What AI actually costs a small product: the model bill was never the big one
At our peak, the AI model bill was about US$400 a month. Servers and storage reached about US$4,000. Six real cost figures from products I built between 2019 and 2026, and where the money went each time.
When business owners ask me what AI costs, they usually mean the model: the price per question, the monthly plan. In the products I've built, that was almost never the largest bill.
At our peak, we spent about US$400 a month on AI model calls. In the same period, servers and storage reached about US$4,000 a month. The part everyone worries about was a tenth of the part nobody asks about.
Six real figures
These come from my own notes and specs, from small side projects to a funded startup.
| Year | What I was running | What it cost |
|---|---|---|
| 2019 | A small app for keeping in touch with friends | Under $60 all-in to set up, then $5 a month to host |
| 2021 | A personality-prediction service, three parts on one server | A $40-a-month server beat a $25-a-month one on the same setup |
| 2024 | A face-analysis model (the baseline in my spec for its replacement) | About $66 a month for the server |
| 2025 (as I recorded it then) | Our AI product at its peak | About US$400 a month on model calls, up to about US$4,000 on servers and storage |
| 2025 | WhatsApp messaging for an AI agent | A $19-a-month plan plus a cheap connector, instead of a $199-a-month option |
| 2026 | A team of AI agents running my company | One subscription of roughly US$100 to US$130 a month |
The pattern in that table is that the recurring AI bill stays small, while everything around it varies widely.
Where the money really went
People. At Mindreader we spent about SGD 200K on engineering over three years. No server or model bill came close. If you're budgeting an AI project, the quote for building it matters far more than the running cost.
Always-on machines. A model you call only when needed costs you per use. A server you keep running costs you every hour, busy or idle. Most of that US$4,000 was machines and storage that were on all the time.
Choosing a tier out of habit. In 2021 I tested two hosting options side by side. The $40 one outperformed the $25 one on the same setup, and the service could handle 50 people at once with answers in under four seconds. The lesson wasn't "pay more". It was that a one-hour test told me more than either provider's pricing page.
The convenient option. The $199-a-month messaging plan would have worked. So did a $19 plan with a connector on its cheapest tier. For the keeping-in-touch app, I cut hosting to $5 a month by moving its scheduled tasks onto a free tier.
How I keep the model bill small
I don't give every job the most capable model. Simple work goes to a cheaper, faster one. The stronger model does the final checks, where a mistake would reach a customer. Run that way, my whole agent team fits on one subscription.
The same goes for scheduled jobs. My recurring lead-finding jobs buy between $0.16 and $0.36 of data per run, and each has a hard cap.
Estimating your own cost
Before you build, write down three lines:
- What runs all the time? Servers, databases and storage. This is usually the biggest recurring line.
- What runs per use? Model calls and data lookups. Multiply the unit price by a realistic month, then cap it.
- Who builds and maintains it? This is the real cost. Get it in writing before you compare model prices.
If a supplier's proposal talks only about the second line, ask about the other two.