Is the AI Boom… a Bubble?

My colleague Berber Jinn covers AI from San Francisco.

But last month, he found himself a world away from Silicon Valley, wearing a hard hat in the middle of the Texas brushland.

I mean, having covered tech for a few years, like I would never have imagined traveling to a massive construction site as part of my job, but that is the moment we are in.

The moment we're in is an AI building boom.

And Berber and a bunch of other reporters had traveled to its beating heart.

A massive new data center being built for Open AI outside Abilene, Texas.

When we got there, it was basically just like a massive construction site.

A lot of workers riding around in buggies, a lot of like pits in the ground that they were digging to lay fiber cables to connect the different data centers.

There were like massive natural gas turbines that were serving as backup power for the facility.

And then there were these eight white data centers that were springing up from the ground.

The supercomputers in this complex will field users' Chat GPT requests and train the next generation of OpenAI's models.

When all eight buildings are complete, it'll be the size of New York Central Park.

This kind of mega construction project is happening across the country.

Tech companies are pouring hundreds of billions of dollars into building a new generation of AI data centers.

It's one of the costliest building sprees in history.

And for market watchers, AI skeptics, analysts, and investors, this fire hose of spending is raising one very basic question.

Will all this investment pay off?

Or are we watching an AI bubble inflate?

Welcome to The Journal, our show about money, business, and power.

I'm Jessica Mendoza.

It's Tuesday, October 14th.

Coming up on the show, the daunting math of the AI building boom.

This episode is brought to you by Credit Karma.

A good podcast helps you connect the dots.

It pulls in interviews, backstories, and multiple POVs to help you see the full picture.

It's what the journal aims for.

Credit Karma does the same thing, but for your money.

Just link your accounts and the app can show you ways to save, pay down debt, and build smarter spending habits.

Get the full picture of your money in one place.

Download Intuit Credit Karma to get started.

This episode is brought to you by Ericsson.

America relies on secure mobile networks.

Across all 50 50 states, America's mobile networks rely on Ericsson.

For over 120 years, Ericsson has been connecting Americans, and now the majority of U.S.

mobile network traffic runs through Ericsson equipment.

These networks are built in Texas and are powering American innovation and growth.

Ericsson, what America's mobile networks are made of.

Learn more at Ericsson.com/slash USA.

AI companies measure the capacity of their data centers in watts, as in, how many megawatts or gigawatts of energy will a data center need to power the supercomputers inside?

When it's finished, the Abilene complex will need about 1.5 gigawatts.

That's approaching the amount of energy generated by the Hoover Dam.

But OpenAI CEO Sam Altman says, even that much computing power won't be enough.

When we come walk sites like this and look at the incredible work, the first thing we think is, this is awesome.

That's Altman in an interview at the Abilene site.

And the second thing we think is we have to figure out how to do much more, faster, better, cheaper, to continue to meet what the demand scales will be years from now.

Altman envisions a future not that long from now when AI will be built into nearly every aspect of our lives.

Here's Berber again.

The idea is that People are going to go from using ChatGPT like any other app to like having an AI agent that will help you book your travel recommendations and help you complete your tours.

It can help you plan a party, send invitations, order supplies.

It can help you understand your healthcare and make decisions on your journey.

And this AI assistant's gonna be running tasks for you while you're asleep.

It can write an entire computer program from scratch to help you with whatever you'd like.

And we think this idea.

Scientists are gonna be using AI to like find a cure for cancer and major governments are gonna be using AI with like precision warfare and corporations will be like using AI assistance instead of like all these low-level employees that they're hiring.

And all of this will need power, like computing power.

Yes, exactly.

To meet what it expects to be a surge in AI demand, OpenAI will need to build more data centers like Abilene, and it's laying plans to do just that.

This summer, OpenAI cut a $300 billion deal with the cloud computing company Oracle to buy more computing capacity.

And on the tour Berber attended, the two companies made yet another announcement.

Today we're here to talk about some very big investments that we're making in Craig.

So OpenAI announced new data centers across the country.

And an expansion to the Abilene data center, which they said would basically bring OpenAI to seven gigawatts of capacity for their data centers.

And did they say anything about how much all of this was going to cost?

OpenAI and Oracle did not announce how much the seven gigawatts of capacity would cost, but OpenAI executives estimate that each gigawatt of capacity costs roughly $50 billion to create the data center and buy the chips to put in them.

So seven gigawatts by that math would be $350 billion.

I was also told on the trip by OpenAI executives that OpenAI would need closer to 100 gigawatts.

100 gigawatts?

Yes, which would put us easily into the trillions of dollars range.

If one gigawatt costs $50 billion,

100 gigawatts would cost $5 trillion.

That's more than the annual GDP of Germany.

And the thing is, it's not just OpenAI that's thinking and spending this way.

Amazon is on track to pour more than $100 billion into CapEx this year.

Microsoft says it plans to spend about $80 billion in its fiscal year to build out artificial intelligence.

Meta says it will invest hundreds of billions of dollars into compute build out, and that is unprecedented.

Multiple multi-gigawatt clusters in the works of the first company.

In the next year alone, Microsoft, Meta, Amazon, and Alphabet are expected to spend nearly $400 billion on AI.

This type of spending and speculative betting is what is defining the tech sector right now.

Our colleague Elliot Brown covers finance, and he's been tracking the AI building boom along with Berber.

Why are they all kind of piling in?

Is it because, you know, OpenAI is spending all this money and Meta looks at them and they're like, oh, we can't get left behind.

We need to spend this money too?

A leading theory for sort of what explains all of this is each one of these big companies has to do this type of large investment because the other one is.

And you can even even see quotes from Zuckerberg and the CEO of Google who have

nodded to this.

You know, the risk of underinvesting is dramatically greater than the risk of overinvesting for us here.

Where they say, well, there could be some overbuilding, but the last thing I want to do is be wrong and left behind.

So the rational thing to do is to spend money now.

But I think, you know, not investing to be at the frontier, I think, you know, definitely has a much more significant downsides.

downsides.

Having said that.

So for the whole AI sector, how much money are we talking about?

The cost, if you look at what has sort of gone into AI infrastructure in the past three years and look a little further out, it's well over the cost of the interstate highway system over four decades.

What do we know about the math?

Like, how much would AI companies have to make for all of this investment to pay off?

Yeah, I mean, I think that's probably the the most concerning part of this and also the part with the least clear answer.

So Bain, the consultants, Bain ⁇ Co., put out a report that estimated you'd need $2 trillion in annual revenue by 2030.

Help me make sense of that ginormous number.

Like, what does that mean?

$2 trillion by 2030.

I just, you know, I've never seen that kind of money.

It is an absolutely enormous amount of money and it's more than the combined revenue of Amazon, apple alphabet meta microsoft nvidia that's basically the entire mega tech sector's revenue the ai industry would have to make that much by 2030 you said yeah which which is uh you know not very far away

how much does ai make now yeah it's not terribly easy to count but it's it's a really small number by comparison uh so morgan stanley put it at in in 2024 the number was 45 billion yi

so

that's 1.9 trillion something short.

What was your first thought when you heard those estimates?

Is there a bubble?

That's next.

Optimism isn't sunshine and rainbows.

It's fixing things.

Changing things.

Changing the way we fix things.

Rolled up sleeves, breaking through,

growing power to meet growing needs.

working to run the world on smarter energy every day, taking power where the grid's never been, then getting up and doing it again.

Because if optimism never stops, then change can't either.

G.E.

Vernova, the energy of change.

This episode is brought to you by FirstNet.

During September 11th, first responders struggled to communicate and respond because the networks and systems they relied on were overloaded or destroyed.

Today, they rely on FirstNet.

FirstNet was built with and for first responders, ensuring they have reliable communication wherever and whenever they need it.

FirstNet covers more first responders than any other network.

That's why it matters for every American.

FirstNet, built with ATT.

Learn more at firstnet.com slash public safety first.

When you think about bubbles in the tech sector, Elliott says there's one pretty obvious place your mind goes.

I think there are a lot of really eerie parallels with the dot-com boom.

In the early 2000s, investors were bullish on the internet, pouring money into seemingly any company with dot-com at the end of its name.

Pets.com, with its loudmouth sock puppet mascot, is a memorable example.

Now that you can get whatever you want at pets.com, it's like Mardi Gras!

Squeaky toys, new collars, leashes.

When the bubble burst, many of those dot-com companies went bust, and investors lost their shirts.

But Elliot says, some of the biggest losers in the dot-com bust weren't actually dot-com companies.

They were telecoms.

What happened in the 90s was everyone saw that the internet was here and it was pretty overwhelming in expectations for its growth.

And telecom companies and fiber builders just started blanketing the country with fiber optic cable.

Telecom companies expected a surge in demand for high-speed internet internet connections, and they poured over $100 billion into building out that infrastructure.

That went on for years, and then when the dot-com boom turned to bust, you suddenly realized that the country was massively overbuilt in terms of the amount of fiber that was needed.

These fiber builders essentially thought that the trees could grow to the sky because the internet was so real.

And in reality, the internet was very real, but it just didn't produce the absolutely stratospheric growth that all of these fiber builders were counting on.

And yeah, I see a lot of that today, where the reality of AI and inevitability that it has some large role in our lives coming forward is sort of justify anything.

By the end of the dot-com collapse, major telecom companies had declared bankruptcy and $2 trillion in market value had evaporated.

Proponents of AI argue that this building boom is different.

The internet took a while to take off, in part because people had to get their homes wired for high-speed internet.

Today, in many parts of the world, you can start using AI instantly.

And while AI isn't profitable yet, revenue is growing.

Take OpenAI, for example.

To OpenAI's credit, they didn't release a product until November 2022, so they're at $13 billion of revenue in three years, which is really fast, extraordinarily fast.

Here's Berber Jinn again.

And so if you like keep the growth rate year in, year out and just extrapolate it over the next 10 years,

it's going to get to a really high number.

But for those concerned about a bubble, there are worrying signs that despite all of the spending, AI might not be taking off in the way that companies need it to.

Take the launch of GPT-5, OpenAI's latest language model.

The hope was that it would be a giant leap forward for AI.

But for many users, GPT-5 landed with with a thud.

On social media, Altman called the GPT-5 rollout bumpy, attributing it to the number of projects OpenAI was juggling.

GPT-5 was not this momentous change in AI.

It was more of this incremental improvement.

At least that's how it was widely viewed.

That, if that continues, is a huge problem because Every time you make a new model, you spend multiple more times the investment to create that model.

Every single new model is even more expensive to build out than the last one, basically.

Yeah, and

many billions and then ultimately going to the tens of billions of dollars.

So therefore, every time you do a new model, you need to get two, three, four times more money out of it.

Making all that money could be tough because for all the people who use AI, only about 3% pay for it, at least according to one study.

And many businesses that do pay for it apparently haven't gotten a lot of use out of it.

One MIT report found that 95% of the organizations surveyed were getting no return on their AI investments.

And remember, this is a sector that needs to make about $2 trillion a year by 2030.

So, okay, if AI doesn't take off explosively, who gets burned?

Yeah, so it's important to look at it that way because the worry is a lot less that AI isn't going to transform the economy.

I think there's a lot of thought that AI will transform the economy.

It's just about when.

And so if it doesn't happen quickly in the next few years, you're gonna have a lot of debt that is gonna be in trouble.

And then you're gonna have a lot of companies who are sort of all in on this, who are gonna be in a lot of trouble.

One company that could find itself in a tight spot is OpenAI.

Under the terms of its deal with Oracle, OpenAI will soon need to pay the company about $60 billion a year.

Today, OpenAI makes about $13 billion in revenue, according to someone familiar with the company's finances.

Elliott says other less well-known players in the AI building boom could also feel the squeeze, like companies that build data centers.

And so this is usually a real estate company, and they actually, you know, do what it sounds like.

They get some steel and concrete and a bunch of backup generators and build a very expensive, very big warehouse.

Building gigantic warehouses is expensive.

So these companies often take on debt to do it.

And then there are the middlemen.

Sometimes you have a third party, a middleman,

which basically leases the data center, buys the chips, and then rents the servers that they build inside to the big tech companies.

That's so interesting.

It's of course there's a middleman now.

Of course there's a middleman.

Those middlemen often take on debt too.

For example, to buy all those chips.

If demand for AI doesn't take off explosively, all this debt doesn't go away.

Berber, have tech leaders addressed the bubble talk?

Yes.

A lot of tech CEOs have sort of acknowledged the moment we're in, right?

Where there's a lot of, there are a lot of assumptions that are being made and a lot of gambles that are being made with their money.

But I think like in any bubble, like people like to say, we're not the dumb money.

The dumb money is somewhere else, right?

Like people are going to lose money, but it's not going to be us.

Right.

Like you talk to every layer of the stack, right?

Like the open AIs of the world, the cloud providers, the data center investors, the data center builders, the land and power people, like all of them say like, it's going to be a bubble, but we're not going to lose the money for these XYZ reasons.

That's the kind of psychology also behind how these investors and CEOs rationalize the amount of risk they're taking.

Sam Altman specifically has acknowledged the possibility of a bubble.

He said that investors are overexcited about AI and that some would, quote, get burned.

His argument is that over a long stretch of time, right, it's going to be worth it.

And in the case of fiber, it was.

The long-awaited surge in high-speed internet demand did eventually arrive.

And when it did, all that cable infrastructure the telecoms had built finally got put to use.

We're still using it today.

The infrastructure survived, even if the companies that built it didn't.

But in the case of AI, it's not clear all this infrastructure will prove as useful over the long haul.

So you put a fiber cable in the ground and you sit there for 15 years and don't use it, and then you turn on a switch and you can use it.

The problem with the chips is that they get better constantly.

And so if we don't need all this capacity, but we need it in 10 years, it's kind of like having a bunch of iPhone 4s lying around in a warehouse.

If the AI revolution doesn't arrive or arrive quickly enough, it's not clear what happens to all these new expensive data centers like the one in Abilene.

I think there's a risk that there are just a lot of these shells of buildings that like aren't operative.

I mean, this is kind of right.

Like this is like a real bare case scenario, but I mean, it's giving like zombie mall vibes, you know.

It would be pretty hard to convert to housing.

Do we get any better at identifying bubbles?

Or do we have to learn this lesson again every time?

I think manias, bubbles, are just an incredibly ingrained part of history.

Like if you look back to basically any transformative technology that requires infrastructure investment, we've had infrastructure bubbles in all of them.

You know, the internet, we've talked about electricity, electricity took ages before it was really adopted by the economy and revved up manufacturing.

It took like 50 years.

Canals, railroads, railroads had multiple bubbles in them.

So these things are all really useful, but the initial builders that got really excited about them all got burned.

That's all for today, Tuesday, October 14th.

The journal is a co-production of Spotify and the Wall Street Journal.

Additional reporting in this episode by Robbie Whelan.

Thanks for listening.

See you tomorrow.

At the University of Maryland, brilliant minds unite across disciplines to ignite AI innovation.

We apply visionary technology to track wildfires, spot hidden illnesses, protect our food supply, and make AI's benefits accessible to all.

While preparing tomorrow's leaders to explore AI's potential and put people first, we're shaping the future of AI for discovery, for community, for the public good.