The last ten years of advice about infrastructure could be reduced to one word: rent. Rent the servers, rent the storage, rent the GPU time, and let someone else carry the depreciation. That advice was right while the expensive thing was idle hardware. A machine sitting dark costs money, and renting meant you only paid for the light. The advice is aging badly now, because AI load does not sit dark. It burns.

The bill now follows the work

Cloud pricing was built around people. Traffic spikes at lunch, goes quiet at night, and a site might double in a good month. Renting smoothed all of that. You paid for what you used, and the flexibility was worth the margin.

Inference and agent workloads break that shape. A model that answers a request burns compute in direct proportion to the request, and an agent that runs all day burns all day. There is no quiet period left to smooth. The bill stops being a cost you can tune with autoscaling and becomes a line that rises exactly as your revenue does. At that point the cloud margin stops being the price of flexibility and starts being a tax on every unit of work.

The break-even is arithmetic

A high-end GPU rents for a few dollars an hour. Bought outright, it costs somewhere in the mid five figures and keeps running for two or three years before it is worth replacing. Run it flat out and the arithmetic flips fast: a card that would cost a few thousand dollars a month to rent pays for itself in under a year of steady use. After that it is nearly free for the rest of its life.

The catch is the word steady. The rental rate bakes in the assumption that the hardware is sometimes idle, because the cloud carries that idle time across its whole fleet and charges the busy machines for it. A workload that runs around the clock is subsidizing the clouds idle capacity. A workload that spikes once a month is getting exactly what it pays for. Owning only wins for the first kind.

A machine that runs flat out is cheaper to own. A machine that sits dark is cheaper to rent.

What owns well

Three things own well, and they share one property: predictable, always-on load.

  • Base inference: a model answering a steady stream of requests never stops, so its hardware never idles.
  • Scheduled training: a job that runs every night for six hours is booked capacity, not a burst.
  • Long-lived agents: a process that never sleeps belongs on a box you own, not a meter that never stops.

Spiky work stays on the meter. A one-off fine-tune, a seasonal spike, a weekend experiment, these are what renting is for. The retreat is a split. Own the base load, rent the burst.

Owning has a second price

Owning has a second price the arithmetic hides. Someone has to rack the machine, feed it power and cooling, keep the drivers current, and plan for the day it dies mid-job. That work is a real cost, and it is the reason the rent-everything default lasted a decade. The cloud took the ops burden and spread it across a fleet. The retreat only makes sense where the compute bill has grown big enough that paying for that ops burden yourself still comes out ahead.

The cloud is a middleman

Duelling Hares automates the middle for a living, so this is a familiar story. The cloud provider owns the hardware, the power, the cooling, and the network, and rents it all back at a margin. That margin was worth paying while the flexibility did real work. Once a workload becomes predictable, the margin is just rent paid for a door.

The retreat from rented compute is the same move the invoice layer and the registry make. Margin survives where the work is, not where the door is. A team that runs its own inference owns the whole cost curve, and the curve is what decides whether the business works. The middleman who only owned the door gets squeezed, the same way the tollbooth did when AI removed the doorway.

Renting stays right for the spike, the experiment, and the one-off. Owning is right for the load that never stops. The only real change is that the decision is now a calculation instead of a reflex, and the line moves with the bill.