What's behind an AI data center

Every time you ask an AI a question, a data center somewhere does the work. Ask people who makes money from that building, and almost everyone gives the same one-word answer: Nvidia.

They're right. Nvidia just reported $96.2 billion in revenue for one quarter, roughly double a year earlier.

But Nvidia is only the first layer. Go down one level, then another, then another, and you end up somewhere nobody expects: a seasoning company, a dish-soap company, a nuclear plant with a famous name, and a shortage of electricians.

Let's go down.

Nvidia doesn't actually make chips

You probably didn't know this: Nvidia designs its chips, but it doesn't build them. Almost every top AI chip in the world is manufactured by one company in Taiwan, TSMC.

And TSMC can't make them alone either. The most advanced chips are printed with extreme ultraviolet (EUV) light, and only one company on Earth makes those machines: ASML, in the Netherlands. Each machine is the size of a bus. ASML's sales just came in above its own forecast, and it plans to raise its capacity for these machines by 30% for 2027.

So the most valuable chip company in the world depends on one factory company, which depends on one machine company.

The chip is useless without its memory

An AI chip needs a special kind of stacked memory sitting right next to it, called HBM. Three companies make it: SK Hynix, Samsung and Micron.

Here's how tight memory is: on September 7, Samsung and SK Hynix were holding less than 10 days of finished memory in stock, lower than in any earlier shortage. SK Hynix and Micron have already sold all the HBM they can make this year.

The squeeze reaches ordinary gadgets too. Memory makers are sending less to phone and PC makers, and that memory is expected to keep getting more expensive.

A seasoning company holds the whole thing together

This one surprises almost everyone.

The chip and its memory have to be joined on a base layer, and that step is called packaging. This summer, demand for TSMC's advanced packaging was running about 20% ahead of what it could supply.

That base layer is built with a thin insulating film. About 95% of that film comes from Ajinomoto, the Japanese company best known for MSG.

It's not a coincidence. Ajinomoto says the film traces back to a byproduct of making MSG. From there its chemists moved into resins and electronic materials, and the film went on sale in 1999.

The chips talk to each other with light

An AI data center isn't one chip. It's tens of thousands of chips working as one. Much of their talking runs through fiber-optic cables, driven by tiny lasers.

You probably didn't know this either: Nvidia was so worried about running out of lasers that in March it put $2 billion each into Lumentum and Coherent, two laser makers, to secure its supply.

Then go one level deeper. Those lasers are built on a material called indium phosphide, and that's short too. Lumentum's CEO says the shortage is now worse than the memory shortage.

Nvidia isn't the only chip in town

Google, Meta and OpenAI don't want to depend only on Nvidia, so they design their own AI chips. The company that helps build them is Broadcom.

Broadcom's AI chip sales just grew 221% in a year, to $16.7 billion in one quarter. Its biggest custom-chip customer next year is expected to be Anthropic, which buys Google's chips: 5 gigawatts' worth in 2027 and another 10 gigawatts in 2028. Those 10 gigawatts alone are roughly the output of ten large nuclear reactors, for one AI company.

Now you have the chips. You can't turn them on.

This is where the story flips. The hardest part of an AI data center today isn't the chips. It's the electricity.

Three companies make most of the world's large gas turbines: GE Vernova, Siemens Energy and Mitsubishi Heavy. By this spring, GE Vernova had 100 gigawatts of turbines under contract, and its production slots are nearly full through 2030.

Then there are transformers, the big boxes that step grid power down so a building can use it. The wait for a large one can now be as long as four years, and the US depends heavily on imports. The result: close to half of the US data centers planned for this year are expected to be delayed or canceled, mainly because transformers, switchgear and batteries are in short supply.

So they bring their own power

If the grid can't deliver in time, tech companies are working around it.

Oracle signed up for up to 2.8 gigawatts of fuel cells from Bloom Energy. The fuel cells sit next to the building and turn natural gas into electricity.

And nuclear is back. Constellation Energy is restarting a reactor at Three Mile Island. Yes, that Three Mile Island, site of America's most famous nuclear accident in 1979. The reactor coming back is Unit 1, which sits next to the one that failed and was not affected by the accident. It's expected back in service in 2027, and Microsoft signed a 20-year deal to buy its power. Meta has signed nuclear deals that could total more than 6 gigawatts.

Every wire is copper

All that power has to travel through metal. One estimate puts an AI data center at 30 to 47 tonnes of copper for every megawatt, counting only what's inside the building. For a one-gigawatt campus, that's tens of thousands of tonnes. On September 7, copper hit its highest price ever. That was pushed by tariff fears, on top of mines struggling to keep up with data centers and power grids. That's good business for copper miners like Freeport-McMoRan.

The dish-soap company went into AI

All that electricity turns into heat, and the newest AI chips run too hot to cool with air alone. So new AI data centers pipe liquid straight onto the chips.

Guess who wanted in: Ecolab, the company behind the cleaning and sanitizing products in restaurant kitchens and hotels. In July it closed a $4.75 billion purchase of CoolIT, a company that makes liquid cooling for data centers.

The other big name here is Vertiv, which supplies power and cooling equipment for these buildings. In the last quarter of 2025, its orders were up about 252% from a year earlier, and its backlog had reached $15 billion.

The last shortage is people

You can buy every part above and still be stuck, because someone has to wire the building.

US data center construction may need 349,000 to 499,000 more workers this year. Electrical work is about half of all the labor on these projects, and a lack of electricians is forcing builders to turn work away. Contractors like Comfort Systems, EMCOR and Sterling have more work than crews. Comfort Systems' finance chief put it simply: there's "plenty more work we could take" if they had the people.

So whose data center is it?

When you read "Meta is building a data center," it means Meta decided to build it and will run its AI there. It doesn't always mean Meta owns the building.

• Meta's Hyperion campus in Louisiana is Meta's AI data center, but funds managed by Blue Owl own 80% of it. Meta keeps 20% and leases the whole campus.
• OpenAI's Stargate site in Abilene, Texas is leased by Oracle for 15 years, and Oracle rents the chips inside to OpenAI.

In both cases, investors own the buildings and collect rent. The tech companies sign the long leases, fill the buildings with chips and run the AI.

Nvidia doesn't build data centers. It helps pay for them.

One last reveal, back where we started.

Nvidia doesn't own the giant data centers its customers run. It sells the chips inside them. But look at where some of its buyers' money comes from:

• This year Nvidia put $30 billion into OpenAI, one of its biggest chip buyers.
• It owns shares in CoreWeave, a company that rents out Nvidia chips, and it must buy any computing power CoreWeave can't sell, through 2032.
• In August, Nvidia agreed to back up to $105 billion of lease obligations on a giant OpenAI data center campus in Ohio. Nvidia pays only if OpenAI can't. The buildings will run Nvidia chips.

So some of the money goes in a circle. Nvidia invests in OpenAI, OpenAI pays for computing power, and that money flows back into Nvidia chips. The economist Noah Smith compares it to vendor financing: a supplier putting up money so its customers can keep buying its product.

Everyone else borrows

The 14 largest publicly owned data center operators are expected to spend close to $750 billion this year, up from a little under $450 billion last year.

A lot of it is borrowed, and often not by the tech company itself but by the building's owners. An Oracle and OpenAI campus in Michigan needed $16.3 billion of financing, mostly bonds backed by the building itself. The bond fund PIMCO bought about $10 billion of them, after US banks had been pulling back from Oracle data center deals.

So the next time a headline says a company is spending $50 billion on AI, picture what that money actually buys. Machines from one company in the Netherlands. Memory that's already sold out for the year. A film made by a seasoning company. Turbines whose factory slots are full through 2030. Copper at the highest price ever paid for it. Electricians who are already booked.

AI sounds like software. It gets paid for in steel, copper, gas turbines and 20-year leases. That's why the buildout can't simply speed up when the money shows up, and it's why so many companies you've never thought of are now riding on it.