Crusoe cancels its $1.25B Boom turbine order

Crusoe walked away from a $1.25 billion order for 29 jet-derived gas turbines. That is roughly 1.2 gigawatts of generation that was supposed to start arriving in 2027 and now is not on one of the largest AI data center builders' near-term plan. If you are budgeting AI compute costs two years out, this is the layer your token price actually sits on, and it moved this week.

What actually happened

Per TechCrunch:

• The deal. Crusoe had signed as the first customer for Boom Supersonic's Superpower line: 29 turbines at 42 MW each, $1.25 billion, first deliveries planned for 2027. It is cancelled.
• Why. Crusoe spokesperson Andrew Schmitt: "While Boom has been a great partner, the partnership isn't the right fit today." Boom CEO Blake Scholl said the turbines are "no longer part of Crusoe's near term primary power mix at Abilene/etc.," and added, "We're grateful for the help Crusoe gave us in shaping Superpower and continue cheering for their successes."
• Crusoe is not short of money. It raised $3.9 billion this month. It still plans to use turbines — from other suppliers.
• What Crusoe is running instead. Its Abilene campus is a 1.2 GW grid-powered site built for Oracle and OpenAI, with a separate 900 MW Microsoft facility using on-site gas turbines under development.
• Boom's side. Superpower shares about 80% of its parts with Symphony, the engine Boom is building for its Overture supersonic jet. Scholl says roughly 250 MW goes to other customers in 2027, targeting 1 GW by 2028.

This is not a failure story about either company. It is a repricing. A first-of-its-kind turbine with 2027 delivery lost to grid power and proven suppliers, in a decision made by the buyer with the most pressure to get electrons fast.

Why data center power deals matter for your business

Nobody running a 20-person company cares which turbine spins in Abilene. Everybody running one cares what a million tokens costs in 2028, and that number is a power number wearing a software costume.

The practical lesson is about commitments, in both directions. Vendors are asking small businesses to sign annual and multi-year AI contracts priced on the assumption that their own costs fall on schedule. The generation those forecasts depend on is being contracted, cancelled, and recontracted in billion-dollar chunks right now. Treat any multi-year AI price as a bet on someone else's power procurement, and pay the small premium for monthly or annual-with-out terms until the supply picture stops churning.

Three things we do on every build because of exactly this:

Keep the model call behind one interface. One adapter, provider chosen by config. When pricing moves — and it moves in both directions; we have written about mid-contract cuts as well as hikes — switching is an afternoon, not a rewrite.

Measure cost per accepted output, not per token. Token prices are the noisiest input in your stack. What a finished, reviewed, shipped unit of work costs is the number that survives a repricing.

Do not architect around cheap inference you have not been quoted. If a feature only pencils at half today's price, it is a roadmap item, not a plan.

Compute has a landlord, and the landlord has a power bill that is still being negotiated.

Sources: TechCrunch: Crusoe abandons $1.25B plan to use Boom turbines at AI data centers.