Crusoe raised $3.9 billion. Crusoe also just walked away from a $1.25 billion turbine order without taking delivery of a single unit. Those two facts sitting next to each other tell you something useful about how AI infrastructure actually gets built, and it isn’t the story the press releases were telling six months ago.
The short version: Crusoe, the Denver-based AI data center builder, cancelled its agreement to buy 29 natural gas-fired Superpower turbines from Boom Supersonic, its neighbor across town. The deal is dead before hardware moved. Boom, which had named Crusoe as the launch customer for its new stationary power business, no longer has one.
I build bots for a living. I don’t pour concrete or wire switchgear. But I’ve been burned enough times by a dependency that looked certain right up until it wasn’t that this story reads less like energy news and more like a familiar failure mode at a much larger scale.
Launch customers are a load-bearing assumption
Boom’s power play depended on one buyer to prove the product. That’s the same shape as a side project that depends on a single unreleased API, or a bot architecture built around one vendor’s model that’s still in preview. It works beautifully in the planning doc. Then the other party reprioritizes, and your entire roadmap is a diagram of something that doesn’t exist.
What makes this case instructive is the timing. No turbines were delivered. The relationship ended at the stage where everything is still contracts and CAD files. In software terms, that’s a dependency that never made it past the integration test, which is the best possible place for it to fail. Expensive, embarrassing, survivable.
The alternative is the version where 29 turbines are installed, sites are designed around their specs, and then the supply relationship sours. Anyone who has inherited a codebase glued to an abandoned framework knows which scenario they’d rather debug.
What this means if you’re building on top of the AI stack
Most of us reading a site like this are several layers up from turbines. We write agents, wire tools together, ship chatbots. But the power decisions at the bottom of the stack eventually show up as capacity, pricing, and region availability at the top. When a data center builder rethinks how it generates electricity, that ripples into what inference costs six quarters from now.
The practical takeaways I’d pull from this, as someone who ships bot code:
- Treat every external provider as replaceable, on purpose. Not because you distrust them, but because a $1.25 billion agreement between two well-funded companies can evaporate. Your model provider integration should sit behind an interface you control.
- Prefer boring, delivered technology for anything load-bearing. A new turbine line and a new model API share a trait: impressive specs, no production track record. Use the new thing where failure is cheap.
- Write down your assumptions where code review can see them. “We assume this endpoint stays free” belongs in a comment or an ADR, not in someone’s head. Crusoe presumably had its assumptions documented. That’s why it could exit cleanly.
- Design for the exit. The cheapest cancellation is the one you planned for before signing. In bot terms: know what it costs to swap your vector store, your orchestration layer, your hosting, before you need to.
Reversing a decision is a skill
The part I actually respect here is the willingness to kill something big and public. Sunk cost pressure on a billion-dollar plan with a hometown partner must be enormous. Cancelling makes headlines that no one enjoys.
Bot builders face a small version of this constantly. You spend three weeks on a retrieval pipeline, then realize a simpler approach beats it. You built an elaborate agent loop and the task needed one well-shaped prompt and a database query. The instinct is to keep going because you already paid for the ticket. The better move is to accept the write-off and pick the approach that actually works.
Boom, meanwhile, is in the harder position. Losing your named first customer for a new product line is a credibility problem as much as a revenue one, because the next prospective buyer now asks why the last one left. That’s a real cost, and it doesn’t show up on a balance sheet.
The unglamorous foundation
AI conversation tends to stay in the top layer, where the demos live. Underneath sits a much slower world of turbines, transformers, permits, and interconnect queues, and it moves at a pace that makes model release cycles look frantic.
Stories like this one are a reminder that the assistant answering a question in 400 milliseconds rests on physical infrastructure that someone is still arguing about how to power. The plans down there change. Build your own stack so it can absorb that without a rewrite.
🕒 Published: