In 2024, Crusoe went from harnessing oilfield flare gas to mine bitcoin to going all in on the first phase an AI data center so big it needed a codename that sounded like an inside joke: Project Ludicrous, OpenAI and Oracle’s Stargate partnership in Abilene, Texas. Investors—who have shoved more than $4 billion into the company and valued it at $10 billion last fall—applauded the audacity: build data centers bigger and faster than ever before. Now, Crusoe is going small.

Crusoe is using $200 million of that funding in the next year to expand its investment in smaller, prepackaged, modular data centers, it told Forbes. That includes unveiling a new 350,000 square foot factory in Brighton, Colorado to make its modular units, branded Crusoe Spark. Unlike the Abilene partnership, where buildings drawing more than 100 megawatts each are stitched together to form a massive 1.2 GW data center, Crusoe Spark units are about one megawatt each and a little larger than a shipping container. That makes them more suitable for inference, the unglamorous, revenue-adjacent work of actually running models like the one behind ChatGPT. Crusoe says it hopes to crank out 100 modular data centers per year, with the first ones due this summer.

In other words: Crusoe isn’t abandoning hyperscale. It’s buying itself a faster, steadier way to add capacity—and get paid.

“We’re going to go big on Spark, but we’re still in for the big hyperscale units” says Crusoe’s cofounder Cully Cavness. “I want to be clear that we’ll do both. It's a barbell strategy, but this is now saying, ‘okay, five production lines, 200 employees, 350,000 square foot factory, a couple hundred million dollars,’ Spark is going to be a real product offering.”

Crusoe sees its Spark units as addressing the next phase of AI infrastructure: not just massive training campuses like its Stargate Abilene data center, but smaller distributed, modular data centers that can be loaded onto semitrucks and then stacked together like Legos, Cavness said. Modularity means that Crusoe can deploy data center capacity in chunks, on schedule, instead of betting on a single massive build that can hit repeated “unexpected” delays because the grid can’t, the crews can’t, or the locals won’t.

Have a tip about Crusoe or another data center company? Contact Anna Tong at atong@forbes.com or 650-468-3913 on Signal and Phoebe Liu at pliu@forbes.com and phoebe.789 on Signal.

For now, all of the capacity from Crusoe’s Spark units gets sold to customers—including AI startup Decart, which uses it to power AI video generation—through the company’s cloud offering.

Use cases could be companies that are prioritizing speed and want to put their AI data centers near their customers for faster response times, or enterprises like hospitals that need to run their AI systems onsite, in a private data center because of security requirements, Cavness said. Crusoe is also hoping to deploy Spark units on-site and sell that compute through a new product offering called Edge Zones, though it hasn’t done so yet.

Crusoe’s announcement comes days after Bloomberg and The Information reported that Oracle opted not to expand its partnership beyond the initial 1.2 gigawatt Stargate data center Crusoe is building in Texas.

Cavness told Forbes the reports were “much ado about nothing,” and said that Crusoe and Oracle were moving forward with the initial eight data centers in Abilene as planned and noted that Crusoe had never announced an expansion with Oracle.

“There was a lot of speculation and people kind of making up explanations…a lot of these reports are not really based in fact,” he said.

The Information reported Wednesday that Microsoft and Meta have been considering leasing the extension to the Abilene campus.

Regardless of what is happening with Abilene, modular data centers are on the rise because hyperscale is turning into a national pastime of delays, power bottlenecks, and public pushback. Public real estate firm JLL projects that annual global sales of the smaller, prepackaged data centers will reach $48 billion by 2030, up from $11 billion in 2025. And as AI demand shifts from training models to using them, smaller units can sit closer to users, which means faster responses and less hauling data across the country. (CoreWeave chief development officer Brannin McBee told Forbes last fall that its data centers were split an estimated 50-50 between training and use; that could trend to 15-85 in five years.)

Still, Crusoe is entering an already-populated market. Cavness argues Spark is built more specifically to manage AI use cases faster and more efficiently, but incumbents like Vertiv, Schneider Electric and Equinix have been selling a version of this for years, and newer entrants InfraPartners, Fermi and BluSky AI are showing up with fresh logos and the same basic promise: compute, faster, in a box.

“Modular [data centers] are here to stay,” says Sean Farney, vice president of data center strategy at JLL. They make it easier to get to revenue sooner, lifting: “all customer segments, products and anything you can build and get to go live.”

It’s a smart move for Crusoe if the doomers are right that expensive hyperscale data centers could get overbuilt—if the gigawatt land grab turns into a wasted opportunity. In that scenario, the winners won’t be the companies with the biggest construction site; they’ll be the ones with something fast, efficient and shippable.

© Forbes