I expect more applications to use open-weight or self-hosted models over the next six months. But I doubt “everything moves local” is the likely outcome.

Self-hosting has an obvious appeal when an application handles a large, predictable volume of routine work. You can control the deployment and reduce dependence on one API provider.

The catch is that model weights being available doesn’t make operations free. Someone still pays for hardware, idle capacity, deployment, monitoring, and maintenance.

For a small team with uneven traffic, an API could remain cheaper overall.

The setup I’d bet on is mixed: a smaller model handles extraction, classification, and straightforward requests; a frontier API handles difficult cases. The split changes as models, prices, and workload volumes change.

The useful comparison is total cost per accepted result.…