Twenty-nine turbines. That was the size of the order Crusoe placed with Boom Supersonic, a $1.25 billion commitment to buy natural gas-fired power plants for AI data centers. As of this week, that order no longer exists. Crusoe walked away before a single turbine was delivered, and Boom lost the first customer it had ever named for its power business.
If you follow AI agents the way most people do, through chat windows and browser tabs, a canceled turbine contract sounds like someone else’s problem. It isn’t. It’s one of the clearest windows we get into what actually keeps your agents running.
Why a supersonic jet company was selling power plants
Boom Supersonic is best known for trying to bring back fast passenger flight. Building a supersonic airliner means building engines, and engine technology doesn’t care much whether it’s bolted to a wing or a concrete pad. Boom took that work and spun up a stationary power line called Superpower, turbines that burn natural gas to generate electricity on site.
Crusoe, based in Denver like Boom, builds AI data centers. It recently raised $3.9 billion. Two neighbors, one with power generation hardware to sell and one with an enormous appetite for electricity, striking a deal worth $1.25 billion. On paper it reads like a tidy match.
Then Crusoe ended it. No deliveries, no installed capacity, no track record either way.
What this has to do with your AI agent
Here’s the part that tends to get lost in coverage of AI infrastructure. When you ask an agent to sort your inbox, summarize a contract, or research a supplier, you are placing an order for electricity. The agent isn’t thinking in some abstract cloud. It’s running calculations on physical chips inside a building, and those chips draw power continuously.
Agents draw more of it than a single chatbot reply does, because agents don’t answer once and stop. They plan, act, check the result, and try again. A single request can trigger dozens of model calls behind the scenes. Multiply that across millions of users and you get an electricity problem that looks less like a software budget and more like a utility one.
That’s why AI companies are buying turbines. The grid in many places cannot deliver new power fast enough, so data center builders are generating it themselves, on site, often with natural gas because it can be deployed relatively quickly.
Reading the cancellation without overreading it
We don’t know why Crusoe pulled out. The verified facts are narrow: the plan was abandoned, nothing shipped, Boom lost its launch customer. Anything beyond that is guesswork, and I’d rather give you a clear picture of a small fact than a confident story built on air.
What the cancellation does tell us is something about the shape of this moment. Deals of this size are being announced at a pace that makes them feel settled, and some of them are not. An agreement to buy 29 turbines is a plan, not a power plant. The distance between those two things is where a lot of AI infrastructure news currently lives.
For Boom, losing your first named customer before delivery is a real setback for a new product line, because early customers are how unproven hardware earns credibility. For Crusoe, having $3.9 billion in fresh funding means options, and changing your mind is one of them.
Three things worth taking from this
- Announced capacity is not delivered capacity. When you see a headline about a company securing power for AI, check whether anything has actually been built. Often it hasn’t yet.
- Agent costs are physical. The price you pay per token traces back to hardware, cooling, and fuel. When those get more expensive or harder to source, it eventually shows up in what agents cost to run.
- The supply chain is improvising. A supersonic aircraft company selling generators to a data center startup is a sign of how unusual the current demand is. Normal suppliers can’t meet it, so unusual ones step in, and some of those arrangements won’t hold.
The unglamorous layer
Most writing about AI agents focuses on capability. Can it book the flight, file the ticket, write the code. Those are fair questions. But every capability sits on top of a stack that ends in a turbine, a transformer, or a substation, and that stack is where the genuine constraints are right now.
A $1.25 billion deal collapsing before delivery is a useful correction to the idea that AI scaling is purely a software story. Chips need power. Power needs infrastructure. Infrastructure needs years, permits, and suppliers who deliver what they promised.
Your agent will keep working tomorrow. But the reason it can is being negotiated, financed, and occasionally canceled in rooms very far from your screen, and it’s worth knowing that layer exists.
🕒 Published: