Every developer has done this. You spec out a project, pick the most interesting piece of tech you can find, spend three weeks wiring it in, and then quietly rip it out because the boring option turned out to be the right one. The difference is that when I do it, the sunk cost is a weekend. When Crusoe does it, the sunk cost is a $1.25 billion purchase order for 29 gas turbines derived from supersonic jet engines.
In September 2026, Crusoe walked away from its agreement to buy Boom Supersonic’s 42-megawatt Superpower turbines for AI data centers, including its campus in Abilene, Texas. Crusoe had signed on as Boom’s first customer for the stationary turbine business, with first deliveries slated for 2027. That deal is off. The stated reason is a strategic shift in energy strategy.
Same month, Crusoe closed a $3.9 billion Series F at roughly a $30.9 billion valuation. It also stepped back from a planned large-scale AI campus in Wyoming. So this is not a company running out of money. This is a company with more money than it had last quarter deciding it liked its original plan less.
Why this matters to people who pick tools, not turbines
I review AI toolkits for a living. Most of what I write is about whether a framework holds up under real load, whether the docs match the API, whether the vendor will still exist in eighteen months. Turbines feel far away from that. They are not.
The pattern here is identical to the one I see in software procurement every week: a team picks the exciting adjacent-industry solution because it solves a real constraint in a clever way, then discovers that clever and available are different things. Boom’s pitch was genuinely appealing. Supersonic engine technology repurposed for stationary power, sized at 42 megawatts a unit, aimed squarely at the problem everyone building AI capacity has right now, which is that grid interconnection queues are long and compute demand is not patient.
Twenty-nine units. First deliveries in 2027. On paper, that is a power plan. In practice, you are betting your data center’s energy supply on a product line from a company whose core business is building an airplane that does not fly commercially yet.
The dependency question nobody asks early enough
When I evaluate a tool, the question I keep coming back to is not “does this work” but “what happens to me if this vendor changes direction.” Crusoe just answered that question from the other side of the table. It was the anchor customer, the validating logo, the reason Boom’s stationary turbine business looked like a business. And it left.
There is a version of this story where Crusoe looks reckless for signing in the first place. I do not think that is the read. The more useful read is that Crusoe kept the option to reverse. It did not build the entire Abilene power architecture around an unshippable product and then find out in 2027. It reassessed and ate the reversal cost while the reversal was still possible.
That is the discipline I wish more engineering teams had with their stack. The decision to adopt something new is easy to make and hard to unmake, and most teams treat the unmaking as failure. Crusoe treated it as arithmetic. The power math changed, so the answer changed.
What I take from it
- Novel supply is not supply. A 42-megawatt turbine that exists in a press release is not a 42-megawatt turbine. Same goes for the API endpoint that is “coming in Q3.”
- Being the first customer is a position, not a prize. You get pricing and attention. You also get to absorb every delay the vendor has not discovered yet.
- Funding does not validate architecture. Crusoe raised $3.9 billion and cancelled a marquee deal in the same month. Capital and correctness are separate variables.
- Reversibility has a price and it is usually worth paying. Crusoe kept an exit. Most of the worst tooling decisions I have reviewed had no exit at all.
The part I am still watching
Crusoe has not said publicly what replaces those turbines, and I am not going to guess. The Wyoming pullback alongside the turbine cancellation suggests a broader rethink of where and how the company builds, but the specifics are not on the record.
What I can say is that this is the cleanest recent example of a large infrastructure buyer publicly unwinding a headline commitment without a crisis forcing it. In a market where nearly every AI announcement is additive, a subtraction is informative. Somebody ran the numbers again and did not like them, and that is a more honest signal than most of what gets published.
Build in the ability to change your mind. The companies spending billions are doing exactly that.
🕒 Published: