# Your AI Model Lives on One Computer. Your App Lives on Another. Now What?

You have a few computers that can run AI models. The app that wants an answer is usually on a different computer. NVIDIA PAIR and a hosted service both solve that by sending the whole question to one computer that already has the model.

They do not glue several graphics cards into one bigger card. A question is never split across machines. If you want two computers to share the work, you put a copy of the same model on both, and each question goes to one of them.

The difference is where those computers are allowed to live.

NVIDIA PAIR is for computers on the same office network. The hosted service is for a customer whose computers, and the people using them, are in different places. The people asking questions never hear a product name. They get a web address and a key.

## PAIR: several computers, one office

Picture a laptop, a desktop, and a studio PC, all on the same network. You type on the laptop. The desktop and the studio PC have the big model. A smaller PC on the same network only has a small model.

You want the app on the laptop to keep using a normal local address, the kind Ollama, LM Studio, or llama.cpp already use. PAIR decides which of the other computers should answer.

Every computer runs the same PAIR software. There is no central server. You pair them with a short code. After that, they only trust each other. The app talks to PAIR on the laptop. PAIR looks at which computers are on, which ones are already busy, and which ones actually have the model you asked for. It sends the whole question to one of them. The answer streams back. The app just sees a normal reply.

![](https://cdn.hashnode.com/uploads/covers/663395db8aebfc729a69d3d4/8eb74d66-ef0f-44b7-a48c-6f9f642e9933.png align="center")

Follow the thick arrow. That is the only computer that sees this question.

1.  The app talks to PAIR on the same laptop. That first step never leaves the laptop.
    
2.  PAIR on the laptop chooses. It does not ask the desktop to pass the question along to someone else.
    
3.  The question crosses the office network once, encrypted, to the studio PC. A computer you did not pair cannot take that step, even if it is on the same Wi-Fi.
    
4.  The studio PC runs the model. The program that runs the model is not open to the rest of the network. Neighbors have to go through PAIR.
    

A second question can go to the desktop instead, because that machine also has the big model. A question that names the small model goes to the small PC. The computers are interchangeable only for the models they share.

This fits when you are sitting at one of the paired computers and every graphics card is on that same network. A laptop in a coffee shop cannot point at the office and get in. The questions stay in the office.

## Several computers, and the callers are somewhere else

Now the computers with the models and the people asking questions are not in the same place, and they are not the same person.

A customer has a home PC, an office PC, and a spare PC. Each one can run models. They may sit behind a normal home router, with no public address. The customer also has people who need answers: their own product, and a teammate with a script. Those people should reach this customer's computers only. A second customer's computers should be invisible to them.

A service in the middle connects them. Callers never see that service's name, and they never see which PC answered.

The customer has an account. With it they do two things.

**Add computers.** They create a short code that expires quickly and can be used once. A small program on each PC uses that code, then opens a connection out to the service and keeps it open. The PC does not need anyone on the internet to be able to connect in. The program on the PC only talks to Ollama on that same PC. It tells the service which models are installed and how busy the PC is. If the connection drops, it calls back.

**Add callers.** Each caller gets their own key. With that key they can ask which models are available and send a normal chat request. The service picks one of this customer's computers that is online and has that model, preferring the one with the least work already going. The answer streams back. A caller's key cannot add computers or manage other callers. Turning a key off cuts that person or that PC off.

![](https://cdn.hashnode.com/uploads/covers/663395db8aebfc729a69d3d4/1d96d60c-dbab-4849-b889-e27c36c8668a.png align="center")

Read the thick arrows as two separate questions, one for each customer. A's question never lands on B's computers.

1.  Before any question is asked, each PC calls out and leaves the line open. That is why a computer at home can answer someone in another city. Nobody on the internet can open a connection straight to Ollama.
    
2.  A's product, or A's teammate, sends a normal chat request to the service. The key says which customer they belong to. They never see the home address of the PC.
    
3.  The service checks the key and looks only at that customer's computers. It sends this question to the home PC, because the office PC is busier and the spare PC does not have the model.
    
4.  The home PC runs it on local Ollama. The answer comes back the same way: PC, service, caller.
    

Customer B is connected to the same service, with their own lab PC and desk PC. Their app gets one of those. There is no line between A's computers and B's computers.

The service has to see the question in order to pass it on. Customers are kept apart from each other. The service itself is on the path.

## Side by side

|  | NVIDIA PAIR | The hosted service |
| --- | --- | --- |
| The setup | Several computers on one office network, all running PAIR. | A customer's computers call out to a shared service. Callers use a key. |
| Who can ask | An app on one of the paired computers. | Anyone with that customer's key, from anywhere. |
| How a question picks a computer | The computer you are sitting at chooses one paired machine that has the model and is less busy. | The service chooses one of that customer's online computers that has the model and has the least work going. |
| What the other computers are doing | They can take the next question, if they have the same model. | Same, inside that customer's group. Another customer's computers are a separate group. |
| Where the question travels | Across the office network, encrypted, then back. | To the service, then down the line the PC already opened. |
| What runs the model | Ollama, LM Studio, or llama.cpp. PAIR can install and start them. | Ollama, already running on the PC. |

## What extra computers do not give you

Three computers do not become one bigger computer. A large model still has to fit on the one machine asked to run it. Copying it onto the home PC and the office PC is what lets those two share the queue. One copy means every question for that model waits on that one machine.

PAIR looks at how many jobs are waiting and a rough sense of how hard the graphics card is working, and it can try the next computer if the first one says it does not have the model. The hosted service looks at how many questions it has already handed to each connected PC. Either way, a long answer can still land on a computer that looked free a moment earlier. Neither system asks "is this model already loaded in memory?" before it chooses.

## Drawbacks

**PAIR only works inside the office.** Every computer has to be on the same network, and the app has to run on one of the paired computers. A script on a server, or a teammate at home, cannot send a question in. You also install and run PAIR on each machine. The privacy is real: the question stays on that network. The reach stops at that network too.

**Callers of the hosted service know nothing about it.** They are handed a web address and a key, the same kind of chat request they already send to any model host. They are not told that their question was routed, which of your PCs ran it, or that a PC at home dropped offline. When something fails, the error looks like the model host failed. Your customer cannot see that the office PC was busy and the home PC disconnected.

That invisibility has a cost for you as well:

*   The service reads every question and every answer as it passes through. Your customers do not. You are trusting the service with the text.
    
*   If a PC drops while it is answering, that question fails. It is not moved to the next PC that has the same model.
    
*   "Least busy" means the fewest questions already handed out. A PC with one long answer looks freer than a PC with two short ones. A free graphics card, or a model already loaded in memory, does not change the choice.
    
*   Each PC has to keep its line open. Asleep, offline, or disconnected means it is out of the pool. Callers have no address they can use to reach that PC themselves.
    

## Which one fits

Use PAIR when the computers are in one place, you are sitting at one of them, and the question should stay on that network. The app keeps talking to a local address. PAIR makes the other computers in the room available behind it.

Use the hosted service when the computers belong to a customer and the people asking questions are somewhere else. Each PC phones home. Each caller gets a key that works only on that customer's computers, and that is all they know. The service is what connects the two sides, so the question passes through it.