Picture a procurement meeting in Denver. Two local companies, both flush with attention, both selling a version of the future. One builds AI data centers and needs power yesterday. The other builds supersonic jet engines and has decided that the same core technology can sit on a concrete pad and spin a generator instead. Someone draws a line item worth $1.25 billion. Twenty-nine natural-gas turbines. Handshakes, press release, done.
Then nothing gets delivered. And now the deal is gone.
Crusoe, the Denver-based AI data center builder that recently raised $3.9 billion, has ended its agreement to buy Boom Supersonic’s stationary Superpower turbines. No units shipped. Boom loses the only customer it had publicly named for its power business. That’s the whole story, and it’s more instructive than most stories ten times its size.
Why I Care About This on a Tools Blog
I review AI toolkits. Most of what I test is software you can install in an afternoon and rip out by dinner. So why spend words on gas turbines?
Because the failure mode is identical, just with more zeros. Every week I get pitched a tool that exists mostly as a roadmap. The demo works. The docs describe features in present tense that are actually future tense. The pricing page implies a scale of operation that nobody has tested. And somewhere there’s a launch customer whose logo does an enormous amount of work convincing everyone else that the thing is real.
The Crusoe-Boom arrangement was that, at industrial scale. A named buyer, a big number, and zero deliveries. When the buyer walks, you find out what you actually had: a plan.
Announced Capacity Is Not Capacity
Here’s the pattern I keep running into when I evaluate AI infrastructure claims. Companies talk about compute and power the way a startup talks about its feature set — as a pipeline, not an inventory. Contracted, planned, announced, in development. Those words are doing real work and most coverage flattens them into “has.”
Crusoe didn’t lose 29 turbines. It lost the option to buy 29 turbines that did not yet exist as delivered hardware. Boom didn’t lose revenue it had booked and banked. It lost a name it could put on a slide. Both of those are genuine losses, but they’re losses of expectation, not of assets. If you were modeling either company based on that press release, your model was measuring optimism.
This is exactly the reading skill I wish more people applied to AI tool announcements. When a vendor says its platform “supports” twelve model providers, ask which ones have been used in production by a paying customer. When they say they’re “partnered” with a major cloud, ask what shipped. The gap between a signed agreement and a working system is where most of the disappointment in this industry lives.
The Diversification Tell
There’s a second thing worth flagging, and it applies to software vendors just as much as to aerospace firms. Boom builds supersonic aircraft. Stationary power generation is adjacent in the sense that both involve turbines, and completely different in every way that matters — customers, service expectations, uptime requirements, regulatory path, sales cycle.
I see the software version of this constantly. An observability company that suddenly ships an agent framework. A vector database that announces a full orchestration layer. The engineering may genuinely overlap. The operational discipline usually doesn’t. And the first customer for a side business is buying the reputation of the main business, which is not the same as buying a proven product.
When a company’s new line of business loses its anchor customer before delivery, that’s information. Not proof of failure, but information. It tells you the side quest hasn’t been validated by anyone actually operating it.
What I’d Actually Do With This
My practical takeaway, the same one I apply to every toolkit that lands in my inbox:
- Ask what has shipped, not what has been signed. Deliveries beat agreements. Working deployments beat deliveries.
- Treat a single named customer as a single data point. One logo is a pilot. It can vanish, and when it does, the product is back to zero references.
- Discount adjacent product lines heavily. Core competence doesn’t transfer as cleanly as pitch decks suggest.
- Watch cancellations more closely than announcements. Nobody writes a splashy press release about a deal that quietly died, which is precisely why the quiet death carries more signal.
The AI buildout is going to produce a lot of these. Enormous numbers, enthusiastic partnerships, and a steady trickle of arrangements that dissolve before anything gets installed. That’s not a scandal. It’s what happens when demand forecasts outrun supply chains and everyone signs early to hold a place in line.
Just don’t confuse the place in line with the product. Crusoe and Boom didn’t fail at building turbines. They didn’t get far enough to try.
🕒 Published: