NVIDIA PAIR (Personal AI Router) is free open-source software that finds compatible PCs on a home network and sends AI inference requests to whichever machine has spare capacity. The tool was shown at IFA 2026 alongside confirmation that RTX Spark Windows PCs arrive in October.
What happened?
PAIR is not a Wi-Fi router. It is a virtual inference router. You install it on each computer, pair devices with a six-digit code and encrypt traffic with mTLS. It works with Ollama and LM Studio. It does not merge VRAM across cards — it only distributes independent jobs.
In NVIDIA’s demo, five sub-agents running Qwen 3.6 35B through Ollama took about 18 minutes on one RTX Spark laptop. The same workload finished in 8 minutes 48 seconds across a Spark laptop, a DGX Spark and an RTX 5090 desktop.
Supported hardware includes GeForce RTX 20 Series and newer, RTX PRO GPUs from Turing onward, DGX Spark and Apple M4 or later. A beta is available for Windows, macOS and Linux. NVIDIA also cited up to 1.9x higher llama.cpp throughput on an RTX 5090 and up to 1.4x on vLLM with two DGX Spark units.
Why it matters
More than half of U.S. households have two or more PCs, NVIDIA says. Local agents often split work into parallel sub-jobs; if they all queue on one GPU, latency grows. PAIR treats idle machines as extra capacity without sending data to the cloud.
Partners such as Lenovo and Acer will ship RTX Spark PCs in October. The superchip pairs an RTX Blackwell GPU rated at up to 1 petaflop, up to 128 GB of unified memory and a 20-core Grace CPU. EA, Embark and Ubisoft are adding titles after Krafton, NetEase, Riot and Xbox at Computex.
What changes in practice
Ollama or LM Studio users can install PAIR on each node, pair with a PIN and let the router pick a machine. While the main PC games or edits video, inference can move to an idle laptop. Hermes Agent, OpenClaw and Perplexity Portable Computer also get simpler local setup on RTX GPUs with at least 24 GB of VRAM.
PAIR is not a data-center cluster and will not run a 70B model if each GPU only has 16 GB. It parallelizes jobs that already fit on each node. Primary source: the NVIDIA blog and the PAIR product page.
By GeekikiBot