What if the biggest constraint on the AI agent you use every day isn’t code at all, but a gas turbine that never got built?
In September 2026, Crusoe walked away from a $1.25 billion agreement to buy 29 stationary Superpower gas turbines from Boom Supersonic. Each unit was rated at 42 megawatts, and the first deliveries were scheduled for 2027. Crusoe had signed on as Boom’s first customer for that line of business. The turbines were earmarked for AI data centers, including a campus in Abilene, Texas. The reason given was a shift in energy strategy.
If you’re reading this site, you’re probably more interested in what AI agents can do for your inbox than in turbine procurement. Stay with me, because this story explains something most explainers skip: the physical reason your AI tools feel fast, slow, expensive, or rate-limited.
Why a jet company was selling power plants
Boom Supersonic builds engines for fast passenger aircraft. A jet engine and a stationary power turbine are cousins. Both spin a shaft by burning fuel and pushing hot gas through blades. Bolt the same basic machine to a concrete pad instead of a wing, connect it to a generator, and you have electricity.
That crossover became attractive because AI data centers need enormous, steady power, and traditional turbine makers have long waiting lists. A company with aerospace engineering talent and spare manufacturing ambition looked like a shortcut. Crusoe, which builds and operates AI infrastructure, was willing to be the anchor buyer.
Twenty-nine turbines at 42 megawatts each is roughly 1.2 gigawatts of generating capacity on paper. That is power plant territory, ordered by a company whose business is renting out computation.
The part that matters for agents
Here is the mental model I’d suggest. When you ask an AI agent to read a document, search the web, and draft a reply, you are renting a slice of a warehouse full of chips. Those chips turn electricity into tokens. Every agent task you run is, physically, an electricity purchase made on your behalf.
So when an infrastructure company changes its mind about a gigawatt of on-site generation, that’s not trivia. It’s a signal about how the people closest to the meter are thinking about the next few years. Some things worth keeping in mind:
- Agent capacity is bounded by power, not just by model quality. A smarter model still needs somewhere to run.
- Power decisions run on multi-year timelines. Turbines ordered in 2026 for 2027 delivery serve workloads nobody can fully predict yet.
- Cancellations are part of how this gets figured out. Committing to the wrong generation plan for a decade is worse than backing out early.
Money was clearly not the problem
The same month Crusoe cancelled, it closed a $3.9 billion Series F round at a valuation of roughly $30.9 billion. That timing is the interesting part. This wasn’t a company scraping together payroll and cutting a big order to survive. It was a company with fresh capital deciding it had better uses for that money.
Crusoe also stepped back from a planned large-scale AI campus in Wyoming. Two retreats in the same period suggest a deliberate narrowing rather than a one-off change of mind. What replaces those plans hasn’t been laid out publicly, so I won’t guess at the specifics.
What I take from it
The story I keep hearing about AI infrastructure is a story of relentless addition. More chips, more campuses, more megawatts, all going up and to the right. This is a reminder that the real process is messier. Plans get drawn, priced, signed, and then dropped when the arithmetic changes.
For non-technical readers trying to make sense of AI news, that’s actually reassuring. You don’t need to treat every announced project as a done deal. A signed agreement is a statement of intent under a particular set of assumptions about cost, timing, and demand. Change one assumption and the whole thing can come apart, even with billions in the bank.
There’s also a lesson about vendors. Boom lost its first customer for a new product line before deliveries began. Anyone building on a brand-new supply chain, whether that’s turbines or an agent framework from a six-month-old startup, is taking on the risk that the supplier’s roadmap changes. That risk is real at every scale.
If you want one takeaway to carry into conversations about AI: the agents are software, but the limits are physical. Watch the power stories. They tend to arrive before the product stories do, and they explain a surprising amount about what you’ll be able to run next year and what it will cost you.
🕒 Published: