Moving a job does not delete it
Renting compute feels weightless, because the fans are in somebody else’s building. The watts are not weightless. A rented GPU hour draws roughly what the same silicon draws under your desk, plus the overhead of the room around it.
That overhead is measurable and, in a modern facility, small — ten to twenty per cent on top of the compute for cooling and distribution. So the cloud is neither free nor automatically dirty. The variable that decides the argument is utilisation.
Where the cloud genuinely wins
- Bursty work. A card that renders one weekend a month spends the other twenty-nine days depreciating; a rented one spends them earning for someone else.
- Cleaner grids. The same job runs dirtier on a still winter evening here than in a region running on hydro — and the scheduler, not the hardware, picks that.
- Peaks you would otherwise buy hardware for. Sixty-four cores needed twice a year is the textbook case for renting them.
- Shared idle. Silicon held near full utilisation across many tenants wastes far less of its manufacture than a machine idling at four per cent under a desk.
Where local wins
Steady daily work is the mirror case. If your card is busy through most working hours, its manufacturing cost is already spread thin, and every hour rented on top adds transfer, orchestration and a second machine’s share of embodied carbon.
Data gravity settles the rest. A cold archive replicated three ways in a data centre runs for years; the same footage on a local disk that spins up twice a year costs almost nothing — and uploading a terabyte has a footprint of its own before a single frame renders.
So the question is never local against cloud, it is busy against idle. Count the hours your hardware genuinely works: above about a third of the working week, keep it under the desk; well below that, rent it and stop buying idle silicon.
