The turbines never shipped.
Crusoe has ended its $1.25 billion agreement to buy 29 natural gas-fired Superpower turbines from Boom Supersonic, the Denver aerospace company better known for trying to bring back commercial supersonic flight. The deal collapsed before a single unit was delivered. Boom, which had named Crusoe as the first customer for its stationary power business, no longer has one.
I spend most of my time thinking about agent architectures, not gas turbines. But these two subjects have quietly merged into the same problem. Every design decision I make about how an agent plans, retries, calls tools, and holds state eventually resolves into a question about watts. And watts, it turns out, are the part of the stack with the longest lead time and the least forgiving physics.
Why a power deal is an architecture story
Agent systems have a demand profile that looks nothing like the web workloads data centers were built around. A search query is a spike. A training run is a long, flat, predictable plateau. An agent fleet is something stranger: bursty, recursive, and self-amplifying. One user request can fan out into dozens of model calls, each spawning tool invocations, each potentially triggering another round of reasoning. Load is a function of task difficulty, which nobody can forecast in advance.
That shape is brutal for anyone trying to size power infrastructure. You are asked to commit capital years ahead of deployment against a demand curve whose slope depends on whether the field settles on short single-pass responses or long multi-step deliberation. Those two futures differ by more than an order of magnitude in energy per completed task. Nobody knows which one wins.
So a company like Crusoe faces a bet with two unknowns stacked on top of each other: how much power will be needed, and where it will come from. Buying 29 turbines from a company that has never shipped one addresses the second unknown by adding a third.
The option value of not committing
There is a version of this story that reads as a failure, and a version that reads as good engineering discipline. I lean toward the second.
Canceling before delivery preserves something valuable: the ability to re-decide. Crusoe raised $3.9 billion, which buys a lot of flexibility, and flexibility is the scarcest asset in AI infrastructure right now. Committing $1.25 billion to a single unproven hardware line from a single vendor would have locked a specific power architecture into place for the useful life of those machines. Every subsequent decision about site selection, cooling, rack density, and even which accelerators to buy would have inherited that constraint.
Software people understand this instinct well. We avoid coupling our systems to a dependency that has not shipped a stable release. Extending the same caution to physical plant is not timidity. It is recognizing that the cost of an irreversible commitment scales with how uncertain the requirement is, and that requirements here are extremely uncertain.
What Boom loses
The asymmetry matters. For Crusoe, this is one supply option among many. For Boom, losing a named launch customer removes the reference deployment that would have made the next sale easier. Power equipment is sold on operating history, not spec sheets. A first customer is how you get one.
That dynamic is worth sitting with, because it is the same dynamic constraining new entrants across the AI supply chain. Buyers under time pressure default to vendors with proven units in the field. New suppliers need buyers willing to absorb first-article risk. When demand forecasts are shaky, nobody wants to be that buyer. The result is a space that talks constantly about building new capacity while concentrating orders on established equipment.
What this means for how we build agents
If the power side of the stack cannot commit years ahead, the compute side has to get better at using less. That reframes a set of research problems I already care about as infrastructure problems:
- Reasoning budgets as first-class parameters. Agents should know how much compute a task is worth and stop. Open-ended loops are an energy liability, not just a latency one.
- Aggressive caching and memoization across agent runs. Recomputing the same intermediate reasoning is the most avoidable waste in current systems.
- Model routing by difficulty. Sending every step to the largest available model is a design choice with a utility bill attached.
- Verification over retries. Cheap checks that catch a bad plan early beat expensive re-runs that discover it late.
- Load shaping. Not every agent task is interactive. Deferrable work can move to when power is cheap and available.
None of that is new advice. What changed is the reason to take it seriously. Efficiency used to be a cost-per-token conversation among people optimizing margins. A canceled turbine order makes it a capacity conversation. If the physical buildout cannot be committed to in advance, the architectural slack has to come from somewhere, and the only place left is in how we write the systems themselves.
Crusoe kept its options open. Agent builders should be asking what their own architectures would cost if the power they are counting on arrives later than planned.
🕒 Published: