NVIDIA unveiled a "local artificial intelligence (AI)" strategy that expands personal PCs into AI agent execution hubs. It introduced the open-source software "NVIDIA PAIR," which uses the idle compute performance of multiple PCs in a home for AI tasks, and it will also launch products in Oct. loaded with the personal AI PC platform "RTX Spark."
On the 4th (local time) in Berlin, Germany, at IFA 2026, Europe's largest consumer electronics and information technology (IT) trade show, NVIDIA is showcasing PAIR, PCs based on RTX Spark, and the latest local AI technologies. At IFA, where AI is spreading to everyday devices such as TVs and home appliances, the company laid out a strategy that PCs, too, will shift from gaming and content creation to devices where personal AI agents operate.
An NVIDIA official in charge of local AI agent products said in a pre-briefing for media, "This year is the year of local AI," and "People have begun using local AI for everyday tasks, and at the same time, new models with greatly improved performance are emerging." The official said downloads of the top 20 large language models (LLMs) on Hugging Face this year have already reached four times last year's level.
An NVIDIA official said, "Agents are a completely new way of computing," and "Instead of users learning how to use applications to produce desired results, agents figure out how to execute tasks once users define what they want." The method is that after the AI infers the command entered by the user, it directly uses multiple applications and tools to carry out the work.
PAIR, introduced for the first time at this IFA, also targets this AI agent use. It automatically finds PCs connected to the same local network and distributes multiple inference requests generated by an agent to devices that can handle them. When a single agent assigns complex tasks to multiple sub-agents, each task can be processed simultaneously on devices such as desktops and laptops. While the main PC is used for gaming or video rendering, AI computation can be offloaded to other devices.
An NVIDIA official said, "PAIR is a personal AI router that intelligently distributes AI inference across all devices on a local network," adding, "It does not combine multiple PCs into one bigger PC; instead, it distributes each inference request to devices that can handle it."
This means it is not a technology that physically combines GPUs or graphics memory (VRAM) from multiple devices to run a single large AI model. Each inference request is handled on one PC that can run the relevant model. Instead, it distributes simultaneous, independent requests across multiple devices to reduce bottlenecks on a single GPU. In NVIDIA's disclosed test, five sub-agents working simultaneously took about 18 minutes on a single device, but when distributed to three devices using PAIR, it was shortened to 8 minutes 48 seconds.
PAIR leverages existing local AI programs "Ollama" and "LM Studio." Each PC runs AI models, and PAIR checks device status, workload, and whether the required model is installed to allocate requests. It supports Windows, macOS, and Linux and has been released as free open source. It supports GeForce RTX 20-series and later GPUs, RTX Pro Workstation GPUs, DGX Spark, and devices with Apple M4 and later chips.
It also reduces the complex setup process previously required to run AI models on PCs. Until now, users had to find appropriate models and inference programs and configure various settings themselves. "Hermes Agent" will support a one-click feature that detects NVIDIA GPUs, selects suitable models, and handles download and setup. Openclaw will also simplify local AI setup on Windows PCs with NVIDIA GPUs that have 24GB or more of memory.
Perplexity's local AI agent "Portable Computer" is also expanding support with RTX GPUs. It can process tasks without sending sensitive documents outside the PC, while using high-performance cloud AI models only when needed. NVIDIA said it boosted throughput on GeForce RTX 5090 by up to 1.9 times through optimizations to the open-source inference program "llama.cpp."
The commercialization timeline for the next-generation AI PC platform RTX Spark has also been firmed up. An NVIDIA official said, "RTX Spark will launch in Oct., and the official product name is 'N1X.'" RTX Spark N1X combines a Blackwell-based RTX GPU with a Grace CPU and supports up to 128GB of unified memory. AI compute performance is up to 1 petaflop, targeting not only gaming and content creation but also the always-on operation of AI agents on PCs.
At this IFA, Lenovo is unveiling "Yoga Pro 9n" and "Yoga 9n 2-in-1," and Acer is showcasing a small desktop product based on RTX Spark. Major PC makers such as Asus, Dell, HP, and MSI are also preparing related products. Electronic Arts (EA), Embark, and Ubisoft have newly joined the lineup of games supporting RTX Spark. Krafton, NetEase, Inc., Riot Games, and Xbox had previously announced plans to support it.