BERLIN — When Nvidia researchers ran a five-subagent AI task on a single laptop at IFA 2026, the job took eighteen minutes. The same task spread across three devices using a new free tool the company released Wednesday finished in eight minutes and forty-eight seconds. The three computers were already sitting on the network.
Personal AI Router, which Nvidia calls PAIR, is the company’s answer to a problem it spent the past year helping create. As local AI inference has moved from hobbyist curiosity to a mainstream workflow, users with a single high-end GPU have started hitting a practical ceiling: AI agents running simultaneously compete for the same hardware, slowing each other down and pushing users back to cloud services they were trying to avoid. PAIR routes around that ceiling by treating every compatible device on a home network as part of a shared compute pool.
The software is free and open-source, currently in beta for Windows, macOS, and Linux. It works through automatic network discovery, finding compatible machines and dynamically routing independent inference requests to whichever device has available capacity. When a device disconnects, PAIR adapts in real time. The two workflow tools it integrates with out of the box, Ollama and LM Studio, are already the most widely used local AI runtime environments, which means anyone currently running local models can try PAIR without changing their existing toolchain.
The timing at IFA, the Berlin consumer technology show where Nvidia also unveiled its RTX Spark laptop line for an October launch, reflects a specific demographic bet. More than half of US households own two or more PCs, the company’s IFA announcement noted. The company’s case is that most of that computing capacity sits idle throughout the day, and agentic AI workloads are precisely the kind that can exploit distributed parallelism without requiring the low-latency interconnects that make distributed computing difficult at larger scale.
Compatible hardware is broad but not unlimited. PAIR works with GeForce RTX 20 Series and newer, RTX PRO workstation cards, Nvidia’s DGX Spark miniature data-center device, and Apple M4 chips. That range stretches back to GPUs released in 2018. Older cards, including the GTX 10 Series still present in tens of millions of gaming PCs, are not supported. Nvidia has not released detailed throughput numbers for mixed-hardware configurations, and the eight-minute benchmark used three identically configured RTX Spark laptops, not the heterogeneous mix of old and new hardware that most households actually have.
Nvidia published inference performance figures alongside the PAIR announcement. The llama.cpp framework delivers up to 1.9x higher throughput on the RTX 5090. Server-grade hardware fares similarly: vLLM reaches 1.2x improvement on the RTX PRO 6000 and up to 1.4x on dual DGX Spark clusters. These are single-model inference benchmarks measured in isolation, not figures from multi-device PAIR clusters under multi-agent load.
The broader software story at IFA centers on eliminating configuration friction that has kept local AI marginal for general users. Three applications now offer one-click model setup on Nvidia hardware. Hermes Agent, from AI safety startup Nous Research, automatically detects the GPU, selects an appropriately sized model, and runs it through a pre-optimized llama.cpp stack without manual setup. Perplexity Portable Computer runs full search-and-reasoning workflows locally while selectively escalating to cloud models when a task requires a frontier-level response. OpenClaw, an open-source project with more than 380,000 GitHub stars and confirmed by Nvidia as the largest AI project on the platform, lets any Windows user with an RTX card and at least 24GB of VRAM launch a capable local assistant in under five minutes.
The privacy argument runs through all three. Cloud AI inference means queries, documents, and context leave the device. A clinician, an attorney, or a financial analyst running sensitive material through a local model keeps it local. As Engadget reported from IFA, the distributed computing logic Nvidia is applying here is a consumer-level extension of what data centers have done for decades: pooling available compute rather than overprovisioning a single machine.
The question the tool’s beta status leaves open is how PAIR behaves under conditions that differ from Nvidia’s demonstration setup. An M4 MacBook operating on battery-saving throttle, a three-GPU configuration where one card is forty percent slower than the others, a device that disconnects mid-task: none of these scenarios have published characterization. The PAIR announcement also arrives the same week the FTC filed a sweeping antitrust suit against Amazon over its digital advertising practices, a reminder that US regulators are actively examining how large technology platforms control infrastructure. Distributed personal AI, where inference happens on hardware the user already owns rather than in a data center operated by a single vendor, positions itself outside that conversation, but only as long as the PAIR software remains genuinely open-source.
Nvidia has not announced a general availability date. The RTX Spark laptops, the hardware most directly designed for this workload, ship in October from Lenovo and Acer. Until then, PAIR is a beta for the technical audience already running local models. Whether it reaches beyond that audience depends on whether the setup experience stays as frictionless as the IFA demonstration suggested when it meets hardware that is not three identical Nvidia machines sitting next to each other.
