One endpoint for several local inference machines
Nvidia introduced the beta of Personal AI Router, or PAIR, on September 3. The Verge reported the launch, and Nvidia's technical documentation explains that PAIR connects compatible computers on a local network and presents applications with one local endpoint. Behind that endpoint, the software selects a machine able to serve each inference request.
PAIR currently works with Ollama and LM Studio interfaces. Nvidia says supported configurations include Windows, macOS and Linux systems, with eligible hardware spanning GeForce RTX 20-series and newer GPUs, RTX PRO workstations, DGX Spark and Apple M4-or-newer silicon. Users must install PAIR on participating machines, explicitly pair them and ensure the requested model exists on an active inference engine.
Routing is not GPU pooling
The architectural limit is important. PAIR does not merge VRAM, combine several GPUs into one accelerator, shard a model or divide a request that is already running. Each request goes to one eligible node from start to finish. Additional machines can increase concurrency when a workload exposes independent requests, but they do not automatically make a single inference faster.
Nvidia says nodes discover one another on the local network, pairing requires user approval, and traffic between paired nodes uses mutual TLS. The scheduler considers reachability, engine state, model availability and pending work. Its Jobs view is the place to verify where requests actually ran.
Practical implications for local AI
PAIR is most relevant to developers and power users running several agents or concurrent local tools. Its proxy approach reduces per-application configuration, but users still need compatible models, sufficient memory on each serving node and a trusted local network.
Because PAIR is a beta and open-source project, the next evidence to watch includes issue reports, compatibility updates, security review and measurements on real workloads rather than vendor demonstrations alone.