NVIDIA’s PAIR (Personal AI Router) project distributes the processing power of different computers on the same network across AI tasks. Windows, Linux, and macOS devices—including a Windows computer with an RTX graphics card, a machine running Ubuntu, and a MacBook with an Apple Silicon processor—can join the same cluster. In the system, which does not require a separate main server, devices both execute tasks and participate in routing; new devices increase the total capacity.
PAIR is not a new inference engine; it is a proxy layer that runs in front of Ollama and LM Studio. The connection is directed to the router without changing the existing client or agent code. Independent tasks from multiple agents are distributed across idle GPUs; the graphics cards are not physically combined into a single GPU.
In an NVIDIA test, a workload with five sub-agents was completed in 18 minutes using a single RTX Spark Laptop, and in 8 minutes 48 seconds using a cluster consisting of the Laptop, DGX Spark, and RTX 5090. PAIR can temporarily exclude a device from tasks when it is under heavy load due to gaming. When the laptop lid is closed, the other devices continue working; once available, the device can rejoin the cluster. The system does not pool VRAM: Two cards with 8 GB of VRAM each cannot be used as 16 GB of VRAM.
Why it matters
This approach allows AI workloads to make use of the spare capacity of different devices on the same network without relying on a single powerful machine. This creates flexibility for multi-agent users who want to use computers running Windows, Linux, and macOS with different hardware together; the fact that it requires no client-side changes also lowers the barrier to integration with existing tools. The fact that devices may be temporarily taken offline for reasons such as gaming or the lid being closed shows that the cluster operates with variable resources rather than as a fixed server. By contrast, because PAIR does not pool the memory of graphics cards, increased capacity does not raise the amount of VRAM that a single task can fit into. NVIDIA’s test shows that the benefit depends on distributing independent tasks across multiple devices.
Background
NVIDIA is not a new name in the FikirPilot archive: We have published 5 articles mentioning the name in the last 90 days; the most recent is dated September 2, 2026.
Term: agent
An AI agent is software that calls tools and carries out multi-step work to achieve a goal instead of producing a single response.