What does it actually cost to run your own AI?
Renting a graphics card by the hour has changed the sums. Here is what it costs to run an artificial intelligence model on your own server, and the one reason to bother.
Around £80 a month for a small server, if the work is scheduled rather than live. The old assumption was that running your own artificial intelligence model meant paying for a graphics card every hour of every day, which put it out of reach of most businesses. You can now rent that card by the hour instead and only pay while the model is working. Work out how often the job runs and roughly how long each run takes, add it up, and that is your card time. Four hours a week is realistic for scheduled reporting work. The moment somebody wants to ask the model a question and wait for the answer, the card has to be there whenever they are, and the price roughly triples. That is a different product, not a bigger version of the same one.
For most businesses, the £20 a month subscription is the right answer, and it is worth saying that out loud before anything else. It stops being the right answer at exactly one point: when the information going into the model is not yours to send anywhere. That is usually where somebody says “we will run our own model then”, everyone looks up what a graphics card costs, and the conversation quietly dies.
It should not. The capability got cheap a while ago. What kept it expensive was the way it was sold, in whole months of hardware you barely used.
Why would I run my own AI?
One reason. Your data.
Not cost, because it is not cheaper. Not quality, because the big commercial models are better than anything you will run on a small server. Not speed. You do it when the information cannot leave your control, which in practice means one of three situations.
You handle records the law treats as extra sensitive, such as health, ethnicity or anything about children. You are a charity or public body that has to write down and defend how a system handles personal data before switching it on. Or your own contracts with your customers forbid sending their information to somebody else’s artificial intelligence service.
One thing it does not buy you. Hosting the model yourself does not put you outside Europe’s artificial intelligence rules, which reach any business whose system or its output is used in the European Union. Those rules govern what a system is used for and what you write down about it, and if anything, running it yourself moves more of that paperwork onto you rather than less.
Where does my data actually go?
Almost every artificial intelligence product being sold to you has a model inside it somewhere. There are only two possibilities and they are very different.
Either the seller runs that model on machines they control, which is an answer you can check, or the product reaches out to one of the large laboratories every time you use it. The second is far more common, and not automatically a bad thing. The business tiers of those services generally do not use your data to train anything, which is worth having in writing rather than assuming.
But the data still travels. And it usually travels to the United States, because that is where those laboratories keep most of their computing.
Why does the USA matter to me?
Sending data from the United Kingdom to the United States is lawful, and most of the country does it every day without thinking about it. Anyone telling you an American service is illegal is wrong, and is usually selling you the alternative. So this is not a warning. It is three things worth knowing before you decide it has nothing to do with you.
Somebody else can be made to hand your data over. American authorities can compel an American company to produce data it holds, including data it holds over here. The agreement you signed with the supplier does not stop that, and the supplier is sometimes not allowed to tell you it happened. Nobody has done anything wrong. You have simply lost the ability to answer for your own records.
The rules keep moving. Two earlier arrangements for sending personal data to the United States were struck down in court, in 2015 and again in 2020, and both times businesses had a few months to redo their paperwork. The one in use now has been challenged as well. That is not a reason to avoid American services. It is a reason to know which of your suppliers you would have to look at if it happens again.
Your clients are starting to ask you. Where is our data processed, and who else touches it, now turns up in contract schedules, tender questionnaires and professional body rules. “I am not sure” is the honest answer for most businesses, and it loses work.
For everyday work, sending data abroad is a sensible risk to take. The problem is that most businesses never took it. It was taken for them, by a product page that never mentioned it.
What does the machine cost?
Two shapes, and the gap between them is the whole point. The figures are ours, rounded, worked out in August 2026 against one United Kingdom provider for scheduled reporting work. They are a starting point for your own sums, not a quote.
Card by the hour. A modest server, with a data centre graphics card attached only while the job is actually running. The server is the fixed part of the bill and the card is the small variable part, so at a few hours of card time a week the whole thing lands around £80 a month. Double the hours and it moves by ten pounds or so, not by eighty.
Card all month. The same card, left attached so somebody can ask a question and wait for the answer. At United Kingdom rates the card on its own is comfortably more than £200 a month, before the server underneath it. That is not this product with the dial turned up, it is a different product.
There is a third shape that skips the card altogether: a larger server running a mid-sized model on its ordinary processor. It works, it is slower, and you pay for every hour of the month whether it is working or not, so it is not the bargain it looks like.
The mechanism surprised me more than any of the totals. You do not have to reserve a graphics card for a month in order to use one, and that puts capable models within reach of businesses that could never justify a permanent one.
One warning before that £80 goes anywhere near a budget.
This works because the model runs in batches on a schedule, not because it is cheap. Nobody sits waiting for it. The moment a person wants to ask the model something and get an answer straight back, the card has to be there whenever they are, and the price stops looking anything like £80. That is a different product, not a larger version of this one. Establish which of the two you are buying before anyone quotes you.
How do I work out my hours?
You do not guess at it. This is where auditing the process before you automate it pays for itself.
By the time you are pricing the server you already know which jobs the model is doing, what schedule they run on, how many times a week that happens and roughly how long a run takes. Multiply it out. That is your card time. Four hours a week is not a rule of thumb, it is what one particular set of weekly reporting jobs added up to, and yours will be different. Add more jobs, or run the same ones more often, and the figure goes up in a way you can predict rather than discover on an invoice.
Then there is the cycle itself.
Stage four is the one people leave out, and it is the one that matters. A card that fails to release turns an £80 month into a £300 one, and the failure is completely silent until the bill arrives. Anything that costs money while nobody is looking needs something watching it.
Two gaps worth naming, because they are the difference between this article and a sales page. Ask any hosting provider whether attaching and releasing a card can be driven by your automation rather than by a person clicking, and get the answer before you commit; that question is sitting with one of our own hosting providers this week. And there are smaller models that run on an ordinary processor with no graphics card at all, which would take the price down again. Whether they are good enough for real work is untested, so treat it as a maybe.
More processing cores buy you speed, not capability. A modest machine runs the same models as a large one, it just takes longer.
What should I ask a supplier?
If you are buying any product with artificial intelligence in it, five questions, and none of them are technical.
- Which model is behind this? By name and version, not “advanced AI”.
- Whose machines does it run on? Yours, or a connection out to somebody else’s.
- Which country is my data processed in, and does it stay there?
- How long is it kept, and where is that written down? The answer should be in a document, not an email.
- Is my data used to train or improve anything? Get the answer in writing either way.
A supplier who can answer all five in a sentence each is worth talking to. A supplier who cannot has not read their own supply chain, which tells you something more useful than any demonstration will. If you want the longer version of that conversation, the questions to ask before you sign covers it, and the Model Hosting Hours Estimate prompt will work out your likely card time from a list of the jobs you want the model to do.
Most businesses that ask those five questions will get five decent answers and carry on paying their £20 a month, which is the right outcome and always was. The difference is that it will have been a decision. Four hours a week is not really a number about graphics cards. It is what it looks like when somebody worked out how much of the machine they actually needed, instead of buying the whole month because that was the only thing on the menu.
- You run your own artificial intelligence model to control where the data is processed, never to save money.
- Renting a graphics card by the hour instead of by the month is what brings self-hosting within reach, at roughly £80 a month for scheduled work.
- It only works in batches. Live question and answer needs the card permanently attached and costs roughly three times as much.
- Card time comes from the audit: jobs, schedule, frequency, run length. Multiply it out rather than guessing.
- Build the check that confirms the card was released, because a card left on is an expensive silence.
- Ask any supplier which model, whose machines, which country, how long, and whether you are training anything.
Drafted by Otto, the Perkins SmartOps AI assistant. Reviewed, edited and published by David Perkins, the human.