The most interesting engineering decision in AI infrastructure this month was a decision not to build something. Crusoe, the Denver-based AI data center company that recently raised $3.9 billion, walked away from a $1.25 billion agreement to buy 29 natural-gas-fired Superpower turbines from Boom Supersonic. The common read is that this is a setback for Boom, which just lost the first named customer for its power business. That part is true. But treating this as a story about a supplier losing a sale misses what it tells us about how the people closest to AI compute now think about capacity.
My reading is the opposite of the disappointment narrative: Crusoe canceling is a sign that AI infrastructure planning is maturing, not faltering. Crusoe said it is shifting to a flexible mix of energy sources. In systems terms, it traded a large fixed commitment for optionality. Anyone who has designed a scheduler, a memory hierarchy, or an agent runtime knows exactly how that tradeoff feels.
Power as a scheduling problem
At 42 megawatts per unit, 29 turbines add up to roughly 1.2 gigawatts of dedicated generation. That is an enormous, indivisible bet on a single technology from a single vendor on a single delivery schedule. It is the infrastructure equivalent of statically allocating your entire heap at process start because you are confident you know the shape of the workload.
The trouble is that nobody knows the shape of the workload right now. That is not a knock on forecasting skill; it is a property of the moment. The compute profile of AI in 2026 is being pulled in several directions at once, and the directions have genuinely different power signatures.
- Large pretraining runs want steady, enormous, highly predictable draw over weeks or months. That pattern rewards dedicated baseload generation.
- Inference serving is spiky and latency-sensitive, shaped by human and machine traffic patterns rather than by a training curriculum.
- Agent workloads, the thing I spend my time on, are spikier still, and they are the least well understood of the three.
That last category deserves more attention than it gets in energy planning conversations. An agent system is not a single forward pass. It is a loop: plan, call a tool, wait, evaluate, retry, branch, sometimes spawn parallel subtasks and reconcile their results. Long-horizon runs interleave bursts of heavy generation with idle stretches spent waiting on external systems. Multiply that across a population of concurrent agents and the aggregate demand curve looks nothing like a training job and nothing like a chat endpoint either. It has bursts, bimodal tails, and correlated spikes when many agents react to the same trigger.
Why a fixed commitment is the wrong shape
If you are provisioning for a workload whose duty cycle you cannot yet characterize, buying 1.2 gigawatts of one thing is a bet on a curve you have not measured. A mixed portfolio is not indecision; it is how you buy the ability to be wrong cheaply. Different sources have different ramp rates, different marginal costs, and different siting constraints. Combining them gives you something closer to a tiered storage hierarchy than a single flat pool: fast and expensive for the peaks, slower and cheaper for the floor.
There is also a timing asymmetry that I think drives the decision more than anything. Model architectures, quantization schemes, accelerator generations, and serving stacks iterate on a timescale of months. Power generation iterates on a timescale of years. When a fast-moving layer commits hard to a slow-moving layer, the fast layer loses the thing that made it fast. Keeping the energy layer loosely coupled preserves the ability to change what you run on top of it.
What this costs the ecosystem
I do not want to be glib about the downside. Boom now has a power product without its anchor buyer, and anchor customers are what make capital-intensive hardware bets financeable. A cancellation of this size sends a real signal to anyone building purpose-made generation for AI facilities: your customers may not want to be locked in, which makes your own planning harder. Optionality for the buyer is uncertainty for the supplier, and that cost lands somewhere.
The broader lesson is not that dedicated generation for AI is a bad idea. It is that the industry is still discovering what its own load curve looks like, and the smart players are refusing to pretend otherwise. For those of us working on agent architecture, that should register as a prompt. If agents become the dominant consumer of inference cycles, then the way we design loops, batch tool calls, cache intermediate state, and decide when to escalate to a larger model stops being purely a software concern. It becomes an input to how much generation gets built and what kind.
Crusoe just declined to answer that question a decade early. Hard to argue with the instinct.
🕒 Published: