NVIDIA PAIR (2026): It Does Not Pool Your GPU Memory

Bottom Line

• PAIR is a scheduler, not a memory pool. It sends each whole request to one machine that can already run the model.
• NVIDIA says so plainly. From the PAIR FAQ: “The devices remain separate systems that handle parallel tasks; PAIR doesn’t combine them into one virtual GPU.”
• Two 24GB PCs do not make 48GB. If a model does not fit on one of your machines, PAIR does not make it fit.
• It is free, open source, and cross-platform. Windows, macOS and Linux; GeForce RTX 20 Series and newer, RTX PRO from Turing onward, DGX Spark, and Apple M4 or newer.
• Buy accordingly: for a bigger model, buy one box with more memory. For more jobs at once, a second machine now earns its keep.

What PAIR Does

NVIDIA’s IFA 2026 announcement describes it in one sentence:

“PAIR automatically discovers compatible PCs on a local network and routes independent inference requests to whichever system has capacity.”

And on the benefit:

“PAIR distributes those jobs across available PCs instead of having them all wait on a single GPU.”

It hooks into tools you already run. NVIDIA states that “it works with Ollama and LM Studio and can adapt as devices join or leave the network.”

So the picture is a household or a desk with three machines — a gaming PC, a laptop, a Mac — and several agents or chats competing for one GPU. PAIR spreads them out. That is a real problem and this is a reasonable answer to it.

What PAIR Does Not Do

The word doing the work in NVIDIA’s sentence is independent. Each inference request is routed, whole, to one system. The FAQ removes the remaining ambiguity:

“The devices remain separate systems that handle parallel tasks; PAIR doesn’t combine them into one virtual GPU.”

The reading that follows, and we are labelling it as a reading rather than a quote: if a request executes entirely on one machine, then that machine must hold the model on its own. Memory does not add up across the network. Nothing in NVIDIA’s material describes splitting a single model’s layers or tensors across two computers.

That distinction is the whole article, because the hopeful version of this story is very appealing and very wrong.

The Real Thing That Splits a Model Across GPUs

Model sharding exists — it just needs the GPUs in one machine, on the PCIe bus, not on Ethernet. That is tensor parallelism, and we have walked through it: vLLM tensor parallel on two GPUs covers --tensor-parallel-size 2, the head-count divisibility rule, and the NCCL errors you will meet.

If two cards in one box is where you are heading, the platform matters more than people expect. Motherboard and CPU for a multi-GPU LLM rig covers the PCIe lane budget, and what PSU for a local AI rig covers the power draw that catches dual-card builders out.

So What Should You Buy?

If the problem is “the model does not fit”: buy memory in one machine. A second box is the wrong purchase. Start from what local LLM fits my machine, then pick the tier from the gear page. The local LLM estimator will size it against your actual hardware.

If the problem is “my agents queue behind each other”: you are the buyer PAIR was built for, and you probably already own the second machine. Install it before you spend anything. It is free, so the experiment costs an evening.

If you were about to buy a second GPU for one machine: read best local LLM for 64GB VRAM first. Two cards in one box does raise your ceiling, unlike two boxes on a LAN.

Honest Caveats

1. PAIR is a beta. Behaviour and hardware support can change. Every quote above was read on 15 September 2026.
2. We have not run it. This page reports what NVIDIA documents, not measured throughput on a real network. When we test it, we will publish numbers rather than adjectives.
3. We found no NVIDIA statement about model sharding either way. The FAQ addresses the virtual-GPU question and says no. It does not separately discuss splitting layers across machines, so we have not claimed it does.
4. “Compatible” is doing work in NVIDIA’s hardware list. NVIDIA notes requirements “may include specific operating system versions, NVIDIA RTX GPUs, GPU memory, system memory, processors, storage, and local network connectivity.” Check your machines before planning around it.

See Also

• vLLM tensor parallel on two GPUs — the real way to split one model across two cards
• Motherboard and CPU for a multi-GPU LLM rig — PCIe lanes decide how many cards you can actually feed
• Hybrid routing: local vs frontier — routing between your box and an API, the other axis
• vLLM vs llama.cpp vs SGLang for one user — why concurrency, not batch size 1, is where runtimes separate
• What local LLM fits my machine — the memory tier lookup
• Best local LLM for 64GB VRAM — the dual-card tier, done properly