The Invisible Trillion

Data Centers: The Cathedrals Nobody Visits

A Tour of the Place Your Internet Actually Lives

When you open Instagram, the photos you see do not come from your phone. They come from a building.

You probably know this in some abstract sense. “The cloud” is a metaphor; “the internet” is a metaphor; everyone has heard that this stuff is stored somewhere. But almost no one has a concrete image of where. The cloud is nowhere in particular, in the same way that money is nowhere in particular. It’s just… there. Out there. In the internet.

The internet, in fact, is a very specific set of buildings. They have addresses. They have parking lots. They have security guards who have badges and bad coffee. If you drove to one, you would see a beige rectangle the size of a big-box store, with a fenced perimeter and a sign at the gate that said something like “Data Campus 4B” and gave you no information about what was inside. You could not go in. There would not be a visitors’ center. There would not even be a receptionist — the building is mostly not staffed.

Inside that building, in rows, are the machines that serve you Instagram. Also Google, YouTube, Netflix, Gmail, Spotify, every video game with online multiplayer, every corporate database in the world, every hospital electronic health record, every airline reservation system, every bank’s back-end, and — the part of the story that has gotten big in the last few years — all of the AI models that the world now runs on.

This chapter is a tour of one of those buildings. Since you can’t actually go inside, I’m going to try to take you there in prose. It’s worth knowing what this place is like, because it’s where most of the things the rest of the book argues about actually live.

Where They Are

Data centers are not random. They cluster.

The biggest cluster in the world is in Northern Virginia, specifically Loudoun County, which sits just west of Washington, D.C. About 70% of the world’s internet traffic passes through Loudoun County on its way to somewhere else. If you have used the internet today — any service, anywhere — there is a good chance your data went through Loudoun County at some point. The county has something like 200 data centers in a roughly 30-square-mile area. The combined electrical load is larger than the cities of Seattle and San Francisco combined. Driving through the area, you see mostly: beige rectangles, fences, transformers, and fiber-optic utility markings along every road.

Other major clusters, in rough order of size: Dallas, Phoenix, the Pacific Northwest (Oregon and Washington), Santa Clara (Silicon Valley proper), Chicago, Atlanta, Dublin (in Ireland), London, Frankfurt, Amsterdam, Singapore, Tokyo, Sydney. Iceland has a small but growing cluster because it has cheap geothermal power and is cold. Norway has one for the same reasons. There is a growing cluster in Abu Dhabi that’s being built for Middle Eastern AI services. There’s talk of one in Greenland.

Data centers cluster for reasons that are mostly not obvious. You might think “they should spread out, because bandwidth.” In fact they cluster, because:

Fiber optic cables are not uniform. Some places have vastly more fiber connectivity than others. Loudoun County has more fiber than many small countries, because it was an early hub and every subsequent network built to connect to the existing networks. Once fiber is there, more fiber comes.
Power is geographically uneven. Data centers need enormous amounts of reliable electricity. Places with cheap power and stable grids (the Pacific Northwest, with its hydroelectric dams; Virginia, with its nuclear plants; Ireland, with its wind) draw more data centers. Places with expensive or unreliable power (California, most of the Northeast) draw fewer.
Cool climates are cheaper to run. Cooling is a huge fraction of a data center’s operating cost. A data center in Ireland or Finland can use outside air for most of the year and barely needs active cooling; a data center in Phoenix has to actively refrigerate 24/7.
Proximity matters for some workloads. Financial trading companies pay enormous amounts to put their servers physically close to stock exchange servers, because even a few milliseconds of network delay can cost them money. This is why lower Manhattan and parts of New Jersey have specialized financial data centers.
Tax incentives. State and local governments offer data center operators huge tax breaks to locate in their jurisdiction. Virginia’s data center tax treatment is famously generous; this is part of why Loudoun County has what it has.
Political stability and rule of law. Data centers hold enormously valuable data. Companies are not going to put them in places where the government might seize them, or where they might be destroyed in a war.

Add all these forces together, and you get: a handful of places with most of the world’s data centers, and a long tail of smaller clusters everywhere else.

All 19 chapters of The Invisible Trillion · Library · All courses