In short
- Your own hours are the largest cost. Cost breakdowns that say otherwise count your time as free.
- You do not need a GPU to start. Hosted models bill per image, and the GPU decision can wait until month three.
- Almost everyone who gives up stops at one of three points: the training set, the first policy change or the volume wall.
- No one can know in advance whether this pays. You can know today what the experiment costs.
On this page
Budget for three things: compute, tools and your own hours. Expect your hours to cost far more than the other two. Most published cost breakdowns show a total that is too low, because they count the subscriptions and count your time as free.
Some decisions come before the money: which country’s rules apply to you, which platform you use, and in which order you decide. They are in getting started with AI personas.
The current price per image and per video, from cheapest to most expensive, is on what generation costs. Those prices change every month, so they are kept up to date there and not in this article.
This article is about costs. It is not about earnings. Nothing below predicts revenue, and you should distrust anything in this field that predicts yours.
Compute, tools and the platform’s cut
Compute: three ways to make images. Most guides only mention two.
The first way is to buy no hardware at all and generate on hosted models. Higgsfield, WaveSpeed and similar services run the current model families in a browser tab. They bill per image or by subscription. One image costs between half a cent and fourteen cents, depending on the model. A first month of serious testing then costs a two-figure sum, not a four-figure one. How to use WaveSpeed AI shows how to check the price of a batch before you run it.
Most persona work starts here now. The reason to start here: you can find out whether you can hold an audience before you spend money on keeping a face consistent.
The other two ways are buying a GPU and renting one. They only matter once you decide to train your own LoRALoRA A LoRA is a small file trained on images of one subject. It teaches a general image model to reproduce that specific face or style. Open the glossary, which is a small add-on model trained on images of your character. Buying means a graphics card with enough VRAM to train without trouble. That is a real four-figure decision. Renting means you pay an hourly rate. Each session looks cheap. By the twentieth retrain, it no longer is.
Owning and renting change how you work. When you own the card, another try costs nothing, so you try again and again. Trying again and again is how you get a consistent character. When you rent, every try costs money. That pushes you to accept a result that is only nearly right. This effect matters more than the price difference. But you only need to decide in month three.
Tools. You need tools to generate, a tool for scheduling and storage for a library that grows faster than you expect. Each one is small. Together they are a cost that comes back every month. Write that number down, so you do not overlook it. Which parts of the stack cost nothing, and what the free licences rule out, is in make an AI influencer for free.
The platform’s cut. Your platform takes a share. Add payment processing. Add the cost of currency conversion if you are paid in a currency you do not spend.
Your hours are the biggest cost
Your hours are the biggest cost. Generating images is the fun part, and it takes the smallest share of them.
The hours go into building a training set that does not spoil the model. Three weeks later you find out that it did anyway, and you retrain. The hours go into writing captions that do not read as generated, into scheduling, and into reading platform rules that changed while you were not looking. And they go into answering the same question in DMs for the fortieth time.
Decide what one hour of your time is worth. Multiply it by your hours, and compare the result with the subscription costs. The result is usually higher than you hoped. It is still more useful than a budget without it.
Is buying a GPU worth it, or should I rent for the first months? Ask it in #question-thread. Someone there has usually had the same problem.
The three points where most people give up
Almost everyone who gives up stops at one of three points. People do not plan for any of them.
One: the training set. You collect a few dozen images, train, and get a face that is nearly your character. Nearly is not close enough. Viewers cannot say what changed from one post to the next, but they notice it. To fix this you have to go back to the dataset. That feels like going backwards, and a lot of people stop at this point.
Two: the first policy change. You build a workflow around how one platform behaves, and then the platform changes. This happens often in this field.
Instagram replaced its “AI creator” label with an “AI-generated profile” label on 31 August 2026, and accounts depicting an AI-generated person that do not set it receive reduced distribution to non-followers on recommendation surfaces.
Instagram Creators — AI-generated profile label (opens in a new tab)checked
That single change broke a lot of working setups overnight. There will be more changes like it, so plan for them.
Three: the volume wall. A workflow built for a few posts a week does not work for a daily schedule. People find this out after they have publicly promised a daily schedule. Most attempts end there.
What did not work
Budgeting for the tools and treating my own time as free
I worked out the cost of a first persona from the subscriptions alone and decided it was cheap. Six weeks later I added up my hours honestly. The real cost was several times higher, at any wage I would accept from someone else.
This mattered for more than the numbers. Because I had told myself it was cheap, I kept going long past the point where I would have stopped and rethought the approach. The time was already spent, and I had hidden that cost from myself by not counting it. The number was not the problem. The problem was that I did not write it down.
How to decide if you can afford to start
Do not try to work out whether this pays. You cannot know that in advance, and the people who sell courses about it cannot know it either.
Work out what it costs to find out. Add up three months of tools, the compute option you picked, and your honest weekly hours multiplied by twelve. You can know that number today. If you can lose that amount and it does not hurt you, you can afford the experiment. If you cannot, no forecast of returns makes it a good idea.
The other half of the calculation is what a month of revenue leaves you after the platform’s cut, and when you can withdraw it. That is on the payout calculator.
Following the rules has a cost too, and people usually find that out late. Read what the EU AI Act asks for before you build a workflow that you will have to take apart again. If you want to know whether you can keep up the weekly hours at all, read where the hours go next.
Sources for this article
The platform rules, laws and numbers above come from these documents. The date says when we last opened each one and checked it.
Questions people ask
Can I start without a GPU?
Yes, and most people now should. Hosted models on Higgsfield or WaveSpeed run in a browser and bill per image. A first month then costs a two-figure sum and not a four-figure one. You only need to think about a GPU once you decide to train your own LoRA. Make that decision in month three, not in week one.
What is the single largest cost?
Your hours, by a wide margin, whatever realistic value you put on your time. Every published cost breakdown that says otherwise has counted your time as free.
How long before it earns anything?
No one can know that. Anyone who gives you a number is guessing or selling something. What you can know is the cost of finding out. So decide whether you can afford the experiment. Do not try to decide whether it will work.
Is a subscription tool cheaper than running things locally?
For the first month, almost always. After that it depends on how much you produce. Work out the switch point for your own output. Do not take it from a comparison table.