200 Megawatts: The Number Behind Every AI Product You Use

200 megawatts. That is the number. Hold it while I explain why a boda-boda rider understands it better than most people writing about AI.

One AI data center campus — not the whole company, not the whole cloud, just one cluster of servers in one location — requires 200 megawatts of continuous electrical power to run. Microsoft is building several of these. So is Google. So is every government that wants an AI programme worth taking seriously.

To give that number a body: 200 megawatts is roughly the sustained energy of 100,000 motorcycle engines running simultaneously. Not parked. Running. Going nowhere. Every hour of every day of the year.

Scene One: Morning — The Fuel Receipt

The rider tops up at six forty-seven in the morning, before the first fare. He knows his fuel cost per kilometre — calculated so many times it is automatic. He knows that a twenty-minute traffic jam burns fuel without income. He knows that if petrol goes up ten shillings per litre, his whole week shifts.

This is not accounting for accountants. It is survival arithmetic, done standing, with a helmet under one arm.

An Nvidia H100 GPU — the chip powering most serious AI models right now — draws about 700 watts continuously while working. One server rack holds eight chips: 5,600 watts. A mid-sized AI inference cluster has thousands of those racks. The fuel never stops. The meter never pauses. And unlike petrol, there is no coasting downhill to save it.

Every prompt you send to a large language model costs electricity. Every image you generate. Every customer-service chat that routes through a model backend. The power bill is real, it is large, and it is somewhere — even when the product feels free.

Scene Two: Midday — Idle Burn

The rider is behind a matatu that stopped for no visible reason. He is burning fuel, making nothing. This is the worst kind of cost: invisible in the moment, brutal when you tally the day.

AI companies have the same problem at a different scale. It is called low-utilisation inference: the model sits ready, nobody is querying it, but the servers still draw full power. The cooling systems still run. The meter does not pause for low-traffic hours.

This is why, quietly and away from the product announcements, the AI industry is in an energy procurement race. Google has signed power purchase agreements measured in gigawatt-hours. Microsoft is funding the restart of a nuclear plant specifically to run AI workloads. Amazon has signed long-term geothermal deals. These are not environmental gestures. These are fuel-cost strategies — the same logic as locking in a bulk petrol deal before prices move.

The press release talks about benchmark scores. The boardroom talks about cents per kilowatt-hour.

The company that secures cheap, reliable power in 2025 has a structural margin advantage that no better model can simply erase in 2027. A slightly dumber model that costs half as much to run will take customers from a brilliant model that charges too much to keep its lights on.

Scene Three: Evening — The Real Count

End of day. The rider writes it down: total fares, total fuel, wear on tyres and chain, time. Whatever is left is real money. Not what he grossed — what he kept. This discipline separates the riders still working in three years from the ones who burned out while feeling busy.

The same filter is coming for AI companies, and it will arrive faster than most expect. The ones who counted power costs early are now building a moat that is almost impossible to cross with engineering alone. You cannot out-research your way past a terrible electricity deal at scale.

For builders working at smaller scale, this same logic flips into an opportunity. A model running locally — on your own hardware, on your own power — costs your electricity bill, not theirs. A small, efficient model on a reasonably tuned machine draws maybe 30 to 80 watts while thinking. That is a number you can fit in a monthly budget. That is a number you can audit. That is a number you control.

This is one of the real reasons local AI is worth taking seriously — not just privacy, not just offline access, but the economics of not renting someone else's power deal at a margin that suits them, not you. Some of the experiments running out of the Ni Biashara lab are built on exactly this premise: what does it actually cost to give a small business a useful AI tool when you count the real wattage, not just the API invoice?

The Thesis

The AI economy will be shaped by power infrastructure decisions being made right now, mostly invisibly, by procurement teams in hyperscale companies that do not write blog posts about any of it.

The company that locks in the cheapest, most reliable electricity for the next decade controls the AI margin. Not the best research lab. Not the most viral product launch. The one with the best power deal.

This is not pessimistic — it is just how every infrastructure industry works. Railways ran on coal contracts. Telecoms ran on spectrum licences. Cloud computing runs on real estate and cooling costs. AI runs on watts. Infrastructure always wins eventually. The question is who owns it.

The boda-boda rider understood this at six forty-seven in the morning. He counted the fuel before the first fare. That is the move.

What to Do With This

• If you build products: Know your inference cost per query — not the API price, but the unit economics underneath. If your business model depends on current API pricing holding, it depends on someone else's power deal holding too.
• If you watch the industry: Track energy announcements alongside model announcements. A new nuclear agreement or a long-term renewable contract tells you more about a company's five-year position than any benchmark chart.
• If you are just curious: Next time a company announces a massive new model, ask one question — where is it running, and who is paying for the electricity?

The number is 200 megawatts. Now you know what it is counting.