For the client
Our own GPU, in plain words
What it means to run the AI on hardware we control, what it costs when it is on and when it is off, the one button that switches it, and what happens if somebody forgets to switch it back.
Abad Naseer · 29 August 2026
The assistant answers residents from your association's own documents. Almost all of that work already happens on our own server. One step does not: turning the passage it found into a readable sentence is done by a language model, and today that model belongs to Google.
We can now do that step on our own hardware instead. A graphics machine, running an open model, that an administrator switches on before a meeting and which switches itself off afterwards.
Nothing about the answers changes. The same documents, the same rules, the same refusals. What changes is whose computer writes the sentence.
Three reasons, in the order they usually matter.
You can say it. "The AI runs on our own hardware" is a different sentence from "we use Google", and for some audiences it is the one that matters.
Nothing leaves the building. When the machine is on, a resident's question and the passage that answers it are handled entirely by computers we control.
It costs almost nothing when nobody is using it. Which is the whole trick, and section 4 is about it.
Press a button, wait a couple of minutes, flip a switch. That is the entire operation.
The waiting is not a wait anybody sits through. Questions asked while the machine is starting are answered normally, by Google, and the screen tells you that is what happened. Residents never see anything different.
The machine costs about eighty cents an hour while it is switched on, and nothing at all while it is switched off. Its disk costs about eight dollars a month either way.
So:
| How it is used | Roughly |
|---|---|
| A few meetings a week | $10 to $15 a month |
| Left running by accident, one month | $580 |
That gap is the only real risk in the whole idea, and it is entirely a question of whether somebody remembers to switch it off.
So nothing depends on anybody remembering. The machine stops itself after twenty minutes with no questions. And because a machine that has locked up cannot switch itself off, a second check runs outside it, on a different computer entirely, and stops anything that has been left on.
This is the part worth understanding, because it is what makes the rest safe to put in front of people.
If the switch says "our own GPU" but the machine is off, still starting, or has stopped responding, the assistant answers anyway, using Google, exactly as it does today. A resident notices nothing.
The administration screen does not hide this. It says, in as many words, that the engine is set to our GPU and that Google is currently answering, and why.
The reason for that honesty is narrow and practical. A demonstration where somebody says "this is running on our own hardware" while it quietly is not is a worse outcome than the machine being off, and it is the sort of thing that gets found out.
There is a working model of all of this on a page you can open on any device. It runs on a sped up clock, so a twenty minute idle timeout takes twenty seconds to watch.
Switch engines. Start and stop the machine. Turn the questions up. Then switch to "our own GPU" while the machine is off and watch the answers keep coming, and the panel say plainly where they came from.
Nothing on that page touches the real system, and nothing on it is charged for.
already decided before any model is involved, by searching your documents. The model only writes the sentence.
documents covers the question, no answer is written at all, by either engine. That was true before and is true now.
separate server and keeps working exactly as it does today.
forever, if that is what you want. The switch is an option, not a dependency.