As AI innovation escalates, it’s predicted the demand for an uninterruptible power supply will increase dramatically. Crusoe, a company specializing in AI infrastructure, uses stranded and renewable energy sources to power facilities while cutting the time to market for new sites by more than half using modular components. On this episode of the At the Edge podcast, Crusoe’s cofounder and CEO, Chase Lochmiller, speaks to McKinsey Senior Partner Lareina Yee about Crusoe’s unique approach to AI infrastructure and how AI will shape the future workforce.
The following transcript has been edited for clarity and length.
Stranded energy’s potential for powering data centers
Lareina Yee: Crusoe is on the forefront of building custom AI data centers. It also manufactures custom-built server system components, and it’s repurposing stranded energy. In short, you have created a verticalized, integrated business that is defining a new category to support the expansion of the AI economy. Tell us about it.
Chase Lochmiller: We’ve aimed to create a category-defining intelligent infrastructure company by building out large-scale factories for intelligence that are powered in unique ways at significant scales and are brought to market quickly.
The infrastructure for AI requires a lot more power than anything we’ve ever seen from the computing infrastructure world. The individual racks to power these AI computing workloads are as much as 140 kilowatts, whereas a legacy rack to support a cloud data center 15 or 20 years ago was two to four kilowatts. The next generations are going to be 600 kilowatts, and then from there, a megawatt.
That amount of power has depleted a lot of the existing energy that’s available to power data centers, and the demand for power for data centers has grown dramatically. That’s left us in this situation where we don’t have the energy that’s needed to power the demand coming from AI, so we need to build it. Crusoe has taken an energy-first approach, which entails building in areas where we can access low-cost, abundant energy resources to accelerate more abundant intelligence.
Lareina Yee: Crusoe captures stranded energy, which means building compute infrastructure on top of oil flares. Can you tell us a bit about that concept and how you came up with this idea?
Chase Lochmiller: My cofounder grew up in an oil and gas family and had spent his career building in the energy sector. I had never set foot in an oil field before we started this business. He told me about this problem that when they drill wells for oil, three things come out of the ground: oil, natural gas, and water. The water and the oil can be trucked to a water treatment facility and an oil refinery, but the natural gas, unless you have access to a pipeline, will oftentimes be burnt off because that’s the most economical thing to do.
A lot of the methane escapes uncombusted. Methane is a potent greenhouse gas that traps about 84 times more heat in the atmosphere than CO2. This became one of the worst greenhouse gas problems, so we thought: Instead of trying to bring that gas to a market where it could be useful, why don’t we bring a market to the gas? We built mobile and modular data centers that we deployed and ruggedized for the oil field. We deployed mobile power generation solutions that would capture the gas and use it on-site to generate power, and then we would run that equipment to power these mobile and modular data centers.
Subscribe to the At the Edge podcast
Lareina Yee: What are other forms of stranded energy that you use?
Chase Lochmiller: One stranded energy resource we’re focused on is the renewable energy in markets like West Texas. In West Texas, there are a lot of independent power producers that were incentivized by a production tax credit to build wind farms and solar farms in areas where they could have a high-capacity factor. They ended up going to areas where there wasn’t necessarily a market for power, and there wasn’t transmission to get that power to another market.
We’ve also considered expanding into Europe, specifically Iceland. Low-cost geothermal energy can be produced there, as well as low-cost hydropower energy coming from the glaciers. It’s far more than the 300,000 people that live there can use. Iceland also has a favorable cooling environment, so you can get high-efficiency data centers with low-power-usage effectiveness [PUE].
Rethinking the approach to building data centers
Lareina Yee: How has Crusoe thought about AI data centers differently?
Chase Lochmiller: Hyperscale data centers have sought to scale our computing needs to power the internet and web applications in central hubs, such as Northern Virginia. The AI data center has become the new unit of compute. When we think about building a data center, we’re actually thinking about building a large, 100-megawatt or gigawatt-scale computer. That’s led to different architectural designs for how we lay out the physical architecture, where the graphics processing units [GPUs] get installed, and how they get networked together.
Our data centers look like the layout of a motherboard. There are large compute data halls that are the wings of the data center. That’s where the high-performance GPUs go. Then there is a central core where the core network and storage live. The GPUs are interconnected through high-density, high-performance networking fabric. So instead of having a network interface card [NIC] on the server, it can be on the physical GPU, so it can directly access the GPU memory and shift data from one GPU to another through this high-performance backend network.
The data center is designed around this concept that all the chips in the data center are thinking as one cohesive unit. This design helps when you train a new breakthrough foundational model. You need to have a whole bunch of GPUs, not just on the same server but in the same data center, that are thinking as one cohesive unit, working toward a single problem. That leads to different architectural choices from a cooling perspective, layout perspective, and density perspective. It ultimately manifested into this data-center-scale computer.
Lareina Yee: Tell us about your project in Texas, Stargate 1, and why it’s important.
Chase Lochmiller: We chose Abilene, Texas, for our site because a lot of wind energy had been developed in the region. But that energy was suffering from a curtailment issue: People in some of the more congested nodes were getting negatively priced power 25 to 30 percent of the time. So they had a large demand for energy in this region and a significant amount of transmission that would take a decade to build. We decided to build a large AI data center here and soak up this underutilized, excess, renewable energy.
Of course, the wind’s not always blowing; the sun’s not always shining. At this scale, it is important for us to have consistent power. To help accelerate the time to market and support firm energy resources, we built a 350-megawatt gas plant on-site, as well. This combination of energy resources brings these facilities to life.
Decreasing the time to market
Lareina Yee: As demand for AI increases, what are the engineering problems and questions you still want to solve?
Chase Lochmiller: One of the things that we’re focused on is how to get to market faster. For this project, there was initially a request for proposal [RFP] for a 100-megawatt data center. That amount of power can power around 100,000 homes in the United States, so it’s a big footprint. There were 32 providers that bid on this initial RFP, and the fastest anybody was able to deliver this data center was in a little over two and a half years. We came in and said we could do it in 12 months. We delivered it in 11. The demand for infrastructure capacity to support the ambitions of AI is insatiable right now, and we’re at a moment where everybody wants it yesterday.
Lareina Yee: How did you approach the build differently to speed up time to market?
Chase Lochmiller: Energy is one of the long lead items, and we were able to access that. We also parallel processed the construction, development, and on-site work as much as possible, and we pushed as much as possible off-site. That is, we build a lot of the components that make the data center modular. I kind of think of it as data center Lego blocks. The individual components that represent the guts of the data center—the electrical systems, the mechanical systems, and the plumbing systems—get manufactured off-site, are shipped, and get clipped together on-site. There is still a nontrivial amount of labor and work that goes into clipping all these pieces together, but it’s far less than if we were to build everything on-site.
This is an exciting time for innovation and development within the data center industry. It’s not that old of an industry, but it’s going through a massive revolution to support AI. Northern Virginia has sort of become the center of the world—it’s where a lot of the internet runs. At the end of 2024, the total capacity in Northern Virginia was 4.5 gigawatts. Our site in Abilene is 1.2 gigawatts. We’ve built all this capacity over the past three-plus decades in Northern Virginia to run the modern internet, and now this single site for a single customer is starting at 1.2 gigawatts and expanding from there. When you think about that, and it’s all going to get built in a very, very tight time frame, the full capacity will be built out over 24 months for the 1.2 gigawatts by mid-2026.
How AI will shift job opportunities and local economies
Lareina Yee: How do you think this infrastructure will impact jobs and the local economy?
Chase Lochmiller: There’s a significant number of construction jobs created during the development phase. With numerous expansions, we think those jobs are going to be sustained. Manufacturing jobs for modular components have also increased, as well as transportation jobs to get those components to site to support the development.
For context, we have about 5,800 workers on-site every day in Abilene, Texas. Abilene has a population of about 120,000, so that’s a substantial portion of the people that live in the area. We’ve been able to source about half of the employees in Texas, and the other half comes from all over the United States. Crusoe has factories in Denver and Tulsa, and we partner with third parties in the United States that help us manufacture the key modular components. We’re also standing up new capacity in other countries, including Canada. We see a huge boom in manufacturing jobs to help support this infrastructure. We plan to hire over 1,000 people in Tulsa to fill skilled-trade jobs, such as electricians, plumbers, welders, and painters, that are important to build the infrastructure for intelligence.
There are many other companies making big investments and managing big development opportunities all over the United States. That will create thousands of high-paying jobs. In the long term, jobs managing the infrastructure are being created, from managing the power plant as new power generation resources come to life to managing the data centers and chips. Every node has tons of networking cables that all have transceivers and optics, and on-site technicians will be needed to replace anything that fails rapidly. That could mean thousands of new technician jobs to support this new industry.
It’s important that we get ahead of training the workforce to be prepared for this. I think there’s an important economic shift underway. We need to reskill the labor force to support the infrastructure needs of AI, which means more electricians, data center technicians, and plumbers. Everything is shifting to liquid-cooled architectures, which require complex plumbing systems that operate within the data center.
Lareina Yee: What types of trade programs and schools would we need to invest in to support this infrastructure?
Chase Lochmiller: For electrical, we need a lot more folks who have low-voltage training because data halls are a low-voltage ecosystem. We also need people with medium-voltage and high-voltage experience because the grid needs to be modernized to include more high-voltage and ultra-high-voltage transmission.
On the data center side, we need people who can do the on-site operations, maintenance, and support to make these systems work. That includes data center technicians, people that understand these high-performance networking architectures and how to operate these intelligent clusters. Being able to support the agentic economy requires a lot of on-site data center technicians to make sure that the infrastructure’s there to support it.
Lareina Yee: Let’s use that to flip to some rapid-fire questions. You have an incredible education background. You actually met your cofounder in high school. You went on to MIT to study physics and math. You then went on to Stanford to do a master’s in computer science and AI. There are a lot of questions at the skills and trade level as well as at the knowledge worker level. What should people be studying and learning to equip themselves for an AI-first economy? What advice do you have for people in high school who are debating college or who are computer science majors?
Chase Lochmiller: This is something I think a lot about, having three young kids of my own and wondering what the economy will look like when my kids are entering the workforce. One thing people need to appreciate about the future is that, at our fingertips, every human will have access to this digital labor, agentic economy. As a result, any individual with any amount of agency can produce an unprecedented amount of economic value. They won’t even need to hire people. They can hire digital labor. From that standpoint, I think it’s not necessarily about having one specific skill but rather developing a sense of autonomy, high agency, focus, and curiosity.
More people will be going to university to get college degrees than will be needed, especially where the labor market is moving. We’re hiring a lot of people who don’t need college degrees; they need trade school degrees or apprenticeships for critical trades to understand how to operate these tools and build these physical components. I think there are huge opportunities for folks to avoid student debt and create great careers for themselves by taking these high-paid skilled-trade jobs.
Bringing a sense of curiosity and autonomy to the future
Lareina Yee: You mentioned that you have three kids. You have climbed Mount Everest. You started this company with your high school friend. You were recently named one of the most important influential AI leaders in the world by Time Magazine. How do you balance it all? More specifically, how do you have time to think?
Chase Lochmiller: I don’t think anybody would look at my life and call it balanced. My life is very busy. I had a lot of hobbies when I was younger, and I have virtually zero hobbies today. I spend almost all my time either working on Crusoe or with my wife and my three kids. I love spending time with my kids and helping to foster their sense of curiosity.
Lareina Yee: What sparked your curiosity to create Crusoe and reimagine how to power AI?
Chase Lochmiller: After grad school, I spent time as an AI practitioner. I had a firm belief that AI was a metascience that could be widely applied to the entire economy, not just for specific use cases. There’s no use case or economic or human experience that can’t be improved by adding artificial intelligence. If there was going to be a massive proliferation of AI, it would require a lot more infrastructure, and the energy needs of it would scale dramatically.
It’s been a fun journey for me because I’d historically worked on more digitally native problems that were more data-focused or engineering-focused, with algorithms and computers. There’s something innately satisfying about building something physical as an entrepreneur. It’s embedded in our humanity or something.
Lareina Yee: In the future, what is the book you’re most excited to read with your kids?
Chase Lochmiller: I was a big Harry Potter fan when I was a kid. I always loved how it exposes you to this magical world and this coming-of-age story of a hero. I’m excited to read it to them as they get older and can appreciate it.
