Vertical 04 of 06
Compute & Workstation Capital
Should I buy high-end local silicon or rent cloud GPU clusters?
The core problem
Local versus cloud is a utilisation question disguised as a hardware question. A $9k workstation that runs 40 hours a month costs more per GPU-hour than on-demand cloud; one that runs 300 hours a month beats it comfortably. The answer flips on hours, power, resale value and how long you keep the box.
What this vertical answers
-
01
Cost per GPU-hour, owned
Purchase price minus resale, plus power and maintenance, divided by the hours you will actually use it before it is obsolete.
-
02
The idle-silicon penalty
Cloud bills only when you run. Local hardware bills every month whether or not the model is training, which is why low utilisation quietly loses.
-
03
Repair vs. replace
The same capital logic applied to the laptop or workstation already on the desk — when a repair is rational and when it is a sunk-cost argument.
Flagship engine
Local Workstation vs. Cloud GPU Break-even
Compute cost per GPU-hour for owned hardware — purchase, power, maintenance and residual value — against on-demand cloud rates at your real utilisation, with the break-even hours per month.
Open the engine →