Inside a data center
When you ask an AI a question, your words travel to a warehouse full of computers, often hundreds of miles away. Here's what's in that building and why it needs so much power.
Watch: the narrated video (3 minutes)
Plain-English explainer by Joe Miller, a nuclear engineer. Not investment advice. Figures are the latest public numbers we could find and change fast; sources are at the bottom of each page.
Your question's trip
- You type a question on your phone. It travels over your internet connection and fiber-optic cables to the nearest data center that runs that AI.
- A computer called a load balancer hands it to a server with free time.
- Racks of GPUs (graphics processing units, the chips Nvidia is famous for) run the AI model: trillions of quick math steps.
- The answer streams back to you, usually in a few seconds. The electricity used becomes heat, and the cooling system carries it away.
The parts of an AI data center
| Part | What it does | Why it matters |
|---|---|---|
| Substation | Connects the campus to high-voltage power lines | Big campuses need their own, and new lines can take years |
| Transformers | Step voltage down for the building | There's a long waiting list to buy large ones |
| UPS batteries | Keep everything running for the seconds it takes generators to start | A blink of the grid can't be allowed to crash the servers |
| Backup generators | Big diesel or gas engines that run during outages | Often dozens per site; neighbors notice the noise and fumes |
| Server racks | Cabinets of computers and AI chips | The reason the building exists |
| Cooling | Air handlers, chillers, cooling towers and liquid loops | Can use large amounts of water in hot, dry places |
| Fiber | Thick bundles of fiber-optic cable | Data centers cluster where fiber and power meet |
Why AI data centers are different
For years, a data center was mostly storage and websites. AI changed that. Training and running big models takes huge numbers of GPUs working together, packed tightly so they can talk to each other fast. That packs far more power, and far more heat, into each rack.
Cooling: almost every watt becomes heat
- Air cooling: cold air blown through the racks, with the hot air pulled out and chilled. Fine for ordinary servers.
- Liquid cooling: cool liquid piped through plates sitting right on the chips. Most new AI halls need it.
- Evaporative cooling towers use water to dump heat outdoors cheaply, which is why water use is a growing fight in dry states.
How much power does one question take?
OpenAI says an average ChatGPT question uses about 0.34 watt-hours, about what an oven uses in one second. That's tiny. But billions of questions a day, plus months of training each new model on tens of thousands of chips, add up to the power plants we cover in your electric bill and Big Tech's nuclear deals.
How big are they?
| Kind | Typical power | Like |
|---|---|---|
| A server closet | a few kilowatts | a few homes |
| A traditional data center | 5 to 50 megawatts | a small town |
| A hyperscale AI campus | 100 to over 1,000 megawatts | a mid-size city |
| The biggest planned campuses | several gigawatts | several big power plants |
One megawatt is roughly the average use of 800 U.S. homes. Campuses of several gigawatts are now being proposed.
Next: Will AI data centers raise your electric bill?
Sources
Checked November 14, 2026.