It’s 2 a.m. and you’re watching a dashboard. Your bot is chewing through a backlog of support tickets, and the p95 latency graph has a hump in it you can’t explain. You didn’t change the prompt. You didn’t change the model. You check the provider status page: all green. So you shrug, bump your concurrency limit down, and go to bed.
That hump had nothing to do with your code. Somewhere several abstraction layers below your retry logic, someone was making a decision about gas turbines.
In September 2026, Crusoe walked away from a $1.25 billion agreement to buy 29 of Boom Supersonic’s 42-megawatt Superpower stationary turbines for its AI data centers, including a campus in Abilene, Texas. Crusoe had signed on as the first customer for that turbine business. First deliveries were supposed to start in 2027. The reported reason for the cancellation was a shift in energy strategy. In the same month, Crusoe closed a $3.9 billion Series F at a valuation of roughly $30.9 billion, and stepped back from a planned large-scale AI campus in Wyoming.
Why a bot builder should care about turbines
Most of us who build bots operate at the top of a very tall stack. We write a tool-calling loop, we tune a retriever, we argue about whether to stream tokens or batch them. The stack underneath us is someone else’s problem until it isn’t.
Underneath the API key is a model. Underneath the model is a cluster. Underneath the cluster is a building. Underneath the building is electricity, and electricity has lead times measured in years and contracts measured in billions. When a company cancels an order for 29 turbines that were going to start arriving in 2027, that’s a change to the physical substrate that your 2028 inference bill sits on.
I don’t say that to be dramatic. I say it because it reframes what “capacity” means in practice. We tend to treat compute as a faucet: turn it on, pay per liter. The Crusoe story is a reminder that the faucet is attached to a plumbing project, and plumbing projects get re-planned.
The part I actually admire
Walking away from a $1.25 billion commitment you publicly signed is not a fun press cycle. Being the announced first customer for a new product line and then not being that customer is worse. And Crusoe did it anyway, apparently because the math on how to power these sites changed.
I’ve shipped enough bots to know how rare that is at any scale. I’ve watched teams keep a vector database they’d outgrown because migrating would look like admitting the first choice was wrong. I’ve kept my own bad abstractions alive for months past their expiry date for exactly the same reason. The sunk cost isn’t the money. It’s the story you told about the decision.
The interesting signal here isn’t “gas turbines are out.” We don’t have enough detail to claim that. The signal is that a company raising $3.9 billion in the same month it cancels a $1.25 billion order is clearly not short on options. It reconsidered while it had room to reconsider. That’s the version of this you want to copy.
What that looks like in a bot codebase
Scale down the lesson until it fits on your laptop:
- Keep your model calls behind one interface. One module that owns provider selection, retries, and timeouts. If you can’t swap a provider by editing one file, you’ve built a commitment, not a choice.
- Write down why you picked something. A three-line decision note with the assumptions in it. When an assumption changes, you get to change your mind without relitigating the whole thing from memory.
- Measure your cost per task, not per token. Token prices move for reasons far outside your control. Cost per resolved ticket or per completed workflow is the number that tells you whether to re-architect.
- Design for degradation. A smaller model, a cached answer, a queued job, a human handoff. If your bot only works when the big model is fast and cheap, you’ve outsourced your reliability to infrastructure decisions you’ll read about in the news.
- Don’t build your product on an unreleased dependency. Crusoe was the first customer for hardware that hadn’t shipped yet. That’s a defensible bet at their scale, with their capital. It’s rarely a defensible bet for a small team betting on a preview API.
The unglamorous takeaway
AI infrastructure looks solid from the outside because we interact with it through a stable-looking URL. The reality is a lot of very large, very reversible plans being adjusted in public. Abilene still needs power. Wyoming’s plans changed. The turbines aren’t coming.
Build like your foundation is being re-poured underneath you, because it is. Then go fix that latency hump — it was probably your retry logic anyway.
🕒 Published: