The Only Thing I Run Locally Is What I'd Be Fired For Losing

I'm not a self-hosting evangelist. I rent the frontier models for almost everything because it's easier and the quality is there. But there's one class of work I refuse to send to anyone's API, and it's the class that would end my career if it leaked.

Customer data, unreleased product details, anything with a person's name attached, that goes on hardware I control, on an open-weight model I downloaded. The slight quality drop is a price I pay for the guarantee that the bytes don't leave the building.

The open-weight layer got good enough this year to make this practical. A trillion-parameter model you can run yourself isn't the frontier, but it's good enough for the sensitive 5% of my work, and that's the part where the trade-off actually matters.

Everything else, the drafts, the summaries, the brainstorming, rides the cloud, because the convenience is worth more than the risk there. The hybrid only works if you're honest about which bucket each task is in.

Don't tear down your subscriptions. Move the one workload where data leaving the building was never acceptable to hardware you own, and leave the rest where it's easy.