Imagine ordering a custom kitchen from a company that has only ever built motorcycles. The renderings look great. The spec sheet is a work of art. Then, months later, the order gets cancelled and you realize the kitchen never existed anywhere except in a PDF. That’s roughly the shape of what just happened between Crusoe and Boom Supersonic.
Crusoe, the Denver-based AI data center builder that recently raised $3.9 billion, has ended its $1.25 billion agreement to buy 29 natural gas-fired Superpower turbines from fellow Denver company Boom Supersonic. No turbines were delivered. Boom, which had named Crusoe as the first customer for its new stationary power business, no longer has a launch customer.
I review AI tooling for a living, which means I spend a lot of my week separating things that exist from things that are announced. This story is the infrastructure-scale version of a problem I run into constantly, and I think it’s worth talking about on those terms rather than as an energy-sector scoop.
Announcements are not products
The pattern here is one I recognize from the software side. A vendor with credibility in one domain announces an adjacent product. The announcement lands with a big-name customer attached, and the size of that customer’s commitment becomes the proof point everyone cites. The dollar figure does the work that a shipped, tested product would normally do.
Boom builds supersonic aircraft. Turbine expertise is genuinely adjacent to that, and the Superpower line wasn’t a random pivot. But adjacency isn’t delivery. A $1.25 billion order from an AI data center company reads as validation, and for a while it was treated that way. Then the order went away, and what remains is the same thing that existed before the announcement: a product line without a customer.
When I evaluate an AI toolkit, the number I care least about is funding raised or contracts signed. What I want to know is whether anyone is running it in production and what broke when they did. Those two questions kill a shocking percentage of the tools that show up in my inbox. This deal would have failed the same test, because the honest answer was always “nobody, because none have shipped yet.”
Why the buyer walking away matters more than the seller stumbling
The instinct is to read this as bad news for Boom, and it clearly is. Losing a launch customer before first delivery is about as rough as a hardware debut gets. But the more interesting signal comes from the other side of the table.
Crusoe just raised $3.9 billion. This is not a company that cancelled a deal because it ran out of money. A buyer with that much capital walking away from a power commitment of this size, before taking delivery, tells you something about how the buyer now views the risk of untested equipment sitting in the critical path of a data center build.
Power is the constraint on AI buildouts right now. If you’re Crusoe, the turbine is not a line item, it’s the thing that determines whether the site produces revenue. Betting that part of the stack on a first-generation product from a company whose core business is airplanes is a different kind of bet than trying a new orchestration framework. One you can rip out in an afternoon. The other you can’t.
The reviewer’s version of this lesson
Everyone building on AI infrastructure right now faces some version of the same choice: take the new thing with the better spec sheet, or take the boring thing with a service history. I lean boring more often than readers expect, and it’s not because new tools are bad. It’s because the cost of a failed dependency scales with how deeply it’s buried in your system.
A few things I’d take from this:
- A large signed commitment is a claim about the future, not evidence about the present. Treat it as marketing until units are in the field.
- Adjacent expertise is real but not transferable on demand. Good at jet engines does not automatically mean good at stationary power plants delivered on schedule.
- The deeper a component sits in your stack, the more boring it should be. Experiment at the edges.
- Watch what sophisticated buyers cancel, not just what they announce. Cancellations are the more honest signal, and they get a fraction of the coverage.
There’s a version of this story where Boom’s turbines turn out fine, find other buyers, and this becomes a footnote about two companies whose timelines didn’t line up. That’s entirely possible, and nothing about the cancellation says the product is bad. What it says is that neither party was willing to find out together at this scale.
For those of us evaluating tools rather than turbines, the takeaway is smaller and more useful. The gap between announced and shipped is where most bad technology decisions get made. A $1.25 billion deal collapsing in that gap is just a very expensive reminder to ask the unglamorous question first: has anyone actually run this yet?
🕒 Published: