Key points
- Free Nvidia tool routes AI inference across computers on a home network
- Three-device cluster cut an 18-minute agent job to 8:48
- RTX Spark Windows PCs arrive in October
On Thursday, September 3, at IFA 2026, Nvidia (NVDA) released PAIR, a free open-source tool that discovers compatible computers running PAIR on a home network and routes inference requests to machines with available capacity.
The name is short for Personal AI Router, and Nvidia calls it a virtual inference router. An agent running on one machine points at PAIR the way it would point at a local model server, and PAIR decides which computer on the network actually answers. Compatible agent software doesn't need a new integration. The target is inference, the running of a trained model rather than the training of one, which now accounts for more spending than model training.
PAIR does not combine the GPUs or their memory, and it cannot split one model request across several machines. Each request runs from start to finish on one node. The speedup comes when an agent generates several independent requests that can run at the same time. In Nvidia's own words, PAIR does not "merge GPUs or pool VRAM into one larger accelerator" and does not "shard a single model or split one inference request across machines."
The beta runs on Windows, macOS, and Linux, in both graphical and terminal versions. It supports GeForce RTX 20 Series cards and newer, RTX PRO workstation cards from the Turing generation on, the DGX Spark desktop, and Apple silicon from the M4 forward. The Apple line means a cluster doesn't have to be all Nvidia hardware.
Machines don't enlist themselves. PAIR discovers candidates over mDNS, and Nvidia said all traffic between two nodes stays blocked until a pairing is approved. Once the connection is made, communication between the nodes is encrypted with mTLS and generated certificates.
After that the cluster is elastic. A node contributes capacity when it has some and drops away when it powers down or hibernates. Nvidia lists five things PAIR weighs before it hands a job to a node:
- whether the node is online and ready
- whether a supported inference engine is enabled on it
- whether the exact model requested is already there
- how much work it is already doing
- how hard its GPU is being pushed by whatever else is open
The models don't have to match across the cluster, because PAIR can route a request to the machine that happens to hold the one it needs.
Nvidia published one measurement in a five-subagent workload run on Qwen 3.6 35B A3B. On a single RTX Spark laptop, the job took 18 minutes on average. On a three-device cluster of an RTX Spark laptop, a DGX Spark, and an RTX 5090, it took 8 minutes and 48 seconds. Nvidia labels that an "unofficial, configuration-specific demo, not a general benchmark or a promise of linear scaling," and says results depend on how parallel the workload is, along with the model, the engine settings, the hardware, the network, and which nodes are free.
