Crusoe Energy got its start harnessing oilfield flare gas to mine bitcoin. So how did cofounders Chase Lochmiller and Cully Cavness end up building the first phase of the world’s biggest AI datacenter – $500 billion Project Stargate – for OpenAI and Oracle?
It’s a windy, 80-degree, early March day at the 1,000-acre construction site in Abilene, Texas codenamed Project Ludicrous, where 2,000 workers toil over what will soon be one of the world’s largest artificial intelligence datacenters. Here under its official name — The Stargate Project — OpenAI will team with Oracle and Microsoft to create what they claim could be the first generation of AI models to fully surpass human cognition. It will be a “factory of factories,” says Chase Lochmiller, CEO of lead developer Crusoe Energy. What he means is that AIs trained here will be used to develop more such systems down the line, producing an endless stream of intelligence-as-a-service.
By the end of this year Crusoe expects 100,000 Nvidia graphic processing units (GPUs) to be installed into each of the first two 500,000 square foot “H”-shaped buildings. Lochmiller, just has to find the right path, both for 8-year-old, $300 million revenue Crusoe, as well as for the Polaris jeep navigating around cranes and trucks on the dusty site. “It’s hard to keep up with the topography of this site. I was here three weeks ago, it didn’t look anything like this,” he says.
Crusoe cofounder and COO Cully Cavness
Cody Pickens for Forbes
Each Stargate datacenter building will draw 100 megawatts of electricity, enough for a city of 100,000. Most of that power will come from the hundreds of 300-foot-tall wind turbines spinning in west Texas, which on most days generate far more electricity than the region needs. Windmill owners with a surplus sometimes even sell power into the grid at a loss, making it up by collecting federal green energy tax credits. “Not a great use of those electrons,” says Lochmiller, 38, who is working to arbitrage that cheap juice for Stargate, which could eventually consume several gigawatts in Abilene. And in case the wind should stop blowing or the Texas grid goes down, Crusoe has set up five gas turbines from General Electric Vernova for backup.
Crusoe raised $3.4 billion in joint venture funding with asset management company Blue Owl Capital and data center-focused investment firm Primary Digital Infrastructure for these first two buildings. But that’s only a sliver of the $500 billion Oracle founder Larry Ellison, Softbank CEO Masayoshi Son and OpenAI CEO Sam Altman promised in January to invest over the next five years to build the future of AI and secure America’s global advantage. Sharing a drone shot of the site on X, Altman enthused over the “big, beautiful, buildings.” That’s the kind of thing that appeals to President Trump, who in his February speech to Congress lauded the investment, “Which they wouldn’t have done if Kamala had won,” he said.
But why pick relatively unknown Crusoe to build it? Oracle, Microsoft and OpenAI wouldn’t comment, but it’s clear, at least to Crusoe’s investors and partners, that Lochmiller and cofounder Cully Cavness had the drive to build datacenters bigger and faster than they ever had before, and faith that they would find enough energy to power them. A successful outcome is far from assured, and some Crusoe clients are frustrated about what they say is the company’s less-than-reliable existing cloud computing offerings. Even if this first phase of Stargate, which will eventually comprise 10 buildings, representing 1.2 gigawatts of computing power at a cost of an estimated $15 billion goes off without a hitch, how could Crusoe possibly keep up the momentum and supply chains for AI’s insatiable appetite for compute and power?
Forbes
Forbes 2025 AI 50 List
The Forbes Artificial Intelligence 50 List of 2025 spotlights promising AI-driven businesses. See the leaders driving the future of the industry.
VIEW THE FULL LIST
Another hitch: President Trump’s sweeping tariffs imposed on some 90 countries last week (now paused), which could make key materials like aluminum and structural steel more expensive and ratchet costs for datacenter developers like Crusoe. The company says it sees it as a competitive moat. “From the beginning, Crusoe made the decision to onshore the manufacturing of critical electrical components including switchgear because we believe in building things in America,” Lochmiller says.
“There’s a shiny factor with Crusoe,” says consultant Daniel Golding, who calls them “the flavor of the month.” After all, Meta, Amazon and Alphabet have separately earmarked more than $200 billion for AI infrastructure. Not all will get built, says analyst Zach Krause at East Daley Analytics in Denver. For instance, Microsoft in March announced a slight pullback in datacenter leasing. Krause figures those that do get completed will add some 80 gigawatts of electricity demand nationwide by 2030, an unprecedented 10% growth, or as much as all of Texas uses on a hot summer day. Anyone who remembers Internet 1.0 will recall the vast overbuilding in fiber optic cables and datacenters that evaporated billions when boom turned to bust.
Lochmiller and Cavness, 37, went to high school together in Denver, spending weekends rock climbing. Lochmiller, 6-foot-1 with a shock of black hair, got into AI when he went to Stanford for grad school in 2013. He was struck by Professor Andrew Ng’s contention that “AI is the new electricity” soon to be just as widely distributed and relied upon by civilization. AI also consumes massive amounts of power, a truth Lochmiller experienced first hand when running up bills to power the stock forecasting and trading algorithms he built at hedge fund KCG Holdings in the early 2000s. Meanwhile, Cavness had become obsessed with locating the world’s cheapest, greenest energy sources, visiting 25 countries like Chile and Argentina, even looking at geothermal in Iceland. On a hiking adventure in 2017, the duo decided to join forces to tap “stranded” energy sources that would otherwise be wasted to power computing. They named their company after Robinson Crusoe, a fictional 18th century character who was marooned on a Caribbean island and used his resourcefulness to survive.
Their first target: redirecting so-called flare gas from oil field operations. Sometimes when new oil wells get drilled, the owners opt to burn off relatively small amounts of natural gas rather than go to the expense of hooking up a pipeline. This wastes energy and produces harmful carbon dioxide. Crusoe’s big idea was to redirect the natural gas from flares and instead use it to power reciprocating engines to make electricity — to mine bitcoin. Early backers were skeptical; Stefan Cohen at Bain Capital led Crusoe’s $4.5 million seed round in 2019, but initially even he wasn’t 100% sold: “The business model itself was not overwhelmingly obvious to us.” Determined to prove it could be done, Lochmiller put in $600,000 of his own money to build out a shipping container crypto mining rig and set it up in a gas field near Denver.
Over time, Crusoe began building compute infrastructure in harsher terrains. Antonio Gracias of Valor Equity Partners, an early backer of Elon Musk’s Tesla, SpaceX and most recently OpenAI rival xAI, first met the team in 2018 when they were installing their first containers in North Dakota’s Bakken oil field. In the middle of a snowstorm, Gracias flew in to see what they were doing. The weather didn’t phase Lochmiller, an accomplished mountaineer who had climbed Everest two weeks earlier. Gracias, who almost set himself on fire on one of the heaters, says, “It was crazy at the time, which is what I loved about it.”
Another early investor, Lee Jacobs of Long Journey, could tell Crusoe had foresight because their original pitch deck, from when the company was focused on bitcoin mining, declared they would forge “the future of compute.” Jacobs doubled down last December, joining Crusoe’s $600 million fundraise. “They always saw what was happening — intelligence as a service delivered down a pipe like electricity.”
The company almost came to a standstill in 2020, when lockdowns spurred by the Covid-19 pandemic temporarily dried up demand for gasoline. Oil companies shut down drilling and with it the supply of to-be flared gas that was core to Crusoe’s business. The company scrambled to find new locations and found a solution in focusing on greener sources of energy.
That set up Crusoe for its leap to AI. In 2023, Crusoe borrowed $200 million to purchase coveted Nvidia H100 GPUs to expand its nascent cloud computing services and hired talent from the likes of datacenter giants 365 Main and Digital Realty. They topped off the year by partnering with Iceland’s AtNorth data center developer to set up their Crusoe Cloud services in the frigid country (making it easier to cool off GPUs). “Crusoe is building datacenters larger than Digital Realty ever did,” marvels former CEO Bill Stein, now a Crusoe board member. Key to their longterm success, he says, will be “signing 15 year leases where counterparty risk is minimal,” like with Stargate.
Crusoe’s shot at the big time came in the first quarter of 2024. Lochmiller had learned that some tech giants, believed to be OpenAI and Microsoft (he declines to say who) were shopping for contractors to build an AI-specialized, hyperscale datacenter, and do it fast. Other developers balked at the ludicrous timetable of less than two years. Land acquisition, permits, supply chain lead times, all looked impossible.
But Lochmiller and Cavness jumped on the RFP. It was less of a pivot, more of an evolution. They already had manufacturing centers in Denver and Tulsa, relationships with important suppliers like G.E. Vernova and new advisors. They were confident Crusoe could build it. As for where, they quickly subleased 1,000 acres in Abilene from developer Lancium, which had already permitted the site and intended to build a crypto mining center.
By June 2024, OpenAI and Oracle had announced a partnership where the artificial intelligence giant would use Oracle Cloud Infrastructure to integrate AI into Microsoft’s Azure platform. Jared Sheiker, managing director at datacenter investor Blue Owl Capital ($250 billion assets) was impressed with the “audacity” of Crusoe. “Locking up a long-term lease with Oracle made this project overnight abundantly attractive to us,” he says. Blue Owl contributed to Crusoe’s $600 million 2024 fundraise and helped arrange $2.3 billion in project financing from JPMorgan for the first two buildings and 200 megawatts of compute. All told Crusoe has now raised $1.82 billion in equity and was last valued at $2.8 billion.
The datacenter buildings will include four data halls of 25 megawatts of compute each. Factories in Denver and Tulsa will manufacture server racks and install GPUs into skid-mounted units that are trucked to Abilene for installation. It’s far better to do this in a controlled environment than a dusty, noisy construction site, says Lochmiller; “You don’t want tradesmen stacked on top of each other.”
The GPUs will use liquid cooling. Two-foot diameter water pipes course around the inside of the building; they will continuously recirculate 1 million gallons of chilled water through a network of pipes reaching into the racks of GPUs like capillaries to draw off heat. It’s a lot of water, but they only have to fill up the system for each building once. By comparison, 1 million gallons is less than what shale gas frackers use to complete a single well.
Crusoe and Co. are far from the only ones building facilities like this right now. Others are building bigger and faster. Elon Musk’s xAI managed to build a massive “gigafactory of compute” in Memphis in just 122 days. Dubbed Colossus, it now has a network of 200,000 GPUs to support its AI supercomputer. Lochmiller says Musk’s facility is starkly different from what Crusoe is doing: building from the ground up or what’s known as greenfield development. “The xAI datacenter, they took an existing building and then retrofitted it to sort of meet the internal electrical needs of the compute racks,” he says. “It was very creative, but just a very different project than what we’ve done here.”
Meta CEO Mark Zuckerberg says he’ll invest $65 billion into AI, mostly for datacenters, this year. His biggest is the Sucré project in northern Louisiana, a 4 million square feet datacenter with 2 gigawatts of compute expected to cost about $10 billion.
Over the longer term, datacenter developers are looking to advances in nuclear power to displace fossil fuels. Amazon has promised $100 billion of AI spending and last year invested $500 million into nuclear startup X-Energy, which aims to build 5 gigawatts of small, modular reactors on datacenter grounds. OpenAI and Softbank are big backers of nuclear fusion developer Helion, which has already presold electricity to Microsoft.
The big question: does training AI truly require so much datacenter capacity? The viral launch of Chinese AI darkhorse DeepSeek, which claimed it built cutting edge AI models that rivaled OpenAI’s top systems at a fraction of the cost and far less compute, has surfaced concerns about the billions of dollars being plunged into AI infrastructure, at a time when companies are yet to reap real ROI.
Lochmiller insists more cost-efficient AI training will unlock new use cases. “What couldn’t be improved by adding intelligence?” he says. He also envisions people will soon pay for intelligence the same way they now do for water or electricity.
But is Crusoe ready? Some customers worry Crusoe’s cloud system is unreliable. In early March one of its centers experienced an outage that lasted 45 hours. One customer, which uses Crusoe Cloud to access GPUs and asked not to be named, was affected by the outage, forcing them to fall back to other cloud providers. “This was a big deal. We were pretty unhappy about it,” they said. “They’re not as reliable.” An industry expert raised an eyebrow as well.“There’s less differentiation in Crusoe than they would want you to believe,” they explained. “But there’s enough demand where anybody that can provide power on site, which is a differentiator.”
Crusoe says its service maintains an uptime of up to 99.98%. And other customers appear to be perfectly happy. “Crusoe is faster than a lot of companies right now,” says Varun Mohan, CEO of AI coding company Windsurf (formerly Codeium.) Crusoe is able to get its customers access to hardware in a matter of three to four hours, and speed is the need of the hour amid a fierce race to launch more capable AI models. Hyperscalers like AWS and Google Cloud, on the other hand, would take three weeks and bind them in years-long reservations.
Nor is Lochmiller immediately concerned about the threat to datacenter demand from new more efficient AI models like China’s DeepSeek. An AI that uses less “compute” and therefore less energy to achieve similar results could imply weaker future demand for datacenters. That risks datacenter overbuild, and bust — like the bankruptcy-inducing overcapacity in servers and fiberoptic cables at the end of the first Internet boom 25 years ago.
Not to worry. After all, Crusoe financed Project Ludicrous on the back of a 15-year contract with Oracle and Microsoft. Satya Nadella, Microsoft CEO, argued on X this year that AI datacenter growth is immune to normal capital cycles due to something called the Jevons paradox. It’s named after William Stanley Jevons, the British economist (d. 1882) who determined that lowering the cost of a fuel (coal in his case) will only cause demand for it to rise. “Jevons paradox strikes again,” Nadella wrote in January. “As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.”
Looking toward Interstate-20 on the southern part of the site and the town of Abilene beyond, Lochmiller marvels at the procession of earth movers hauling tons of north Texas red dirt. They’re working to flatten what used to be known as Rattlesnake Ridge, where the next set of datacenters will be built. It takes an ironically vast amount of manual labor to accomplish the end goal of the AI revolution, which Lochmiller describes as “freeing up all of that mind space, that creative brain power to focus on higher level abstractions.” He expects to more than double the number of workers on site to 5,000 at peak next year.
More from Forbes
Read the full article here