGX Spark, and an RTX 5090, the system achieved a roughly 2x reduction in completion time compared to running the workload solely on the RTX Spark laptop. Early adopters have already reported success with the tool. One Reddit user, Vegetable-Warthog81, noted that PAIR made distributing jobs "pretty painless" when running the Qwen 3.8 27B model via Ollama across three RTX 5090 GPUs for stable, repetitive workloads.

For developers and AI practitioners, PAIR provides an accessible way to maximize existing local hardware without relying on expensive cloud setups or complex clustering software. While it does not allow for running massive models that exceed single-card VRAM limits—for which tools like Petals or Mesh LLM with Skippy remain necessary—it significantly optimizes multi-agent workflows by keeping all available local GPUs busy.