The AI Fuel Bill: What a Boda Rider Knows That Most Builders Forgot

I will start with the confession, because everyone in the trade has a version of it and most people hide theirs. For three months I rented the biggest GPU I could find, paid for it like rent on a house I never slept in, and shipped a model that was no smarter than one I could have run on a quiet machine in the corner of the office. I was paying for horsepower I was not using. The boda rider parked outside my building had already solved this problem years ago. I just had not bothered to ask him.

Watch a rider who actually makes money, not the one chasing the loudest exhaust. He does not buy the most powerful bike on the road. He buys the one that sips fuel, starts in the rain, and survives the potholes on Jogoo Road without a new chain every fortnight. His whole business lives or dies on one number nobody romanticizes: shillings of fuel per trip. Top speed is a story he tells at the stage. Fuel cost is the truth he counts at night.

Two ways to power an AI idea

There are, roughly, two ways to put compute behind your work, and the industry spends enormous energy pretending only one of them is serious.

The first approach is what I'll call chase the biggest engine. You reach for the largest frontier model on the most expensive rented cloud GPU, on the theory that more horsepower is always closer to winning. It feels responsible. It feels like you are taking the future seriously. And for certain genuinely hard problems — novel reasoning, long messy documents, code that has to be right the first time — it is the correct call. Nobody should pretend a small model can do everything.

The second approach is right-size the route. You ask what the job actually needs, pick the smallest model that clears the bar, and run it as close to the work as you can — a modest cloud instance, or a machine under your own desk. It sounds humble. It sounds like settling. It is neither. It is the rider counting fuel.

The frontier model is the bike that hits 140 on a clear highway. Your business is a market route at 7am. You will never see 140. You will see traffic, and you will pay for the engine anyway.

Where the bill actually comes from

Here is the part the demo never shows you. A data center is not magic. It is a warehouse full of chips that drink electricity and sweat heat, and somebody has to pay for both the drinking and the cooling. When you rent a giant GPU, you are not renting intelligence. You are renting a power bill wearing a friendly dashboard.

Most small-business AI tasks — drafting a reply, sorting an inbox, summarizing yesterday's orders, answering the same three customer questions — are short, repetitive trips. They are the boda equivalent of ferrying someone two stages down the road. You do not need the highway monster for that. But if you route every one of those trips through the biggest engine, you pay highway fuel for a market errand, thousands of times a day, and you call it being modern.

I added it up eventually, the way I should have on day one. The expensive setup was doing the same work as a model a fraction of the size, just slower to bankrupt me. The intelligence I was buying was real. The intelligence I was using was a rounding error on what I paid for.

Picking a side

So I will pick a side, because compare-and-contrast without a verdict is just a shrug in a nice jacket. For the overwhelming majority of builders and small businesses, right-size the route wins, and it is not close.

Not because the big models are bad. They are extraordinary. But because your costs are real, your margins are thin, and the power bill does not care how impressive your stack sounds at a meetup. The builder who survives is the one who treats compute like fuel: a cost per trip to be measured, not a flex to be admired.

The discipline looks like this. Start with the smallest model that passes your quality bar. Send only the genuinely hard trips up to the expensive engine — and prove they're hard, don't assume it. Run the boring, repetitive work somewhere cheap and close. Keep the data that matters to you on a machine you control, the way a rider keeps his own logbook instead of trusting the stage to remember his earnings.

That last point is where a private setup earns its keep — a small lab idea we keep circling back to at Ni Biashara: a quiet local notebook that does the everyday AI work on hardware you own, so the only trips that leave the building are the ones worth the fuel. Not paranoia. Just a rider who counts.

What to do now

Before you provision anything this week, do the one thing I skipped. Take your three most common AI tasks and ask the smallest model you can find to do them. If it passes, you just found your fuel savings. If it fails, now you know exactly which trips deserve the big engine — and you'll pay for horsepower on purpose, not out of habit.

The biggest engine is not the goal. Finishing the route with fuel left in the tank is. The rider outside my building never forgot that. It only took me three months and a power bill to catch up.