Take note of the name Stargate because I think you're going to hear a lot about it in the future. A new American company that will invest at least $500 billion in AI infrastructure. The data centers are already under construction. The first ones are being built in Texas, in the city of Abilene, which is our first site.
Abilene is a town in the central plains of West Texas. Here we fight for every good thing that comes our way. Some people used to ask me, 'When will the president come to see you?' And I would answer, 'He still hasn't written to me.' I'm very excited to be able to do this in the United States. I believe this will be the most important project of this era.
Welcome to Abilene, Texas. This is the site of Project Stargate, a very mysterious, talked-about, and speculated project that brings together OpenAI, SoftBank, and Oracle. If everything goes as planned, every building on this red land behind me will be filled with chips that will power the AI revolution.
We are experiencing the construction of the largest infrastructure in the history of humanity. It's going to create new ways to do business and innovate, and it's really going to push humanity forward.
I was in the real estate business. These big, beautiful buildings are going to employ a lot of people.
How do you personally deal with the speed of all this? It feels like everything is moving very fast.
I think if I could transport myself back three years, it would seem like unimaginable progress. This is an explosive and amazing growth of intelligence appearing in our decade. Everything is going to change.
When Stargate was first announced, there was a lot of skepticism. Many believed it was too good to be true. This is your first look.
Thanks for having us. Welcome in.
I know. How many people are here?
Today there are about 2,200 people working here.
The key man in building Stargate is Chase Lockmiller. He is the founder of Crusoe, a little-known data center startup that is overseeing this massive project. We started construction here last June, in 2024. This was just a completely barren field. There was nothing here. There were just a lot of mesquite bushes and shrubs.
Stargate is very mysterious. What really happens here?
Here we will have eight separate buildings. All with large clusters of GPUs fully interconnected.
Yes. Let's take a tour to see some of that. This will take us through the entire 500,000-acre campus.
We're going on an adventure. It's like touring Jurassic Park.
Alright. Let's go out into the field.
It's like a data center safari.
Yes, something like that.
The eight planned buildings will house up to 400,000 chips, making it one of the largest computing clusters in the world. The partners behind Stargate, OpenAI, Oracle, and SoftBank, have already invested $100 billion in the project and claim to be ready to invest five times more as they expand. The first site in Abilene is nicknamed 'Project Ridiculous,' as in ridiculous speed. Ridiculous speed, which is appropriate. Everything is supposed to be finished by mid-2026. They work 24 hours a day.
We work all day. All day. Every day. We are trying to meet the fastest deadline for a 100 MW or larger data center.
And why the ridiculous haste? Because Stargate is part of a much larger AI race. Companies like Microsoft, Meta, Google, Amazon, and xAI are building AI data centers at sites across the United States and beyond. Each company is seeking to make the most capable AI available to the most people and has committed tens of billions of dollars to do so.
We need to be able to increase computing capacity 100,000 times in a few years.
I'll put $80 billion. I will invest $80 billion for building Azure.
Falling behind means being out of the game for the most important technology of the next 10 to 15 years.
Beyond improving AI, many tech leaders are seeking to create AGI, an artificial general intelligence capable of thinking and learning like a human or better. Large language models like ChatGPT have become more capable as the technology has evolved. So the expectation is that adding an astronomical amount of computing will make them even more capable.
Take me back to the beginning, before Stargate. What made you realize that we need more computing, energy, and connectivity, and that you were hitting a limit and needed to scale?
We used to think a lot about the computing capacity we would need to train the models. But we didn't think as much about how much people would use these models. We did consider it, but it turned out people wanted to use them much more than we imagined. And a couple of years ago, perhaps after the launch of GPT-4 in ChatGPT, we realized this had gone from requiring a lot of computing to becoming the largest infrastructure project in history, and from there Stargate was born.
Why did you call it that?
It started as a codename and it stuck. But it means something, right? The design of one of the first data center schematics looked like the Stargate from the series.
How did you get to talk to Masayoshi and Larry Ellison and how did everyone come together?
I took two fairly long trips around the world, and a lot of that research was to talk with developers and governments, but it was also to try to understand the supply chain well. On one of those trips, I met with Masa and we talked about what would be needed to achieve computing at this scale.
Hello, Masa, good to see you. What conversation did you have with Sam that led to Stargate?
Well, we have a shared vision that AGI will arrive soon and we need computing capacity.
Of all the bets you could make, why Stargate?
I firmly believe in the vision that this artificial intelligence, AGI, will change the life of humanity in all aspects. We are very fortunate to be able to witness this enormous revolution and I would like to be part of it. It excites me.
Last month I came here to celebrate your victory and I promised we would invest $100 billion in the project. And you told me, 'Masa, why don't you invest about $200 billion?' It's true. Now I'm back with $500 billion.
Explain to us how you get to needing $500 billion.
That covers what we believe we'll need in the coming years, given our growth projections. The question is, if we could get a trillion dollars today, which is not the case, could we use it profitably in the coming years? I'm not sure, but I believe we can generate $500 billion in value.
Tell me, what are we seeing now? There are many forklifts, cranes, dump trucks, and excavators.
At every step, new creatures emerge. Most of the workers are from Abilene. We are a very large employer here in Abilene, but we have also brought in a lot of labor from out of town and out of state. We have workers from seven different states.
I had never built anything of this scale.
An underappreciated aspect of AI is the physical infrastructure it requires.
This is incredible. I love it. We are building things.
Yes, there are 155 vehicles moving earth and excavating right now.
There is something very satisfying about seeing so much earth being moved.
Tell me, how does it compare to other cranes?
That one can lift a lot, a lot of weight. 600 tons, that's like 100 elephants.
This is happening all over the world. Data centers are being built in Malaysia, Japan, and France.
I think the demand for computing infrastructure is not an American phenomenon, it's a global phenomenon. Many countries have decided that data centers are critical infrastructure that must be accelerated.
If the hyperscalers are going to Malaysia or Japan, are they competition for you?
We don't see it as competition, we see it more as overall growth of the sector. We believe we have many unique ways to do it and a lot of value to bring in the process. And honestly, we're excited about everything happening in the sector.
Lockmiller's path to leading the construction of Stargate has been winding. He climbed five of the so-called Seven Summits, the highest mountain on each continent, and worked in physics and finance. Tell me about your background before Crusoe. You grew up in Denver and started using AI as a trader on Wall Street.
Yes, we were pioneers in deep learning and we used these neural networks to predict the returns of certain assets. We found they were quite effective. I had a strong conviction that AI was going to transform everything and I always felt it was a meta-science.
Lockmiller left finance to get into the crypto world, not to get rich quick promoting meme coins, but with hundreds of computers solving mathematical problems. His idea was to set up Bitcoin mining operations near oil wells, taking advantage of a particularity of the extraction process. When an oil company drills for oil, an associated byproduct is natural gas. And when they don't have access to a pipeline, all that associated gas ends up being flared on site. So, we thought that instead of taking that gas to a market to sell it, we could bring the market to the gas. We could build mobile and modular data centers and bring them directly to the well to power the data center with that energy. It was a good idea at a bad time. The cryptocurrency market had collapsed, so you went from Bitcoin mining to AI. It sounds like going from one fad to another.
Yes, it might give that impression, but the truth is that from the beginning we had planned to build an AI infrastructure business.
How do Bitcoin mining and AI data centers relate?
First, both consume a large amount of energy. Second, both are more location-independent than traditional data centers.
Northern Virginia is widely considered the world epicenter of data centers today. That is data center 1.0. Yes, that's it. The Virginia data center corridor is where the first generation of cloud computing was built near the government agencies that created the original internet. Those data centers are full of CPUs, basic chips that power everyday things like social media, streaming, and the cloud. Stargate, on the other hand, is part of a new era of data centers with GPUs, advanced special chips needed to power AI and perform multiple tasks at once. So, is this the inside of a data center?
Yes, of a data center under construction. Okay, everyone is working. It's a very active construction site. It's a very active construction site. You can take it through here and then down the hallway.
People are working very hard and fast.
Nobody likes it when I show up to visit data centers.
It's not my area of expertise and I still have opinions. So they say, 'Leave us alone.' We're going to go into one of the GPU data halls, where we'll find a very large cluster of Blackwell GPUs. Blackwell is Nvidia's newest chip optimized for AI. Once the eight buildings of Project Ridiculous are complete, each will house up to 50,000 chips that will be installed by Stargate partner Oracle.
Wow, this is where the GPUs will go.
This is where the GPUs are going to go. I was expecting this.
This is where the AI factory will be.
Okay. This is what an AI factory looks like.
And it really is like a factory, but the difference is that the final product being made here is intelligence. And the sooner we start this AI factory, the better for OpenAI. In March, the company launched a new and improved image generator for ChatGPT, which quickly became almost too popular to handle.
I've seen viral moments, but never something with a usage spike like that.
They added a million users in an hour.
Yes, and more in some hours, but yes, it was unprecedented, a madness. And also, creating an image is not exactly a low-compute task, as we do with the new image generator. You tweeted that the GPUs were melting.
I didn't think people would take it so literally.
I understand it was a joke, but it's meta.
They were running at very high temperatures, but the metal wasn't melting. So, we had to do many unconventional things. We borrowed compute capacity from the research area and throttled other functions, because it's not like we have hundreds of thousands of GPUs sitting around doing nothing. If we had more GPUs, we could handle these demand spikes better. Like this one. It's directly related. Yes, more computing capacity, more AI.
Something important to keep in mind is the amount of activity overhead. There you can see those huge pipes that distribute cold water to all the individual chips to cool them. Water is needed because GPUs generate a lot of heat and keeping them cool normally involves an evaporation process that can consume millions of gallons of water a day. This has raised concerns about the sustainability of the AI boom. At Stargate, Crusoe says it has avoided this problem by using what they call a closed-loop system that keeps the water permanently enclosed within the pipes.
You can put in a million gallons and then you don't need any more water.
Yes, that's correct. We fill it once and that's it. And that's easily coordinated with the city to get a million gallons of water on a single occasion. It's very different from asking for a million gallons per hour or some other outlandish request that might be the case if you had an open-loop cooling architecture.
Once all this is built, what is the big vision?
To give people tools so they can do better whatever they want to do. People are going to use it and they're going to be able to unleash their creativity to create all kinds of things they're interested in or just use it to entertain themselves. Personally, the area that excites me the most is AI for scientific discovery. We're not there yet, but we're not far. I think 2025 will be a time when agents do a lot of the work, but tasks we already know how to do. I hope 2026 will be a great year of real new scientific progress.
To achieve that, buildings full of chips are not enough. Energy is also needed, a lot of energy. That could be a problem for the climate and a limiting factor in the advancement of AI. Sam Altman has been seeking more computing capacity and it seems energy is a major constraint. Can you explain to me what the constraints are?
Let me put it in perspective for you. If you look at a data center from 20 years ago, a single rack in the data center was typically budgeted to consume between 2 and 4 kW. Today, with the most advanced Nvidia configuration, we budget 130 kW per rack. We're talking about an increase of nearly two orders of magnitude. So, asking ChatGPT a question uses approximately 10 times more energy than a Google search. It's a total transformation in the infrastructure needed for this workload and that has become the main bottleneck.
So, AI data centers can't be located just anywhere. They need a place with a lot of available energy, and West Texas turns out to be ideal.
We came to Abilene because it's a community very rich in clean, low-cost energy. A lot of wind energy has been developed in the Abilene region and its surroundings, and what they lack is really energy demand. That created a huge opportunity for AI, which is currently hungry for energy.
This is the first 200 MW substation. Right here, an additional 1 GW substation will be developed. So the total capacity of the site will be 1.2 GW.
We're talking about more than 1 GW.
That's right. It seems everyone wants 1 GW or more these days.
If that sounds like a lot, it is. Enough to power about 750,000 homes, 2,600 Tesla Model 3s, or 100 million LED light bulbs. There's a funny quote from Back to the Future where it says 1.21 GW.
1.21 GW is a pretty funny coincidence.
As more data centers come online in the coming years, they will add more gigawatts of demand to an already overloaded grid. It is estimated that data centers could consume more than 8% of U.S. electricity by 2035. While the International Energy Agency predicts that renewables will cover about half of that demand, the rest will come mainly from coal and gas. Crusoe is building its own gas power plant as a backup. Every major tech company has a promise of zero emissions by 2030. Are our AI ambitions going against our climate goals?
It's complicated. My personal perspective is that none of those promises will be met by 2030. And there are ways to adjust those numbers, like, for example, buying a lot of certificates or offsets, but there's no way the energy being used will be completely carbon-free to power all the AI infrastructure being developed. It's simply impossible. So, companies adjust the numbers.
Look, I think what's happening is that many of these goals were ambitious and were set before the demand that AI infrastructure is now generating. So I think AI has really changed its course. Now, my optimistic side says that even though it might get worse before it gets better, I believe there's always light at the end of the tunnel. And I think the fact that AI is an enormous tool for invention is going to greatly accelerate the development of new technologies. The fact that we are building a lot of new infrastructure that requires a lot of energy actually opens a great opportunity for AI to be one of the first to adopt new energy solutions, things like small modular reactors, geothermal, nuclear fusion. AI generates an enormous demand for energy and that puts it in a position to deliberately decide what the energy infrastructure that will power it will be like.
For now, Stargate seems to have the resources it needs to achieve its goals, but with big investments come big risks. Brody Ford of Bloomberg has been following Stargate and the many ways it could go wrong or right for Stargate's partners and for the city of Abilene.
Thanks for coming to get a drink.
