What actually powers the AI agent you talked to this morning? Not magic, not a cloud, not some abstract server humming politely in the background. Electricity. A lot of it. And the story of how that electricity gets made is where a $1.25 billion deal just fell apart.
Here’s what happened. Crusoe, a Denver-based company that builds data centers for AI workloads, had agreed to buy 29 natural gas-fired Superpower turbines from Boom Supersonic, another Denver company better known for trying to bring back supersonic passenger flight. The total value was $1.25 billion. Crusoe has now ended that agreement. No turbines were ever delivered. Boom lost the first customer it had publicly named for its power business.
That’s the whole factual core. But if you’re someone who uses AI agents without thinking much about the plumbing underneath, this is one of the more useful stories to pay attention to.
Why an AI company was buying jet-engine-adjacent hardware
Let me connect the dots, because the pairing sounds strange at first. Boom Supersonic builds engines for aircraft. Crusoe builds buildings full of chips. What’s the overlap?
Turbine technology. The same basic engineering that spins a turbine to push a plane forward can spin a turbine to generate electricity. Boom took that expertise and spun up a stationary power line, which is what “Superpower” refers to: not aircraft engines, but generators that sit on the ground and make power.
Crusoe wanted them because AI data centers need enormous, reliable, on-site power, and getting that power from the existing electrical grid is slow. Connecting a large new facility to the grid can take years of permits, studies, and queueing. If you can generate your own electricity on-site with natural gas turbines, you skip a lot of that waiting. For a company racing to build capacity, speed matters more than elegance.
What this tells you about AI agents
When I explain AI agents to people, I usually start with the software side: an agent is a program that can take a goal, break it into steps, use tools, and act on your behalf. That’s the part you see.
The part you don’t see is that every one of those steps costs energy. An agent that checks your calendar, drafts three emails, searches the web twice, and summarizes a document is making many separate calls to large models running on specialized chips in a building somewhere. Multiply that by millions of users, and you get why a company like Crusoe was willing to sign a ten-figure power deal.
So when you read about AI infrastructure news, you’re reading about the actual physical limit on how many agents can exist, how fast they respond, and how much they cost you to run. Software ambition is cheap. Electricity is not.
Deals fall apart, and that’s normal
I want to be careful here, because we don’t know why Crusoe walked away. The verified facts tell us the agreement ended and nothing was delivered. They don’t tell us whether it was pricing, timelines, technical readiness, a change in Crusoe’s build plans, or something else entirely. Anyone telling you the reason with confidence is guessing.
What I can say is that this kind of reversal is a normal feature of a fast-expanding industry, not a scandal. Crusoe had recently raised $3.9 billion, which means it has the capital to change its mind. Large announced plans get restructured all the time when the underlying assumptions move. The announcement makes headlines; the quiet unwinding usually doesn’t.
The sharper consequence lands on Boom. Losing your named launch customer for a new product line is a real setback, because launch customers do more than pay you. They validate the thing. They give other buyers permission to sign. Boom still has its aviation work, but the power business now needs a new first believer.
What to take from this if you’re not technical
A few things I’d keep in your back pocket:
- AI agents run on physical infrastructure, and that infrastructure has hard constraints. Power is the tightest one right now.
- Announced deals are intentions, not outcomes. A signed agreement is a plan, and plans change before anything ships.
- Watch for what gets delivered rather than what gets announced. Delivery is the only signal that survives contact with reality.
- When AI companies start buying power plants instead of renting server space, that tells you something about how much demand they expect.
The agents themselves will keep getting more capable and more useful. But the ceiling on how far they spread is being set right now in deals like this one, in boring conversations about turbines and megawatts and grid queues. That’s the part of the AI story that rarely makes the front page, and it’s the part that decides what you’ll actually be able to use next year.
🕒 Published: