Nvidia used IFA Berlin 2026 to introduce PAIR, short for Personal AI Router, a free application designed to let households pool the GPU power of multiple PCs into a kind of home AI cluster. The software detects other PAIR-enabled devices on a local network and routes AI agent and sub-agent tasks to whichever GPU is available, rather than relying on a single machine to handle everything.
According to Nvidia, most homes already contain significant unused GPU compute. Wccftech reports that Nvidia estimates U.S. households average only about 17 percent daily PC utilization, translating to roughly 165 TFLOPs of idle compute per home, and that tapping this could be worth over a thousand dollars a month in avoided cloud API costs, against a smaller increase in electricity spending. PAIR is Nvidia's answer to putting that spare capacity to use.
The software works by installing on every participating PC, discovering other devices via mDNS (with an IP fallback, per Tom's Hardware), and pairing them using a secure code and encrypted communication channel, according to Wccftech. It proxies through existing local AI tools Ollama and LM Studio rather than requiring a new API, so agent applications don't need special integration. Nodes don't need identical models installed, though having more shared models broadens which systems can handle a given request, Tom's Hardware notes.
Importantly, PAIR does not merge GPUs into one larger accelerator or split a single AI model across machines — it distributes whole, independent jobs to available nodes, according to Wccftech. It is also elastic: if a household member starts gaming or doing creative work on a PC in the cluster, PAIR will route around that machine rather than reserving its resources, per Tom's Hardware. This means the tool doesn't guarantee consistent performance, but is best suited to tasks without strict deadlines.
Demonstrations shown by Nvidia illustrated the potential speedups. PCWorld reported that a networked task spread across three RTX-powered PCs completed in just over 9 minutes with PAIR, versus more than 18 minutes on a single PC. Wccftech described a separate test running a five-subagent workload on the Qwen3.6 35B A3B model, where a single RTX 5090 took 6 minutes and 18 seconds versus 3 minutes and 48 seconds using a two-device PAIR cluster of two RTX 5090s — a result Wccftech cautioned was a specific demonstration and not a general benchmark or promise of linear scaling.
PAIR is compatible with non-Nvidia hardware as well as Nvidia GPUs. Tom's Hardware reports the client supports GeForce RTX 20-series cards or newer, any DGX Spark or GB10-based system, and Macs with M4-series chips or newer. The app will be available at launch for Windows, Linux, and macOS, released as open-source software under the Apache 2.0 license. Nvidia says it has tested configurations with up to 18 devices, though there is reportedly no fixed limit on cluster size. PCWorld notes the launch comes ahead of Nvidia's RTX Spark PCs, expected in October, and frames PAIR as a way to make local AI more viable if cloud AI pricing or availability becomes less attractive.