When owning your AI stack beats paying per token

Short version: cloud by default, local for the three cases that earn it. Owning your AI stack is not a flex and not a religion. It is a per-task cost-and-privacy decision you should be able to defend on a spreadsheet.

Cloud vs local: the honest comparison

For each task, ask four things. Is the data sensitive? Is the volume high and steady? Must the output stay identical over time? Or is it low-volume and fast-changing? Those answers route the work far better than any preference does.

When each one wins

• Cloud wins for heavy thinking and anything low-volume or fast-changing. You get the strongest model and maintain nothing.
• Local wins on privacy (the data cannot leave your walls), on cost at volume (the same task runs constantly and a fixed machine beats a per-token bill), and on control (the model must stay exactly the same).

I run this split myself, including a local pipeline that turns a photo and a voice clip into a talking-head video on one desktop GPU, with no per-minute fee. Cloud does the heavy reasoning; the repetitive, private, high-volume work stays on hardware I own.

Where it goes wrong

• Self-hosting at low volume, where maintenance costs more than you ever save.
• Choosing local for work that genuinely needs the strongest model.
• A local box with no monitoring that quietly degrades until someone notices the output got worse.
• Privacy assumptions that do not hold, because the data still passes through a logging layer you forgot about.

Self-hosting AI is not a flex. It is a cost-and-privacy decision you should be able to defend on a spreadsheet.

Bring me the workflow you want AI inside. I'll tell you what I'd run in the cloud and what I'd keep on your own hardware.