Local vs cloud: the energy trade-off
Running a model on your own machine isn't automatically greener. Here's how to think about it.
Jamie Owen Updated 31 May 2026
| Task | Local model | Cloud model |
|---|---|---|
| Per-answer efficiency | Often lower | Often higher |
| Uses purpose-built hardware | Rarely | Yes |
| Idle / overhead waste | Your device | Shared at scale |
| Greenest when… | Your grid is clean | Provider grid is clean |
| Data privacy | Stays local | Leaves your control |
It’s tempting to assume that running a model locally is the greener choice because “nothing leaves your machine”. The energy picture is less tidy than that, and sometimes it points the other way.
Why cloud can be more efficient per answer
Large providers run purpose-built accelerators at high utilisation, in data centres tuned for efficiency, increasingly on clean energy. A consumer GPU doing the same work is usually less efficient per answer, and a powerful machine left running draws power whether or not you’re using it.
Where local wins
Local avoids the network and shared-infrastructure overhead, and a small model on modest hardware can be genuinely light. The deciding factor, more often than people expect, is your local electricity grid versus the provider’s. Clean power changes the answer.
How to read the table
These are tendencies, not guarantees. The real footprint depends on the model size, your hardware, utilisation, and the grid behind each option. Local’s clearest advantage isn’t energy at all. It’s privacy.
Don’t pick local for the planet by default. Pick it for privacy and control, and choose the greener option case by case, with your grid in mind.