Who pays for the AI in an AI app — and how do you price it?

October 6, 2026

In most AI apps, the app owner pays for every AI action — a charge per token, per image or per second of GPU time — and that bill grows with every use. The price users see has to cover that cost for the heaviest realistic user, plus the app store's cut, not just the cost of an average user.

Every AI action is a bill

A chat reply, a generated photo or a short video means a machine did real work, and someone pays for it. Hosted model APIs usually charge by the amount of work: text models per million tokens in and out, image models per image, video models per second of video produced. If you run models yourself on rented GPUs, you pay for the time a machine is switched on — on serverless GPU platforms, from the moment a worker starts until it stops, rounded up to the second, including a few idle seconds after each request.

Either way, the cost is tied to use. A user who opens your app once a month costs almost nothing; a user who generates all day can cost more than they pay.

Option 1: pay a provider per use

The fastest way to start. Sign up, get an API key, and a working model is in your app on day one.

You pay nothing when nobody uses the app, which matters at the start.
Cost grows with every use, including use by people on your cheapest plan.
Prices, rate limits and the model itself change on the provider's schedule. A model you built around can be retired, and the replacement may behave differently.
You keep the keys safe, watch rate limits and plan for spikes.

Option 2: run your own models

At steady volume, your own models on your own GPUs usually cost less per result, and you decide when a model changes. In exchange, you take on a second job:

Machines that wake up slowly. When a GPU worker has scaled to zero, the first request waits while it starts and loads the model.
Capacity: enough machines for the busy hour, without paying for them all night.
Keeping quality where it was every time you update a model or its settings.
Watching the bill, since idle time and retries are paid time.

Which one is right?

Neither is right for everyone. Per-use pricing usually wins while volume is low or uneven, because you only pay when something happens. Your own GPUs start to pay off once usage is steady enough to keep machines busy. Many apps mix the two — running the expensive, high-volume steps themselves and calling a hosted model for small supporting steps.

Price for the heaviest user, not the average

Average cost is the number that misleads. If most users barely touch the app and a few use it all day, the average looks cheap while the few heavy users decide whether you make money. Three common answers:

Meter it. Every AI action draws from what the user paid for — credits, a monthly allowance or a usage cap. Cost and revenue move together.
Cap it per plan. Each plan includes a fixed amount of use; heavier users move up a plan.
Unlimited. Simple to sell, but it only works when each action costs almost nothing, or when a fair-use limit quietly does the capping.

Whatever you pick, work out the cost of a plan's full allowance before you set its price, and check it again whenever a model or its price changes.

Don't forget the store's cut

If people pay inside your mobile app, the app store takes a share before you see anything. Apple's App Store Small Business Program charges 15% for developers earning up to $1 million a year. Google Play takes 15% on auto-renewing subscriptions paid through Google Play Billing — in the US, EEA and UK since June 30, 2026, split into a 10% service fee and a 5% billing fee. A plan priced to just cover AI cost on the web loses money once it is sold in an app — so the in-app price has to cover the AI cost and the store's share.

How AshWeb splits it

App owners on AshWeb hold no AI keys and never get a GPU or API bill. The models that make what users see — images, video and chat replies — run on our own GPUs, and supporting steps such as planning a post or writing a caption use a hosted model; we handle both.

The owner pays for AI with credits bought in advance, and can turn on automatic top-up. When a user pays in the owner's app and receives credits, at least that many credits come out of the owner's prepaid balance. Every AI action the user then takes draws from the user's credits. When an owner saves a plan for their users, its price is checked against what the plan's included usage costs, so a plan cannot be set below that floor.

What stays with the owner: deciding who the app is for, what people pay for, and finding those people.

FAQ

Is running your own GPUs always cheaper? No. It pays off at steady volume. At low or spiky volume, paying per use is usually cheaper and much simpler.

What is the most common pricing mistake? Pricing on average use instead of the heaviest realistic user — and forgetting the app store's share on in-app purchases.

Do app owners on AshWeb pay for AI usage? Yes — through prepaid credits, not a GPU or API bill. When users pay in the app, at least the credits they receive come out of the owner's balance, so an owner's prices have to cover that; the platform checks each plan against that floor.