The Staggering Costs of a Single Data Center, From Land to Chips

Bloomberg Audio Studios.

Podcasts. Radio. News.

You're listening to Bloomberg Business Week with Carol Masser and Tim Stenevek on Bloomberg Radio.

All right.

We want to get to one of the most read.

Stories on the Bloomberg terminal today. It is about the staggering costs of a single AI data center from steel, cement, to chips and copper, to.

Land engineering power. And more.

There's a lot that goes into the cost of those data centers. And increasingly, all of those things cost a lot more than they used to.

Okay, but given how expensive everything is, when you add it all up, it's actually not that difficult to see how Meta, Amazon, Microsoft, and Alphabet are spending a whopping $ 750 billion on CapEx alone.

No.

At least when you scroll through this story.

Well.

Right?

Yeah.

You get it.

And I keep thinking about Mandeep's story and research about the spend next year, about $ 1.

7 trillion.

And I see this number, and you realize how much is really because of these four companies in a big way.

Well, Lynn Duan is the byline with an assist from Ian King. Lynn Duan is Bloomberg News' managing editor of technology. She used to be out here in New York. Now she's out there in California. Ian King is Bloomberg News' U.S. semiconductor reporter. They both join us from San Francisco. Lynn, I just love how you open this story, contextualizing just how big a number $ 750 billion is. Can you just lay it out for us?

Well, first of all, thanks for having me. I miss you all out there in New York. I'm sorry I have to dial in by Zoom these days. I felt like that $ 750 billion needed a lot of context. It's just such a massive number, right? As you say, we're talking about heading... into the trillion dollar realm next year.

And so I thought, how can I wrap people's heads around this? I know GDP of a country. Yes, GDP of all of Ireland, of all of Belgium. It's more than half of what all of the companies in the S & P 500 index spent in total last year. It's enough to cut $ 5, 500 in a check for every household in the U.S. So it's a lot. I mean, I think that the only thing that is comparable to this is the build-out of the U.S. highway system. Otherwise, we have never seen such a build-out of this enormity dollar-wise.

Ian, let's bring you into this because, as we know, these data centers, they're complicated. There's a lot of moving parts, and a lot of those parts, whether it's power, whether it's semiconductors, increasingly have gone up. And I think a lot of it is just basic supply and demand. There's just so much demand that the prices of these things have gone up, and it just costs a lot to make them all happen.

Yeah, that's absolutely right. I mean, you've got to think about it. The main component in these buildings is something that's like a couple of square inches.

And you're paying.

$ 30, 000.

For one of those.

It's using as much electricity as several large flat screen televisions. And you're putting 72 of those in a server. So imagine the expense of just putting one server there, of feeding it with power for a year. And obviously, there's more than just one server in these machines, in these giant data centers, as Lynn's story really, really shows very well.

Ian, I'm wondering, though, if this stuff gets cheaper. And remember, we have the Oracle Force Majeure that we saw just a couple weeks ago. But look, as we're reminded all the time, the cure for high prices is high prices. If you have a lot of people going after something, then more people come off the sidelines and try to give the market those things that are expensive because they want in on the action. Can this stuff get cheaper or is this different?

Well, the other $ 30 billion question here is chip plants, right? Cost even more than these things. So you don't just rock up and say, oh, OK, I'll build a chip factory. You can't do that either. It's, you know, two years to build a chip plant and we need multiple chip plants to meet the amount of demand.

Demand that we've got right now.

So the problem isn't just we can't build enough data centers. It's we can't build enough giant chip plants to feed these data centers with enough components. So it's kind of a compounding issue. And there's no real end in sight if demand stays at the current levels.

We're just not going to be able to meet it.

So let's go to, we want to bring up a pie chart as part of this story to just give everybody an idea. And I'll spell it out for those who are listening on radio as well. But it is a pie chart. And a big chunk of that pie, which is in yellow for those who are watching right now, $ 17. 1 billion will go to chips and servers. That's the estimate that we've come up with here at Bloomberg. Then you've got about $ 2. 8 billion in contingency planning. $ 2. 2 billion goes to other. $ 2. 4 billion is construction. About $ 3 billion goes to power. $ 1. 7 billion goes to cooling. Ian, it is disproportionately, as you just laid out, semiconductors, right? That's what makes it all sing, these data centers, and creates

the AI that people are consuming at this point. We go back to you, Tim and I, time and time again, that you remind us that.

There are cycles. And again, you do have demand, as Tim said.

Then you have build. Then you have excess capacity. And then you have some fallout. Is this cycle likely to be like any other? Or no, just maybe the duration of the cycle.

Could be different?

We've never seen anything of the magnitude of this current cycle. You've got one company that's heading to $ 100 billion per quarter in sales at a 75% gross margin. You've got memory chip companies that have demand for the next several years already locked up. So There's never been anything like this in terms of the magnitude, in terms of how long it'll go. If I knew that, I wouldn't be sat here talking to you.

But you do remind us that it's cyclical, right? There's nothing to make you think.

That this is going to be any different this time around.

I mean, the industry has always, always overbuilt, right? We've always had this lag between supply and demand. We've never been able to match it. Anybody who tells you that this is going to be different is probably not really understanding how the industry works. It's just a case of when the pendulum swings back.

Lynn, I want to bring you back in here. We're talking to Lynn Dewan. She's Bloomberg News Managing Editor of Technology at Bloomberg. And also Ian King, Bloomberg News U.S. Semiconductor Reporter. Both join us from San Francisco. The story we're talking about is the staggering cost of a single data center from land to chips. I encourage everybody to read this on the terminal or read it at Bloomberg.com and just play with the interactive element a little bit because it just shows how these costs add up. And there are a lot of costs in there that you wouldn't necessarily associate with this type of CapEx. And that's where I want to send it to you, Lynn. What was the most surprising element of cost that goes into

a data center from your research that you found?

Well, I just want to double down on what Ian said about the cyclical nature of chips. In that way, I used to oversee the energy and commodities team at Bloomberg. In that way, chips have very much become like any other commodity. Every commodity market and energy market is cyclical. You have these cycles of oversupply and then shortage of supply.

And what really stood out to me in doing this deep dive with Cushman and Wakefield and Bloomberg Intelligence is that there is this bottleneck across all raw commodities, not just chips, but natural gas turbines and power transformers. and concrete, and cement, and copper, and steel. The inflationary nature of this boom is across the board. I think that was probably the most staggering revelation to me is it's not just about the chips. It's about the costs of all of these things soaring, and then suddenly you understand why some of these data centers are running into huge problems and potential delays.

This is why, though, Lynn, when we talk about delays, those that purchase land or entities that is already powered, the powered land, like it's something that we've kind of started to learn about, too, in conversations with real estate folks. That is in demand because you've already got access.

Right. And we're not even talking about powered demand. It's a powered land itself soaring in costs. You know, we have firms that are building themselves as sort of the arbitrage, the supplier of powered land, like Tract Capital does. A company that we have profiled has secured more than 30,000 acres of this kind of powered land just to sell it to data center providers. There is one basic reason why powered land is taking off in this way, and that is because there is not enough electricity generation in this world to power all of the AI data centers that are being proposed and built right now. And powered land means that they are already either connected to the grid or already have a power source contracted out.

Hey, one of the things I wanted to ask you, because Ian, Lynn brought up power.

Obviously, that's a big part of all of this.

And that, I feel like over the last couple of years, we've increasingly, when we talk AI, we talk about the.

Power demand and needed.

Data centers, in this story, data centers spend a good 20% to 40% of their energy simply cooling the equipment housed within. And those built to power the development and training of AI systems require more cooling than most.

Chips, though, Ian, are getting better, right?

Smarter, faster, but I assume they're getting more productive, that hopefully that power demand will be reduced, but that's part of what we're trying to see in terms of the development of more advanced chips.

That's part of it.

What we're seeing is a computing problem that is, in effect, inflationary. It's inflationary in cost for all the reasons that we talked about, but it's also inflationary in terms of the power produced.

Yes, they are more efficient. Yes, they do more work.

But guess what? The more work they do, the more useful they become. And AI is a problem that fundamentally, the more data you throw at it, the better the software becomes. So there's an incentive to use more chips, more powerful chips, chips that can move around more data. So you've got like your laptop, right? 20 watts, that's quite a bit of power for something quite small. But we're talking individual chips here that are using more than 1,000 watts of electricity.

Why are we doing that? That sounds crazy.

We're doing it because it makes sense if you want to create these effective AI systems. So for the time being, unless there's a massive breakthrough in the fundamental mathematics, we're going to continue to throw more power around. more powerful chips that require more energy.

At this issue.

Ian, you covered, I ask you this all the time, but I mean, each time we interview you, it feels like things are even more intense or more extreme than they were the last time. So given your decades-long history of covering this, where you think we are in this cycle, does it match anything that you've ever seen in this industry?

Absolutely not.

No, I mean, people have tried to compare it to the dot-com era. People have tried to compare it to when PCs took over from mainframes. We've never seen anything to this extent, and you just need to look at the actual numbers that the companies are putting up there. Even valuations can't keep up with them. Even some of, I mean, Hoctan this weekend, the CEO of Broadcom, called casino-type investing going on. But even with that, some of these companies are relatively cheap still on a price-to-innings ratio.

Lynn just got about 25 seconds here. Anybody that, as you were putting this together, saying, wait, boom, bust, bubble.

Any of that come up?

I mean, there's a reason why the very end of the story is framed around whether all of this investment is going to be worth it. I think that is a huge question mark, and much like Ian, if I knew the answer to that, I wouldn't be on television with you. I will say, I had a Bloomberg reader just ask a question, which is, if we are consuming all these resources with data centers, why don't we just allow human beings to consume these resources instead? That speaks to the boom and bust. We'll see what happens.

Yeah, Michael Burr has done a lot about this, too, about the eventual burst of the AI bubble.

Folks, thank you so much.

Our thanks to Linda Wong, Bloomberg News Managing Editor of Tech, Ian King, Bloomberg News U.S.

Semiconductor Reporter.

The Staggering Costs of a Single Data Center, From Land to Chips

Kapitelklicken springt direkt hin

Shownotes

Transkriptvom Podcast bereitgestellt · Zeitmarken springen hin