What Does It Cost to Run an AI Product? Lessons from Cuplikin

Questions about app costs usually stop at "how much does it cost to build". For AI products, the more important question comes afterwards: what does it cost to run an AI product every month, and does the price users pay cover that cost?

We lived through this while designing Cuplikin, an app that turns long videos into short clips with the help of AI. This article summarises how we think about its cost structure and the lessons anyone planning an AI product or feature can use. We deliberately don't disclose our per-unit costs or margins, since those are part of the secret recipe. What we share is the way of thinking, which is more useful to you anyway.

If you haven't read the cost range for building an app, read that first. This article is about costs after the app is alive.

Five cost components of an AI product

Almost every AI product has five running cost components:

The first two are what make an AI product different from an ordinary app. In an ordinary app, a thousand extra users barely add cost. In an AI product, a thousand extra users can mean a thousand extra model calls, and each call has a price.

The biggest decision: where does compute happen?

For Cuplikin we answered this at the start, and the answer shaped the whole cost structure. The three heaviest jobs in Cuplikin's flow are audio transcription, video cutting, and rendering captions onto video. All three run on the user's device, using Whisper and FFmpeg embedded in the app.

That means the compute that would be most expensive in the cloud is carried by the user's Mac or phone. Our server only handles light, high-value work:

• AI moment analysis. The video is not uploaded; only the minimum data needed is sent to the server. Compact data is far cheaper to process than raw media.
• Accounts, credits, and subscriptions.
• YouTube link downloads via a resolver service, with temporary files deleted automatically.

With this design, the two biggest variable cost components in the table above almost disappear from our side, and user privacy is protected too, because videos and projects stay on the device.

Honest note: this approach isn't free. On-device transcription means the Mac version of Cuplikin supports Apple silicon only, and on phones the process is slower than on a server. We traded some convenience for predictable cost and a privacy promise we can keep. For another product the trade-off could be different. There's no universal answer, only a conscious decision.

Tying usage to price: credits

Once cloud cost is small, what remains is still variable: every moment analysis calls an AI model, and every call costs something. An unlimited flat subscription means a user who analyses a hundred videos a month is subsidised by one who analyses two. That's a fast way to lose money.

So Cuplikin uses credits for cloud features. Three rules we consider important:

1. Credits are only spent when the analysis succeeds. If a job fails or is cancelled, the credits are returned. Users should never pay for a system failure.
2. There's a free allowance. Free users get a monthly credit allowance to try the full flow before deciding to pay.
3. A Pro subscription gives a larger allowance, credit packs for those who need more. Subscription credits expire at the end of the period; pack credits don't expire.

We deliberately don't publish the credit amounts here or in the store description, because we adjust them as usage data comes in. What matters is the mechanism: usage and cost move together.

Server: one machine is enough

Cuplikin's backend, from API, database, and queue to the media worker, runs on a single server with Docker. For the current stage, this is a small, predictable fixed cost. Scaling up is only needed when the job queue starts to pile up, and that's a pleasant problem because it means there are many users.

What founders often forget: server cost isn't just the machine. There's monitoring, database backups, and people's time to maintain it. In client projects, we always include this in the post-release estimate.

Platform commission

Cuplikin is sold through the Mac App Store, the iOS App Store, and Google Play. Each transaction has a commission taken by the platform, within the ranges each platform publishes, typically 15 to 30 percent depending on the programme and revenue scale. This has to go into the pricing calculation from the start, not be a surprise in the first report.

We use RevenueCat to manage subscriptions across all platforms. We consider its fee well worth it compared to building cross-store subscription logic ourselves, which is complex and error-prone.

Cutting cost without cutting quality

A few things we did or learned that you can copy:

• Send as little data as possible. Process heavy media on the device, then send the AI only what it truly needs for the task, not raw files. This is a difference of tens of times in transfer and processing cost.
• Pick the model for the task. Not every task needs the biggest model. We separate tasks that need a strong model from tasks a lighter model handles well.
• Let users bring their own AI. Cuplikin users can connect an OpenAI-compatible provider, or a local model through Ollama. They carry the cost; we don't pay twice.
• Cap before you monetise. A clear free allowance is healthier than "unlimited free" that's quietly throttled later.
• Clean up temporary files. Every file stored indefinitely is a cost that keeps growing. Our YouTube resolver deletes files automatically after download.

What this means for your AI project

If you're planning a chatbot, a document assistant, or an AI feature in an existing app, four questions need answers before you set a price or budget:

1. Where does compute happen? The user's device, your own server, or a third-party API. The answer determines your variable cost.
2. What is the usage unit? Per conversation, per document, per minute of audio. One clear unit makes calculation and pricing easier.
3. What does one unit cost, and how many units per user per month? Measure with real data during the trial phase, don't guess.
4. Does the price cover your heaviest user, not just the average? If not, your most active users are the ones losing you money.

For internal projects, such as a RAG-based document assistant or a WhatsApp chatbot, the same questions apply, just replace "price" with "monthly budget".

Next steps

If you're working out the cost of running an AI feature or product, we can help build a simulation with your own business data, including the local-versus-cloud compute decision. Start with a free consultation or chat on WhatsApp. Our AI engineering services are on the services page.