If the AI Bubble Pops, This Is What Survives

0:00 · $750 billion in one year

Seven hundred and fifty billion dollars.

That's how much four companies, Meta, Amazon, Microsoft and Google, are expected to spend on buildings and equipment this year alone. Mostly on AI data centres.

It's more than the entire economy of Ireland. It's enough to write a cheque for five and a half thousand dollars to every household in America.

And a growing number of people think it's a bubble.

But here's the thing about bubbles. You can't hold one. You can hold what the money buys. So instead of asking whether the stocks go up or down, let's do what engineers do. Let's take one apart.

0:52 · What a gigawatt is

First, a unit. The new AI campuses are measured not in servers, but in gigawatts.

One gigawatt is enough electricity to power about seven hundred and fifty thousand homes at once. Or, if you watched our last video, about seven hundred and fifty Tesla Roadsters, all running flat out at the same time.

Over the next four years, more than sixty gigawatts of this kind of capacity is expected to be added in the United States alone. That would triple what exists today.

1:28 · Exploded view: where $30B goes

So what does one gigawatt cost? An estimate for a site near Abilene, Texas, where some of the world's largest computing campuses are going up, puts it at around thirty billion dollars, at the high end.

Let's draw that thirty billion as an exploded view.

Balloon one: the chips. The servers, the GPUs, the memory, the storage. By McKinsey's global estimate, this is roughly two thirds of all the money. It's the part everyone talks about.

Balloon two: power and cooling. Turbines, transformers, switchgear, cables, chillers, pumps. This is where the copper lives.

Balloon three: the shell. Land, concrete, steel and the people who pour and weld it.

Now here's the first surprise. These three balloons age at completely different speeds.

2:28 · The chips age fastest

Start with the chips.

The chips are the most expensive part, and the fastest to become old. A new generation of AI hardware arrives every year or two, and each one makes the last look slow. Memory chips used in data centres now cost more than five times what they did a year ago.

The research firm Gartner put it bluntly this month: spending on rented AI computing will grow too slowly for most providers to recover their investment before the current generation of chips is replaced.

In other words, the most expensive part of the building may be obsolete before it has paid for itself.

3:12 · Gas turbines: sold out to 2030

Now walk outside, to the part nobody photographs.

A gigawatt has to come from somewhere, and in America, a lot of it is coming from natural gas. That means gas turbines: essentially a jet engine bolted to the ground, spinning a generator.

The biggest maker, GE Vernova, had a hundred gigawatts of turbines on order or reserved at the start of the year. By July, it was a hundred and sixteen. It expects at least a hundred and twenty-five by December, and says it's mostly sold out through twenty thirty.

Sold out. If you want a large gas turbine today, you're joining a queue that ends next decade.

3:59 · The 2.5-year wait for a transformer

Then there's the transformer.

Every bit of power that enters a data centre goes through one. It's a big steel tank full of oil, with copper coils wound around a core of special steel. It steps the grid's voltage down to something the building can use.

It has almost no moving parts, and it is one of the hardest things on the site to get.

The average wait for a large power transformer in the US was about two and a half years in the latest widely cited survey, and the bigger units that connect power plants to the grid took even longer. That survey also estimated the US is short by around thirty percent of the power transformers it needs.

So you get a strange picture. A building full of chips that might be outdated in a few years, waiting for a transformer that took two and a half years to arrive.

4:58 · Water and the neighbours

Then there's water. A typical forty-megawatt data centre can use around sixty thousand gallons a day to stay cool. And there are the neighbours. Just last month, Oracle issued a force majeure notice on a giant New Mexico data centre, because a gas pipeline it needed was held up by regulators.

The concrete, the copper and the turbines are now the bottleneck. Not the chips.

5:28 · The $5 trillion gap

So, is it a bubble?

Here's the maths Gartner published this month. To pay for everything that's been announced, plus the interest, plus a normal twelve percent return, AI data-centre owners would need to earn about eight and a half trillion dollars in revenue by twenty thirty.

Its forecast for the AI revenue those data centres will actually earn? No more than three point six trillion.

That's a gap of nearly five trillion dollars. Gartner thinks more than three and a half trillion dollars of this infrastructure could face write-downs by twenty thirty. And it predicts that more than eighty percent of the smaller AI cloud companies, the so-called neoclouds, won't survive on their own.

A lot of this is being built with borrowed money. One credit tracker counts at least one point three trillion dollars of debt already raised for data centres. And as one Columbia economist put it, with that much debt, you don't need a big shock. Ten percent will wipe out the equity.

6:40 · We've been here before: dot-com fibre

But here's the twist. We've been here before.

In the late nineteen nineties, telecom companies raced to bury fibre-optic cable across America. They dug trenches, laid far more glass than anyone needed, and borrowed heavily to do it.

Then the bubble burst. More than sixty telecom companies failed in just a few years. By one estimate, around ninety-five percent of that fibre was sitting dark, with no light going through it at all.

But the fibre didn't go anywhere. It was bought up cheaply, lit up, and it became the backbone for streaming, smartphones and the cloud. The companies lost. The infrastructure won.

7:29 · What's left if it pops

Now look at our exploded view again.

The chips are like the dot-com companies: expensive, fast-moving, and quick to age. The turbines, the transformers, the substations and the concrete are like the fibre. They're slow to build, and they're built to run for decades.

So if this bubble does pop, it may leave behind something strange: an enormous amount of electrical infrastructure, in places that never had it before.

8:00 · The honest part

Now, the honest part.

Nobody knows how this ends. Bank of America's chip analysts argue the data shows scarcity, not oversupply. The orders keep coming, and they don't expect a collapse. And the International Energy Agency says the energy needed for each AI task is falling at a rate it calls unprecedented in energy history.

Even so, the IEA expects data centres to use about nine hundred and fifty terawatt-hours of electricity by twenty thirty, roughly double last year, and about three percent of all the electricity on Earth.

And none of this is financial advice. We draw machines, not stock charts.

8:49 · Railroads, and your answer

But one comparison stuck with us. In the eighteen eighties, America spent about two point four percent of its economy building railroads. By twenty twenty-eight, AI spending is expected to pass that. Many railroad companies went bust. The railroads are still there.

So here's our question for you. If the AI bubble pops tomorrow, which part of this drawing do you think will still be running in twenty fifty?