I fine-tuned a 7B model to turn raw account data into structured JSON briefs.. it's up on Hugging Face, here's how it went and what I still haven't measured.
Data in an AI stack doesn't live in one place. It sprawls across caches, logs, vector stores, and traces. Residency is a surface, and you can't prove a surface from a diagram.
Everyone's composing small specialized models now, and everyone's moving them onto hardware they own. Almost nobody is measuring what happens when those two collide.
Connection isn't the hard part anymore. Reasoning about resources is — the resource-decision gap the current protocol stack leaves wide open.