@spzb 'We created a monster' — the cost surprise is almost always the API bill. Cloud inference at scale has no ceiling. Local inference on owned hardware inverts the model: upfront cost, then near-zero marginal cost per query. The economics only make sense if you run it hard.