AI / ML Nov 17, 2026 • 14 min read

What's actually inside a data center buildout?

Amazon, Microsoft, Alphabet and Meta are on track to spend over $700 billion on capex this year, mostly on AI data centers. I kept nodding along to that number without knowing what it buys. So I took one apart: the dirt, the wires, the chips, the pipes and the debt. There's a tiny campus you can click in the middle, and a quiz at the end where you answer in your own words.

Imagine someone hands you $40 billion and says: build me a gigawatt.

Where do you even start? You'd probably think about chips first. That's what the headlines are about. Nvidia, GPUs, H100s, Blackwell. But if you start with the chips, you lose a year. Maybe three.

The people who actually build these things start somewhere much more boring. They start with a question a farmer could answer: is there a big power line near this field?

That's the whole article in one sentence. An AI data center is an electricity problem wearing a computer costume. Once that clicked for me, every other weird thing about the buildout started making sense: why they're in West Texas, why they buy gas turbines, why the chip lasts 5 years in a building that lasts 20, and why some very smart investors are nervous.

The unit is the megawatt

Data centers used to be priced by the square foot, like an office. Not anymore. Everyone in the industry now talks in megawatts (MW) and gigawatts (GW). Ask how big a campus is and you'll hear "1.2 gigs", not "4 million square feet".

Why? Because floor space is cheap and power is scarce. A rack of Nvidia's GB200 NVL72 holds 72 GPUs and draws about 120 kW at peak. An old enterprise server rack drew 5 to 10 kW. Same steel cabinet, 15x the power. You run out of electricity long before you run out of room.

So here's what one gigawatt actually turns into:

One gigawatt, decoded
1,000 MW from the grid ≈ 830,000 homes 833 MW reaches the chips 167 MW cooling + losses (PUE 1.2) ~6,900 racks at 120 kW each ~500,000 GPUs, 72 per rack
Round numbers. A gigawatt is roughly a big nuclear reactor's output, or a mid-size city.

Hold onto that picture, because it's the whole business. Power comes in. Some of it gets wasted keeping things cool. The rest becomes math. Everything else in a buildout exists to make that pipeline bigger, faster or more reliable.

The recipe, in the order it actually happens

Here's what's inside a buildout, in roughly the order the money gets spent.

1. Land (but really, power)

Ten years ago you'd put a data center near fiber and near customers. That's why Northern Virginia is full of them. Today the first filter is: how many megawatts can I get here, and when?

The demand for that answer is wild. One Texas utility, Oncor, has received about 103 GW of requests for new load, about 82 GW of it from data centers. Texas's all-time peak demand is in the mid-80s of GW. One utility is being asked to roughly add another Texas.

So land that already has a utility promise attached, "powered land", sells for many times what the same dirt is worth as farmland. What you're really paying for is a spot in line.

2. Power in

Electricity takes a long trip to reach a GPU, and it changes shape at every step. Here's the chain:

How a watt gets from the grid to a GPU
TransmissionBig towers from the grid345 kV
→
SubstationHuge transformers step it down34.5 kV
→
SwitchgearRoutes power around campus34.5 kV
→
Hall transformerSteps down again near the building480 V
→
UPSBatteries that never let it blink480 V
→
BuswayBars over the racks480 V
→
Rack power shelfAC turns to DC~50 V DC
→
GPUFinal conversion on the board<1 V
Backup on the sideDiesel generators start in about 10 to 15 seconds if the grid drops. The UPS batteries carry the load until they do.
US voltages. Every box in this chain is something you have to buy, wait for, install and test.

Look at that chain and ask the obvious question: which box takes the longest to get?

It's the transformers. Big power transformers have been taking well over two years to arrive. They're hand-wound, enormous, and every utility in America wants them at the same time as every data center developer does.

3. Backup

A bank's servers can never go down. So traditional data centers buy two of everything ("2N") and rows of diesel generators, each a couple of megawatts. A big campus has dozens or hundreds of them.

AI training changes this a little. A training run saves checkpoints, so if power blinks you lose a few minutes, not customer money. Some AI campuses buy less backup on purpose. It's one of the few places in this whole story where someone saves money.

4. The building

The building is a big, boring, windowless box. Concrete slab, walls that are often cast flat on the ground and tilted up by crane, a roof, a fence and some very serious door locks. It's the cheapest part of the building, and the part people picture.

The expensive part is what goes inside the walls: switchgear, cable, pipes, pumps, cooling gear. Builders call it MEP (mechanical, electrical, plumbing). Cushman & Wakefield puts all-in development at about $17.6 million per MW in 2026, not counting chips, up 21% in a year.

The labor is mostly electricians. Crusoe's Stargate campus in Abilene, Texas, 8 buildings and 1.2 GW, had about 2,000 workers on site a day, heading toward 5,000. In a town of about 125,000 people.

5. The chips

This is where most of the money goes. At roughly $3.5 million per NVL72 rack including its share of networking (a rough number), the chips on a 1 GW campus cost more than everything else combined. More on that below, because it's where the risk lives.

Racks also need to talk to each other. A training cluster is only useful if 100,000 GPUs can act like one computer, so the hall is laced with fiber, switches and optical transceivers. There's more fiber inside a big AI hall than in the line coming into it.

6. Cooling

We'll get to this one. It deserves its own section.

7. Fiber out

The orange line along the road. Two routes from different directions so one backhoe can't take you offline. Training sites care less about it than you'd think. Sites that serve answers to users (inference) care a lot more, which is why they sit closer to cities.

Why can't you just plug it in?

This was my first question, and it's the right one. You have the land, you have the money. Why does it take four years?

Because the building and the power run on two different clocks, and you only make money when both are done.

One big hall, waiting for the grid
012243648 mo
Rough timeline. The building is done around month 37. The power shows up around month 48. The red line is when chips can run.

That gap is brutal. A finished building, billions of dollars, sitting dark for most of a year because a utility study and a transformer haven't shown up. Every month dark is a month the chips you ordered are getting older without earning anything.

So builders got creative. Three moves you'll see everywhere:

  • Bring your own power. Put gas turbines or big engines on site and run "behind the meter" until the grid catches up. Crusoe did it in Abilene. xAI brought in trailer-mounted turbines in Memphis to get its first cluster running fast.
  • Buy your place in line. Pay up for powered land, where the utility has already said yes.
  • Build in a factory. Whole electrical rooms and cooling skids get built off site and trucked in, so the building clock shrinks too.

And then the second-order problem: turbines are sold out too. GE Vernova's gas turbine backlog hit about 100 GW in early 2026, with new orders quoted around three years out. When everyone takes the shortcut, the shortcut gets a waiting list.

Go poke at it

I built a tiny version of a campus so you can see how the pieces sit next to each other. Click anything, or hit "Take the tour". Yellow dots are power flowing in. Blue and red dots are coolant going in cold and coming out hot. There are four calculators inside: size the campus, pick a cooling system, race the build schedule, and see whether one GPU pays for itself.

Drag to move, right-drag to turn, click the scene to zoom. On a phone, tap the pins or open it full screen. Open full screen ↗
Tiny Data Center. Toy model, not to scale. Try the "Construction" and "The bill" panels first.

Every watt in comes out as heat

Here's a fact that sounds obvious and isn't: a chip doing math is a very expensive space heater. 100 MW of electricity in means 100 MW of heat out, all day, every day. Cooling is just moving that heat outside.

You can't blow enough air across a 120 kW rack to cool it. So AI halls run liquid straight to the chips:

The cooling loop
cold plate GPU in the rack hot coolant cold coolant CDU swaps heat between two loops facility water Dry coolers big radiators, almost no water Cooling towers evaporate water, less power or heat to the sky
Two loops, so the coolant touching the chips never mixes with the water outside.

The industry keeps score with one number, PUE: total power coming in divided by power that reaches the computers. A perfect site would be 1.0. The industry average has sat above 1.5 for about six years. New AI sites aim for 1.1 to 1.2.

PUE sounds like a nerdy efficiency stat. Look at what it does to revenue, though. If the utility gives you 300 MW and not one watt more, PUE decides how many GPUs you get to plug in. Going from 1.5 to 1.15 is roughly 37,000 more GPUs on the same wire. Same land, same substation, about 30% more stuff to rent out.

The tradeoff is water. Cooling towers are power-cheap and water-hungry. Dry coolers barely touch water but burn more electricity on hot days. Which one you pick depends on whether your town is short on water or your utility is short on power. Lately, the answer is often both.

Where the money goes

Okay. Let's spend that $40 billion.

A 1 GW campus, roughly $39 billion
$17.6M per MW of IT load (Cushman & Wakefield, ex chips) ~$3.5M per 72-GPU rack, rough
How long each piece lasts
Building + power gear
~20 years
Typical lease
10 to 15 years
Chips (accounting)
5 to 6 yrs
New Nvidia generation

The red sliver is about one year.

The expensive part is the part that ages fastest.

That second chart is the one I keep coming back to. You're putting the most expensive thing (chips) inside the longest-lived thing (the building), paid for with something in between (a 10 to 15 year lease or a bond), while a better chip comes out every year or so.

Now think about a single GPU. Its share of the campus is roughly $49k of chips and networking plus $29k of building. If you rent it at $2.90 an hour and it's busy 70% of the time, it brings in about $17,800 a year. Take out power and staff and you have about $15,400 of cash. After the building's share, the chip pays for itself in about three and a half years.

Fine, if the chip is still worth renting at a good price in year four. Is it?

Nobody knows. Accountants at the big clouds usually spread server costs over 5 to 6 years. If the real useful life of a GPU turns out to be closer to 3, the same campus that looks like a money printer at 6 years loses money every year. You can try this yourself in "The bill" panel of the miniature. Drag the chip life slider from 6 to 3 and watch.

The whole buildout is a bet on two numbers nobody can see yet: the price of a GPU-hour in 2029, and how long a chip stays worth renting.

The part where I think about cycles

Howard Marks has a habit I love. When everyone agrees on something, he asks a different question: not "is this true?" but "is this already in the price?" He calls it second-level thinking. First-level: AI demand is enormous, so build data centers. Second-level: everyone else knows that too, so what happens when they all build at once?

Let's run it.

Everyone agrees demand is huge. I think it is. The four biggest spenders guided to roughly $720 to $745 billion of capex for 2026, up from about $410 billion in 2025. Some analysts expect more than $1 trillion in 2027.

So everyone builds. That's what makes the supply chain jam: transformers, turbines, electricians, powered land. Prices for all of them go up. The cost per MW jumped 21% in a year.

Then the supply arrives at once. This part already happened once in miniature. H100 rental prices were around $8 an hour at some providers during the 2023 scramble. By December 2025 a market index had them near $2.00. By late October 2026, $2.88, up 44%. Shortage, glut, shortage again, in three years.

The pendulum
us? Shortage prices spike, everyone builds, lenders say yes to everything Glut prices fall, chips sit idle, debt still has to be paid
The pendulum rarely rests in the middle. Where it is today is the hard part to see from inside.

History rhymes here. Railroads in the 1800s and fiber in the late 1990s were both real technologies with real demand. Both got overbuilt. Lots of investors lost money, and the country got cheap rails and cheap bandwidth that the next generation of companies built on. You can be right about the technology and still lose money on the buildout. Those are two different bets.

So the questions I'd ask, in Marks mode:

  • Who's lending, and on what terms? When Big Tech pays from cash flow, a bust hurts their stock. When the money is debt, a bust can break things. More of this buildout is now debt: bonds, private credit, joint ventures that keep borrowing off a company's balance sheet. Meta's Hyperion campus in Louisiana used a roughly $27 billion deal with Blue Owl.
  • What has to be true? GPU-hour prices stay healthy for years, the chips last at least 5 years, and demand grows faster than the supply everyone is building at once.
  • What happens if AI gets way more efficient? In January 2025, DeepSeek showed a cheaper way to train a strong model and Nvidia lost about $590 billion of market value in a day. The bull answer is that cheaper AI means way more AI gets used, so demand goes up. The bear answer is that the same work needs fewer chips. Both can be true at different times.

I don't know where the pendulum is. I'm pretty sure nobody does with any confidence, which is itself useful to know. What I do know now is what the money buys, and what has to go right for it to pay off.

What I took away

  1. It's an electricity business. The unit is the megawatt. The bottleneck is the grid. The chips are the expensive passenger.
  2. Time is the hidden cost. A building that waits a year for power is a year of aging chips and interest.
  3. Cooling is revenue. Better PUE means more GPUs on the same wire.
  4. The mismatch is the risk. 20-year buildings, 15-year leases and 5-year chips, financed with more and more debt.
  5. Being right about AI is a different bet from being right about the buildout.

Check your thinking

No multiple choice. Type your answer in your own words, like you're texting a friend who asked. You'll get a nudge back: what you nailed, and a question about what you might be missing. Try again as many times as you want, or peek at how I'd answer it. Nothing leaves your browser.

Last one

In one or two sentences: what's the most surprising thing you learned about data centers today?

Share on X

If you think I got something wrong, especially on the money side, DM me on X. I'd rather learn it now than in the next cycle.

Tags:
AI / ML Infrastructure Investing
Published Nov 17, 2026