Colossus 1

Tech billionaires have made an enormous bet on AI, the biggest bet that tech has made on anything in a very long time. Reporter Sruthi Pinnamaneni goes on a journey – from Data Center Alley to the shadow of Colossus – to see what the money’s being spent on, and to learn what happens if things go wrong. Support Search Engine! https://searchengine.supercast.com/

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:25PJ Vogt?

  This is Search Engine. I'm PJ Vogt.

• 0:29PJ Vogt?

  Lately, the canniest 15-year-olds I know are putting all of their allowance towards buying fractions of Nvidia stock. It only goes up, they tell me, with the wisdom of people who have finally figured everything out in life. My teenage investment advisors inform me that it keeps going up because the AI companies have promised that one day soon we'll reach a kind of machine superintelligence that, among other things, justifies these vast oceans of investment. It's funny, I do actually believe AI is a very useful, certainly socially transformative technology for bad and for good. I believe that, and I also believe that we're probably in a bubble. Because a bubble would just mean that, once again, investors have gotten a little too

• 1:14PJ Vogt?

  excited about a genuinely promising new technology. The sheer fortunes being spent on AI already are so large, it's hard to find the right precedence to compare to. And what we know is that the size of the bet means we're all, to some degree, in this. Whether this pays off, or crashes, or lands in between, the AI bet is one of the big stories of our time. But as big as it is, it can also feel abstract. And so a reporter here wanted to go see it. Visit the most prominent real-world manifestations of all this money. The data centers, the buildings powering AI, which have already begun to transform American towns and cities.

• 1:55PJ Vogt?

  She learned things that surprised me about what's happening right now, about the untested assumptions that have led us here, and about how it might look if things go wrong. The story begins in the part of America that has built the most data centers, even before the starting pistol of the AI race had been fired.

• 2:15PJ Vogt?

  Our reporter is Shruti Pinamaneni. I'm gonna let her take it from here.

• 2:20Speaker 01

  Chapter one, The Gold in the Ground.

• 2:27Speaker 01

  Ashburn sits on the very northern tip of Virginia, right by the Potomac. Driving around on the main roads here, everything looks so flat. It's more lawn than trees, lots of short fences behind which are these very square, very squat, windowless buildings.

• 2:45Speaker 02

  So you can see that, you know, we're on Loudoun County Parkway now. And more data centers here on the left.

• 2:53Speaker 01

  My guide today is Buddy Reiser, shaved head, dark suit. And this county that he's driving me through in northern Virginia, Loudoun County, it has the highest concentration of data centers in the world. Second place is a sleepy little town named Beijing.

• 3:09Speaker 02

  Everything you're seeing is data center. Data center here, data center there, data center there. This is data center alley proper.

• 3:17Speaker 01

  Honestly, I've read a lot about this before getting here, but the scale of just the construction is mind bending. Like, everything is being built right now, and that's on top of a whole lot that already existed.

• 3:32Speaker 02

  That's correct. I mean, often when I take people on tours, they think, wow, I've heard a lot about it, but this exceeds my expectation. And what you have to remember is that this is one of five clusters here, right? There's the Sterling Cluster, there's the Leesburg Cluster, there's the...

• 3:51Speaker 01

  Everywhere you look, there's either data centers or data centers under construction. Cranes abound. This building frenzy, it has certainly gotten an adrenaline shot from AI. Some of the new unfinished buildings we were seeing were AI inference centers. But Buddy told me that construction here has been going nonstop for many years. That really, it began back when we were all first starting to store our data in the cloud. Back then, this was the late 2000s, Google had announced Google Docs. You didn't have to save documents on your hard drive anymore. Amazon had announced Amazon Web Services, meaning startups could borrow Amazon servers. They didn't have to own their own anymore.

• 4:33Speaker 01

  And the invention of the cloud would be a huge deal. It would change almost everything about the way we use the internet. There was money to be made here, and Loudon needed that money. The county was reeling from two successive crises, the dot-com crash and then the mortgage crisis. And in the wake of those disasters, Buddy showed up.

• 4:58Speaker 02

  I arrived here in 2007. And you have to remember, we were pretty desperate then, right? We were laying people off. We were raising taxes.

• 5:09Speaker 01

  And was it that you were losing property taxes because people were leaving because of what? Like their companies had collapsed?

• 5:18Speaker 02

  No, it was, you know, that all, that whole fake mortgage bubble. 81% of our tax revenue was coming from homes, residences. So when the housing bubble burst, you know, all of a sudden, we lost about a third of our tax base. So that was how I came here. I was our first real proactive salesperson.

• 5:40Speaker 01

  There was a reason I'd come here to meet Buddy and have him tell me this story.

• 5:46Speaker 01

  The thing that he was doing back then, which is selling companies on Loudon, inviting big business to town. This happens everywhere. Every town in America needs money, tax revenue to fund schools, roads, fire stations. And so they invite companies to come set up. They try to negotiate good terms. How good a job they do at that? What kinds of corporations arrive? That can really define how life feels in a place afterwards. The story of how things played out in Loudon would end up contrasting with the story of another city I'd visit later on, where these same buildings, data centers, would come to represent something very different. But for now, Loudon.

• 6:30Speaker 01

  Buddy and his quest quickly learned a couple helpful facts about Loudon. AOL used to be headquartered here, and during the telecom bubble of the late 90s, a lot of fiber optic cable had been laid in the ground. These are the long glass tubes that carry internet traffic using infrared light. And after the bubble burst, much of Loudon's fiber just sat there, unused. And so what was here, underground, was the data equivalent of miles and miles of pristine, mostly empty highway.

• 6:58Speaker 02

  This area right here, this intersection of Loudon County Parkway and Waxpool, is basically main and main for the internet. There's more fiber in the ground here than anywhere in the world.

• 7:11Speaker 01

  Wow, right here?

• 7:12Speaker 02

  Right here, this intersection. When you think about fiber, it's about capacity, and it's about latency, and it's about being close to that that center hub. So this land around here is some of the most expensive land you'll find.

• 7:28Speaker 01

  And it's because the proximity means speed.

• 7:32Speaker 02

  Proximity means speed.

• 7:33Speaker 01

  I do want to point out, there's a bunch of ducks on this intersection data center alley.

• 7:39Speaker 02

  Yeah, it's this here.

• 7:41Speaker 01

  Are those ducks or geese?

• 7:42Speaker 02

  Geese. Here on the left, here's another data center that's being.

• 7:47Speaker 01

  So Loudon was trying to recover from the popping of past bubbles, but those bubbles had left behind infrastructure that was actually very valuable, ready to be taken advantage of. The high-speed fiber in the ground, but also these big, empty buildings that had been built before the dot-com bubble burst. And so Buddy presented his idea to the board. Let's find some rich tech companies to fill the buildings, and that is where we will get our tax base.

• 8:14Speaker 02

  I remember sweating through my suit, thinking, you know, how am I trying to sell this? And one of the the board members looked at me and said, Mr. Reiser, I don't understand why we need data centers if everything's going to the cloud. And I'm just like dumbfounded because, well, obviously, the cloud is in the data centers. It's not really in the clouds, right?

• 8:35Speaker 01

  Right.

• 8:40Speaker 01

  It's funny, to Buddy, this is just an amusing anecdote, but I found myself thinking about it a lot here in Virginia. What we call a new technology really shapes how we understand it. The name, the cloud, picked probably to make us feel safe, to make us feel like we're not putting our data and trust in other people's computers. The name also has this side effect of making you forget that the internet isn't a thing that's invisible in the sky. It's very physical. It's made of cables and servers that all have to be put somewhere. Buddy explained this to the board member, and then he hit the road, pitching big tech companies on Loudon.

• 9:19Speaker 02

  We've never given $1 of local incentive money to data centers to bring them here, not $1. What we did, part of our pitch, was we understand data centers. We know how to get them to market. We're going to make it fast. We're going to make it predictable. And that was the pitch. That's what I was out on the road selling to, you know, Amazon and Microsoft and Google and all the rest of them.

• 9:47Speaker 01

  The pitch worked. Every tech company under the sun would arrive, the big ones, the new ones, many you've never heard of. And as the cloud grew, invisible to most of us, people in Loudon could see all these concrete boxes blooming across the landscape.

• 10:04Speaker 02

  When the cloud started to take off, you know, I really say about 2017 is when that really took off. Then we had a real big growth spurt there. Then, you know, when COVID hit and everybody started working from home and we were, you know, stuck watching TV and we were streaming videos and we were communicating and learning and working and everything through the internet, all of a sudden, there was another growth spurt.

• 10:31Speaker 01

  I had no idea. You were seeing the internet from a completely different angle than the rest of us.

• 10:37Speaker 02

  Yeah, and we were seeing that demand. And there's not been a single day without data center construction in Loudon for going on two decades now.

• 10:45Speaker 01

  That's pretty crazy.

• 10:46Speaker 02

  You know, fast forward from 2007 to 2025, we've been able to lower our tax rate by 48 cents on the dollar. We had 10 straight years where we were building elementary schools, middle schools, and one high school. 10 straight years we were doing that. And that money comes from the data centers. You know, we'll get a billion dollars in local tax revenue from the data centers this year.

• 11:12Speaker 01

  I do need to fact check Buddy here. The county made almost a billion dollars from data centers this year. 895 million, to be exact. That enormous amount of money is perhaps why, even though you do find people in Loudon complaining about the data centers, nobody wants the new transmission line going through their backyard, the general sentiment seems tolerant.

• 11:37Speaker 01

  Chapter two, Inside the Cloud.

• 11:41Speaker 03

  No, Vincent, you can describe what you see, but you will never, ever say the name of what you see.

• 11:49Speaker 01

  Michael Whitwalk, general manager of the Sabi Data Center in Ashburn, he's actually much friendlier than he sounds over here.

• 11:56Speaker 03

  That area still is off limits, no matter what you want to say.

• 11:58Speaker 01

  That's not an issue.

• 11:59Speaker 03

  No, not what you say.

• 12:01Speaker 01

  The main thing that Michael is trying to keep secret is the names of the companies that are renting server space here. I asked why those names were so secret. It was indicated to me via hand gesture that that is also a secret. Oh, it's secret. Okay, that's fine. But Michael was otherwise happy to show us how this cloud data center functions.

• 12:18Speaker 03

  And what I'm going to show you today.

• 12:20Speaker 01

  The giant block of a building is organized in layers, kind of like an onion.

• 12:25Speaker 03

  So we're here on our loading dock, adjacent to our generator yard.

• 12:29Speaker 01

  The outer layer outside the building is all the power infrastructure. So there's this power substation pulling electricity from the grid.

• 12:37Speaker 03

  If you look over top of those colas, you'll see a giant substation. This is the Waxpool Substation.

• 12:41Speaker 01

  Virginia's data centers use. A ton of electricity. Nearly 40% of the state's estimated power use this year will be just for data centers. However, most of the air pollution being created by that power generation, that pollution is not actually happening in this county. It's happening elsewhere, in other parts of Virginia or other states like Ohio, Pennsylvania, because that's where the grid is pulling power from.

• 13:07Speaker 04

  Sunny outside.

• 13:08Speaker 03

  Very sunny.

• 13:09Speaker 01

  So here, today, the skies are blue, the air is clear, and the people of Loudoun County are not dealing with the side effects of their power needs.

• 13:21Speaker 01

  Inside the building, Michael shows us the next layer of the onion, the cooling layer.

• 13:26Speaker 03

  So we have now gone into what we call the Cray Gallery. It's computer room air handler. These are the--

• 13:34Speaker 01

  A long room filled with industrial air conditioning fans, each one about my height, with a maze of metal tubes snaking cold water in and hot water out. And are there usually people working in here, or is usually this empty?

• 13:46Speaker 03

  It's empty until we have maintenance going on in here.

• 13:53Speaker 01

  Have you seen the TV show Severance?

• 13:56Speaker 03

  I have not.

• 13:58Speaker 01

  We look like we're walking through the sets of Severance. It's just like these endless halls with fluorescent lighting, and every door looks like the other door. And I imagine it's easy to get lost.

• 14:10Speaker 03

  This is our longest hallway in the facility? I'd say this is a quarter mile.

• 14:16Speaker 03

  This is a nice long--

• 14:18Speaker 01

  I thought about what it would be like for some future archaeologists finding the remains of this building, long after we're gone, after the internet's been shut off. What would they think this is for? I realized the place reminds me a little of photos I've seen of the insides of pyramids. A labyrinthine, enormous tomb. Do a lot more people want to see data centers these days?

• 14:40Speaker 03

  It's always been a hotspot, but--

• 14:44Speaker 03

  just makes it that much more like, hey, what's going on?

• 14:47Speaker 04

  What is going on? This is very loud.

• 14:49Speaker 03

  You're in what we call our MPOE. It's our main point of entry. This is where--

• 14:54Speaker 01

  This is the next layer, where all the fiber enters the facility.

• 14:59Speaker 01

  And finally, we get to the most important layer of all. So we're entering a data hall. The heart of the building, the place where the computers actually live.

• 15:08Speaker 03

  I'm gonna give you an up and close personal on a place that everybody doesn't get to see, so you're special. I'm gonna give you your 30 seconds.

• 15:19Speaker 01

  I get 30 seconds to stand here amidst the tall white stacks. Walking behind one of those stacks, I see an actual server rack. A metal-framed rectangle, the kind you'd glimpse in the IT closet of your office, holding this stack of servers. An unremarkable hunk of metal, but this is where the cloud lives.

• 15:40Speaker 01

  It's funny, I spend most of my workdays living inside of a Google Doc, but before this story, I'd never thought about the backend that enabled that. Like, every time, at the office, I type a single letter, say, an X, into a Doc, that X gets transmitted over my Wi-Fi to the router, where it jumps to my internet service provider's fiber, onto a fiber-optic highway, and ultimately, to a data center like this one, to a room like this one, to a server like the one I'm looking at.

• 16:15Speaker 01

  The data center is infrastructure. And like most infrastructure, or water lines, or sewage pipes, it's both an engineering miracle, but also so reliable as to render itself boring. It works, so we don't notice it.

• 16:32Speaker 01

  For years, that's what data centers were. Boring miracles that nobody, except, of course, Buddy Riser, ever really thought much about. And then, in late 2022, a nonprofit, whose stated mission is to ensure that artificial general intelligence benefits all of humanity, quietly released a chatbot that would upend our understanding of the internet. ChatGPT.

• 17:03Speaker 01

  And almost overnight, a data center would take on a different meaning. Instead of being boring old pieces of internet infrastructure, now, data centers are being used to train AI, which makes them territory in a war. An enormous, expensive, ego-driven war waged by tech billionaires.

• 17:26Speaker 01

  After the break. Prometheus. Stargate. Colossus.

• 17:58Speaker 01

  Chapter three. Arms Raise.

• 18:02Speaker 01

  So when you first got assigned the cloud computing beat, what did you think?

• 18:06Speaker 05

  I thought that I did not know how I was going to make cloud computing interesting to the reader or to myself. And so it was sort of a let me do this niche boring thing for a little bit before I switch beats. And now I'm like, this is going to be my beat for a long time.

• 18:25Speaker 01

  Anisa Gardizi is a reporter at The Information. She'd been assigned the cloud beat in early 2023. Lucky for her, this was the exact moment when this formerly sleepy part of the tech world was beginning to explode.

• 18:38Speaker 05

  OpenAI had just a couple of months earlier released ChatGPT. And everyone was going crazy. How did they do that? This is such a breakthrough. We have to catch up.

• 18:48Speaker 06

  Tonight, we're taking a deep dive into the world of AI with a special focus on ChatGPT.

• 18:55Speaker 07

  ChatGPT, which stands for Generative Pre-trained Transformer. And it's fully powered through artificial intelligence. Just rolls off the tongue. Uh huh. Joining us now to talk about.

• 19:05Speaker 01

  For most consumers, honestly, even for most tech reporters, ChatGPT came as a total surprise. But like most overnight successes, it had been building for years. Back in 2012, a small group of academics had trained a model that was unusually good at classifying images. It could distinguish, for instance, between different breeds of dogs. The academics trained this primitive neural network using more than a million images. But their big innovation was that they'd handle the computing load with a particular kind of chip: a graphic processing unit, a GPU. A GPU specifically from a company called Nvidia, which most people at the time associated with video games.

• 19:49Speaker 01

  The academics built their little neural network using just two of these Nvidia chips. They won a contest, and they published a paper announcing that this was possible. OpenAI spent the years between that paper and ChatGPT's release applying that insight that you could build more powerful neural networks by feeding them more training data and wiring together more and more GPUs to process it all. In 2012, the academics had used two Nvidia chips. By 2020, OpenAI had partnered with Microsoft to connect 10,000 Nvidia chips. Just a couple of years later, they were reportedly up to 20,000 chips, which is how they were able to train ChatGPT 3.5, the chatbot that would take the internet by storm.

• 20:34Speaker 05

  Once AI started taking off, it became clear that to train these massive models that companies like OpenAI needed to train, they needed an enormous amount of GPUs all in the same space.

• 20:47Speaker 01

  The data center setup that Microsoft had built for OpenAI, it was worth hundreds of millions of dollars and packed with GPUs. And that was now the exact kind of setup that everybody else wanted. And what would have happened in the earlier cloud era is that the biggest cloud providers like Google and Amazon would have just built those kinds of data centers and rented compute to everybody else. But now, in the AI era, these companies seemed more preoccupied with competing with each other.

• 21:18Speaker 05

  OpenAI developed ChatGPT on Microsoft's cloud. And so you had Amazon and Google looking around the room and figuring out how they could catch up with ChatGPT. And so the issue that they were facing was that they were buying so many Nvidia GPUs, but they had to decide, how many are we going to rent out to our customers? And how many are we going to keep for our internal AI teams? Because Microsoft just one-upped us with ChatGPT.

• 21:45Speaker 01

  I see. So it's like, they realize that's a way to almost kneecap their competitors? Like, if you're Google, you might be thinking, I have GPUs, but if I save more of those GPUs for myself, for my own developers, then in a way, that is like a competitive advantage.

• 22:04Speaker 05

  Exactly. They were very much, you know, some people would use the word hoarding, but they were definitely keeping large amounts of the Nvidia GPUs for their AI research teams.

• 22:15Speaker 01

  Again, ChatGPT was a huge surprise to us normal people. But what you can't lose sight of is what it meant to companies like Google and Meta, giants that typically dominate the internet, but who this time had been caught half asleep. Overnight, the big tech companies wanted their own powerful chatbots, and they were hungry now for GPUs to try to quickly train those chatbots up. This is where our current data center construction boom begins.

• 22:43Speaker 04

  Amazon plans to spend $10 billion to build data centers. They proposed a $165 billion data center. We'll invest hundreds of billions of dollars into computer.

• 22:55Speaker 01

  To invest an initial $100 billion, but it could be up to half a trillion over the next four years. These new AI data centers would be gargantuan, much bigger than any built-in Loudon in the cloud era. And they would have grand names pulled out of myth and science fiction, the kind of names you'd expect to see slapped on the side of a Star Destroyer. This is the site of Project Star.

• 23:19Speaker 04

  The project is called Project Jupiter. Project Zodiac. Prometheus, coming online, he says, in 2026. There's also Hyperion that he says will scale to 5 gigawatts over several years. It's hard to put that in context.

• 23:34Speaker 01

  These data center projects together will cost trillions of dollars, all spent really on the same gamble. The tech moguls have tied the future of their companies, as well as an oversized portion of the American economy, to the notion that all of this, in the end, will make sense. And while I do not have access yet to information about the future, I started to take a closer look at these enormous data center projects, trying to figure out for myself, does it make sense? Could it make sense?

• 24:08Speaker 01

  I ended up focusing on one, the biggest and ballsiest of all data center projects, the first of its kind from a total upstart in AI, a tech CEO who has never met a race he did not want to win.

• 24:27Speaker 04

  Elon Musk is starting a new company to get into the field of artificial intelligence. Musk announced the launch of XAI today. He says-

• 24:36Speaker 01

  Chapter four. Elon enters the ring.

• 24:42Speaker 01

  In the summer of 2023, eight long months after the big ChatGPT release, Elon unveils his AI startup.

• 24:49Speaker 06

  According to its website, the new company will work with Twitter, Tesla, and other companies to understand the true nature of the universe.

• 24:56Speaker 01

  If the AI race is defined by founders making incredibly hyperbolic claims about what their chatbots are going to do to society, Elon, of course, out-hyperbole them all.

• 25:06Speaker 06

  So the mission of XAI and Grok is to understand the universe. We want to understand the nature of the universe so we can figure out what's going on, where are the aliens. What's the meaning of life? How does the universe end? How did it start? All these fundamental questions were driven by curiosity.

• 25:23Speaker 01

  To most of the people who'd been following the AI race, Elon Musk did not seem like the entrant to bet on. It wasn't just the overly lofty promises. It was that unlike the competitors, Elon was a have not. He did not have his own data centers. He did not have a big head start. Here's Anisa.

• 25:42Speaker 05

  One of the biggest questions I had was, how could they possibly catch up to OpenAI? And I think a lot of people were sort of questioning like, OK, how is this startup that doesn't have a lot of money going to get enough GPUs to train AI that's even competitive? And if they don't have a large data center and they don't push out their training runs as quickly as possible, OpenAI is going to get even more steps ahead of them. And so Musk had to prove that even though they were late to the game, that they could catch up and be competitive.

• 26:17Speaker 01

  I should say here, we emailed XAI for comment on this story, and we received their standard auto-response, quote, "Legacy media lies." They were not going to talk to us. However, Elon Musk has talked publicly about XAI a fair amount. For instance, the day he announced his startup, he seemed to echo what Anisa just said, that he was beginning this race from far behind.

• 26:41Speaker 06

  XAI is really just kind of starting out here. So it's not a, you know, it'll be a while before it's relevant on a scale of the sort of OpenAI, Microsoft AI, or the Google DeepMind AI. Those are really the two big gorillas.

• 27:01Speaker 01

  For Elon to catch up to the big gorillas, he had to very quickly train up his young chatbot, Grok. Elon believed that the more GPUs Grok was trained on, the more intelligent it would be. So he wanted 100,000 GPUs, just to start. On another live broadcast, Elon talked about what happened next. How he went around to companies like Oracle to ask how quickly they could give him the setup that he wanted.

• 27:26Speaker 06

  We actually weren't intending to do a data center ourselves. We were gonna just, we went to the data center providers and said, how long would it take to have 100,000 GPUs operating coherently in a single location? And we got timeframes from 18 to 24 months. So we're like, well, 18, 24 months, that means losing as a certainty. So the only option was to do it ourselves.

• 27:51Speaker 01

  The only option was to do it themselves. It was a moonshot. Elon, obviously, had never built a data center himself. And because he wanted to build something faster than anyone thought was possible, he decided to try something new. He would find an existing building, an abandoned factory with immediate access to a lot of electricity. And then he would engineer a way to cram that structure full of Nvidia chips. For that plan to work, the factory would need to be in a place willing to cooperate, and cooperate fast.

• 28:30Speaker 01

  How you doing? I'm good, how are you? Good, nice to meet you, I'm person. Hello, hello. Recently, producer Garrett Graham and I flew down to one such city, Memphis, Tennessee, where we visited a white spacious office. Through a big window, I could see the river. Yes, this is the Mississippi. It's funny, we've been in Memphis for 48 hours and hadn't seen the Mississippi yet. I've been trying. We were about to meet the person in charge here. So let me start by just having you introduce yourself. So tell me your name and what you do.

• 29:03Speaker 06

  Ted Townsend, president and CEO of the Greater Memphis Chamber.

• 29:07Speaker 01

  Ted, a very stylish man, dapper suit, prayer beads on his wrist. Ted is to Memphis, a version of what Buddy Reiser is to Loudoun County. His job is to convince companies to set up shop here. And in particular, there was an empty factory by the Mississippi River that he and his team had been trying to fill.

• 29:25Speaker 06

  We'd been marketing this site in Memphis that was home to Electrolux.

• 29:31Speaker 01

  A very large factory, previously used for producing kitchen ovens.

• 29:36Speaker 06

  Eight hundred thousand square foot facility that they had decommissioned and moved out of. Great space. It was probably one of the hottest buildings in the Southeast market. We were showing it to a lot of companies, a lot of major brands.

• 29:52Speaker 01

  And give me a sense, just for a little bit of context, because again, it's my first time in Memphis. How much did this city need economic development?

• 30:00Speaker 06

  Oh, it's in desperate need of that. You know, when you consider going back to the 2008 recession, you know, it took Memphis 10 years to recover from that recession, and that was four years longer than the national average. And we've had generational poverty in some neighborhoods in our community, and we want to solve for that.

• 30:24Speaker 01

  Memphis did have some old school industry. There's a steel plant in town, an oil refinery, FedEx is headquartered here. But Ted is particularly interested in tech, the part of the American economy that year after year still seems vibrant. He was interested in tech, but tech companies had not been as interested in Memphis until March of 2024, when Ted learned that a mysterious tech company was asking after the old oven factory. As his meeting invitations started getting accepted, he could see the email domains at Neuralink, at SpaceX, at X.com. Ted says this first encounter with reps from across Elon's Hydra-headed tech empire,

• 31:08Speaker 01

  it was exhilarating. And within days, he got word that Elon himself wanted to meet.

• 31:14Speaker 06

  It was incredibly thrilling. You know, you, you open up your, in my case, my iPad and the, the boxes start to populate, you know, in the virtual meeting and there in the top left corner of my screen is Elon Musk and it was a seminal moment. It was a pass or fail, right? Someone like Mr. Musk moves quickly. He discerns information rapidly and he was focused on making a decision very quickly. So I was told by his team to, you know, do your thing Ted, sell Memphis and then hand it over to the utility partners to talk about power. And so I, I began and Elon quickly cut me off and said, I don't need to be sold on Memphis. I want to talk about power.

• 31:58Speaker 06

  And I, I thought I messed up, but handed it over very quickly to my great partners and we ended it. I was still a little uneasy about how abruptly I didn't get a chance to sell Memphis, but I got over that when it was Sunday night, we got the call that he had given the green light to his team to go with Memphis. So in a week's time, we had won that project. You know him for X, Tesla and SpaceX, but now Elon Musk is coming to Memphis to build the world's largest supercomputer.

• 32:32Speaker 01

  An unlikely story if there ever was one. This is a Chapter five. Play to win or don't play at all. Elon Musk would name his data center Colossus, reportedly after a science fiction novel from the sixties in which scientists build a supercomputer named Colossus, which later enslaves humanity. The Memphians were just going to call it a supercomputer. Whatever it was, whatever it was going to be, a science fiction future had arrived in the Mississippi Delta.

• 33:01Speaker 07

  Ted Townsend, president and CEO of the Greater Memphis Chamber says this is single-handedly the largest multi-billion dollar

• 33:09Speaker 01

  Investment by a new market company in Memphis history. Ted held a triumphant press conference.

• 33:16Speaker 06

  This monumental recruitment investment marks a pivotal point in our city's trajectory and will drive continued entrepreneurship and ingenuity, propelling Memphis to the forefront of global innovation and competitiveness. So Memphis, are you ready?

• 33:37Speaker 06

  Let's go.

• 33:46Speaker 01

  Colossus could bring tens of millions in much needed tax revenue for Memphis. But for all of this to work, Memphis had to move fast. Not city government fast, not even tech company fast. They had to move at Elon Musk's definition of fast. The city and county would grant permits at record speed, put agreements in place for power and water. Elon's people would gut the inside of the oven factory and bring in racks to house Nvidia GPUs. Even before Ted had publicly announced the project, he remembers touring the construction site with Elon and being astounded at the rate of progress.

• 34:23Speaker 06

  Everyone was just amazed at what was built. You know, Elon did have his phone out and he was taking videos and I was texting my team back here at the chamber. I was like, has he posted? Has he posted? So he did not, but I was invited into a room where it was after the tour. And everybody sits and Elon is standing at the front. And he's pacing just a little bit and there's a hush that falls over the room. And he says, we're going to haul ass in Memphis. And he thrust his fist in the air and he says, LFG. And he said it just like that. He didn't, you know, he used the letters. It gave me chills. As a Memphian, I was witnessing history.

• 35:07Speaker 06

  And then from that one visit, he decided that very day, we're going to double it. We're going to use the rest of this building to double it in size. So you're talking 200,000 Nvidia chips.

• 35:22PJ Vogt?

  I got to visit Memphis.

• 35:23Speaker 06

  Yeah, yeah.

• 35:24PJ Vogt?

  You're going big on compute.

• 35:26Speaker 06

  Yeah.

• 35:26PJ Vogt?

  And you've also said play to win or don't play at all. So.

• 35:29Speaker 06

  Yeah.

• 35:29PJ Vogt?

  What does it take to win?

• 35:32Speaker 06

  For AI, that means you've got to have the most powerful. And compute.

• 35:38Speaker 01

  Elon appeared on the Lex Friedman podcast that summer in 2024. He talks about the Colossus build. He describes being on the ground there and how he actually did some of the cabling himself.

• 35:48Speaker 06

  I try to do whatever the people at the front lines are doing. I try to do it at least a few times myself. So connecting fiber optic cables, diagnosing a faulty connection. That tends to be the limiting factor for large training clusters is the cabling. So many cables.

• 36:06Speaker 01

  This one episode is eight and a half hours long, which leaves the two of them plenty of time to nerd out on the complexity of Colossus's cabling.

• 36:15Speaker 06

  That was a crazy cable layout.

• 36:18PJ Vogt?

  It looks pretty cool.

• 36:19Speaker 06

  Yeah.

• 36:20PJ Vogt?

  It's like, it's like the human brain, but like at a scale that humans can visibly see. It is a brain.

• 36:27Speaker 06

  I mean, the human brain also has a massive amount of the brain tissue is the cables.

• 36:33PJ Vogt?

  Yeah.

• 36:34Speaker 01

  There's something about just listening in on Elon taking a break from his data center sprint to sit for one of Lex Friedman's interviews. The two expressing wonder in matching monotones. It's easy to make fun of them, and I might be just a little bit. But the truth is, the thing that they are describing is remarkable. I looked at a photo posted on the XAI account showing the Colossus interior, this meticulous Byzantine setup. And while I did not see a human brain, I do see something new and impressive. Rows of tightly packed, dense, yellow cables which line the ceiling. It's a much more complex interior than the data center I saw in Loudon, with a much more elaborate cooling system.

• 37:19Speaker 01

  GPUs run very hot. And so instead, Colossus requires liquid cooling. Cold water is piped directly into the servers. And inside those servers sit the Nvidia chips, sleek, sophisticated black squares. The whole setup looks very expensive, and it is. In a pre-AI cloud data center, a typical rack full of CPU servers might cost a few hundred thousand dollars. A single rack here, packed with Nvidia H100 GPUs, might cost a few million dollars. A single rack. XAI never said publicly what all this cost, but a JP Morgan report estimated that Colossus took between $9 and $11 billion.

• 38:03Speaker 01

  And up to $7 billion of that was just on the AI hardware.

• 38:09Speaker 01

  It surprised me to learn this, an essential fact about these AI data centers. The GPUs really are the most expensive part. They can be more than half the total cost of the data center. More than the land, the power infrastructure, the cooling. It's really the chips. But in the immediate press all around Colossus, the thing people seemed to focus on was just how quickly this data center had been completed.

• 38:32Speaker 05

  The company says that they got it done in 122 days.

• 38:36Speaker 01

  Again, Anisa, the reporter from The Information.

• 38:39Speaker 05

  A lot of people at the time and still today debate just how quickly they did it and what was really up and running when they said the 122 days. But the impact of that data center on the entire industry was very palpable. People were making phone calls to their contractors. There was a group that flew a spy plane over the Colossus data center to see what was going on.

• 39:01Speaker 01

  A group that flew a spy plane? I'm sorry, what? Like, do you mean like one company? Or do you mean a bunch of CEOs got together and flew a spy plane?

• 39:10Speaker 05

  One, one data center developer flew around the site and snapped pictures of what was going on because it was that important. They were like, this is going to be worth our time and worth the money. We, we need to see how Musk did this because people were seriously so shocked.

• 39:29Speaker 01

  For Elon's competition, his speed was shocking in an inspiring way. But in Memphis, that speed would soon read differently. Elon had made a crucial decision that had helped him get Colossus operational as fast as he did. And that decision became a flashpoint. Tinder for a huge fight that would swallow entire neighborhoods in Memphis.

• 40:07Speaker 01

  Underneath it all was a question. What happens when the city you live in becomes a battleground in somebody else's AI race? That answer, and a look at the billions being poured into Colossus, where that money is coming from and what it means for the rest of us. All of that in our next episode, which is available right now.

• 40:50Speaker 01

  Search Engine is a presentation of Odyssey. It was created by me, Shruti Pennemaneni, and PJ Vogt. Garrett Graham is our senior producer. Emily Multer is our associate producer. Theme, original composition and mixing by Armin Bazzarian. Fact checking this week by Mary Mathis and additional production support from Kim Kupel. Special thanks this week to Chris Gilberti. Our executive producer is Leah Reese-Dennis. Thanks to the rest of the team at Odyssey, Rob Morandi, Craig Cox, Eric Donnelly, Colin Gaynor, Maura Curran, Josefina Francis, Kirk Courtney, and Hillary Schuff. If you would like to support our show, get ad-free episodes, zero reruns, and some extras,

• 41:34Speaker 01

  please consider signing up for incognito mode. You can join at searchengine.show. Please follow and listen to Search Engine wherever you get your podcasts. Thank you for listening.