Why Data Centers Are So Big: A Gigawatt Campus Costs $38 Billion, and the Land Is Under 1% of It

Nvidia agreed in August to backstop the financing for an OpenAI data center campus in Pike County, Ohio, and the campus was sized in gigawatts. It starts at 4.25 gigawatts with an option on 3.75 more. Meta's Hyperion campus in Louisiana is planned at 5 gigawatts. A single gigawatt is about what one big nuclear reactor puts out, so these are buildings that want a power plant each.

So why are data centers so big now? You can read most of the answer straight off the bill. Epoch AI (a research group that tracks AI compute) modeled a one-gigawatt AI campus in May and priced the build at $37.9 billion. The land and the utility hookup together come to under 1% of that. They're also the two things that take years to get, while the chips show up in months. So once a developer has a grid connection, the sensible move is to put as many chips behind it as the substation can carry.

What a gigawatt data center costs

Most of the $38 billion is computers. Epoch built its model on Nvidia's GB200 NVL72 racks, and it puts $21.2 billion into servers. The facility comes to $11.4 billion, and that's everything that keeps the servers alive (the shell, the electrical rooms, the cooling plant). Networking adds $4.9 billion, which pays for the switches and optics that let a few hundred thousand chips behave like one machine.

Then there's the land, which Epoch has at $172 million. The utility works come to $164 million on top of that. Put together they're smaller than one month of the rent Google pays SpaceX for GPUs in Memphis.

Other estimates land on either side of Epoch. Bernstein put a gigawatt closer to $35 billion, while Nvidia told investors last year it was $50 billion to $60 billion (Nvidia sells the priciest line on the bill, so you'd expect its number to be the big one). SpaceX's Colossus 2 is a decent real-world check. Epoch's directory has it costing $35.8 billion, and it runs about 946 megawatts of IT load. That's right around the model.

Why nobody builds a small data center anymore

The expensive part of the bill wears out fast. Epoch writes the servers off over five years and the facility over fourteen, so one building goes through two or three generations of chips. The permission to draw electricity stays with the building through all of them.

That permission is the slow part. A large substation transformer now takes more than 160 weeks to arrive, and the steel inside it comes from one American company. A new grid connection also has to clear the interconnection queue (the utility's waiting list of projects asking to plug in, studied in batches). Texas got so backed up it froze 474 gigawatts of requests. When xAI built its own 150 megawatt substation in Memphis, it took 97 days against the two and a half years a utility normally spends, and people wrote about it because it almost never happens.

Put the cost and the wait next to each other and the size follows. If the connection costs almost nothing and takes three years, you don't ask for 40 megawatts. You ask for everything the transmission line near you can carry, because you only want to wait once. It's also why Nvidia has been buying stakes in powered land developers, the companies that line up acreage with a grid agreement already signed.

Training pushes the same direction. A frontier model trains on chips that talk to each other constantly, so they need to sit in one place on one network. You can't spread half a million GPUs across ten towns and train one model on them. (Inference, the part where the model answers your questions, is more forgiving about where it sits.)

What's inside one gigawatt

A GB200 NVL72 rack holds 72 GPUs and draws about 120 kilowatts. Epoch assumes a PUE of 1.14 (power usage effectiveness, what the whole building draws divided by what the computers draw). So about 877 megawatts of a gigawatt reaches the racks. By my arithmetic that's roughly 7,300 racks, and a little over half a million GPUs. Anthropic described Colossus 1 as more than 220,000 GPUs on 300 megawatts, which is more chips per megawatt than my estimate. I'd guess the gap is older Nvidia chips in Colossus 1 that draw less apiece.

Every one of those racks has to keep running if the grid drops, so campuses line up diesel generators outside. Amazon's site in Manassas, Virginia has 93 generators at 2.5 megawatts each, with fuel on site for up to 96 hours. Virginia had permitted more than 10,500 generator units by the end of 2025. If you scale Manassas up to a full gigawatt you get around 400 of those machines (my arithmetic again), and nearly all of them exist for the day the grid fails.

People ask about water a lot, and it's the smallest line in Epoch's running costs. The model puts water at $6 million a year. The whole campus costs $8.5 billion a year to own and run. Newer AI halls cool the chips with liquid that loops through a closed system, so the same fluid goes round and round. A proposed 1.2 gigawatt campus in Delaware City put its cooling water at about 12.7 million gallons a year. Electricity costs a lot more, at $594 million a year.

Who books revenue on a $38 billion campus

Follow the bill down and the order lines have names on them. The $21.2 billion server line is mostly Nvidia, which sells the GPUs and a lot of the rack around them. The $4.9 billion networking line splits between Nvidia's own networking gear, Arista switches and Broadcom switch chips.

The facility line is where size shows up for everyone who isn't Nvidia. Each gigawatt spends $11.4 billion there, on switchgear and transformers from Eaton and GE Vernova and on the cooling loops and power distribution Vertiv sells into the hall. The generators come from Caterpillar and Cummins. Caterpillar's Power & Energy segment grew 17% to $8.2 billion in the second quarter. Cummins posted record revenue of $9.5 billion and raised its full-year growth guidance to 10% to 13%. (I didn't expect the AI trade to run through an engine maker in Columbus, Indiana.)

The landowner gets paid $172 million once. The gear makers get paid on every megawatt the connection allows, so a bigger connection is a bigger order for them.

The chips on a $38 billion campus get replaced every five years. The grid connection under them is the part that took three years.

Tags: pillar, gigawatt, data-center-costs, power, diesel-generators, nvidia